. and is best when . =ε and c = (with . % correct predictions). this suggests that choice of these parameters is important in the selection of the svm used for prediction. the predictive accuracy of undifferenced svms that exclude the country dummy variables are reported in the section headed undifferenced model with no country dummies in table . the percentages of correct predictions are between . % and . % ( . % - . %) for the general (parsimonious) undifferenced svm without country dummies. this is a substantially worse predictive performance relative to when country dummies are included in svms. this confirms the findings obtained from the ordered choice models’ predictive performance and further emphasises the importance of accounting for country effects when modelling and predicting bank ratings. svms with undifferenced variables are re-estimated using the single country index (employed in the ordered choice models) to capture country effects instead of the individual country dummies and their prediction accuracy is reported in the section headed undifferenced model with single country index of table . the predictive accuracy is substantially lower for the svms using the single country index (between . % and . % for the general model and . % and . % for the parsimonious specification) than for the svms with unrestricted dummies but are similar to those produced by the ordered choice models (that also use this country index). hence, the superior predictive performance of svms over ordered choice models seems to be because the former can estimate the coefficients of the country dummies simultaneously and unrestrictedly, whereas the ordered choice models cannot. to explore this issue further we calculate the percentage of correct predictions from ordered choice models that include a new country index constructed using the weights obtained from the general undifferenced svm with unrestricted dummies. the weights for this model are reported in table . predictive performance is reported for ordered choice models that include this new country index in the row entitled with country svm of table . the simple and spearman rank correlation coefficients are . and . , respectively, for the correlation between the original country index and the one based upon the svm weights. hence, whilst they are highly correlated there are still notable differences between the two indexes. the original index is replaced with the new index in the general models and the general-to- specific methodology is applied to obtain parsimonious models. the predictive performance measures fall in the range of . % to . % with the parsimonious model securing the highest accuracy. it is noticeable that all of the ordered choice models that use the index based upon svm weights have a superior predictive accuracy than the best performing ordered choice model that employs the original index. however, these ordered choice models can still be substantially outperformed by the undifferenced svm that estimates the country dummies unrestrictedly. this further highlights the advantage that the svm has in terms of its ability to freely estimate the coefficients of all of the country dummies. . conclusions using data on banks from around the world we compare two different methodologies, ordered choice models and svms, for predicting and identifying the significant determinants of bank ratings. the ordered choice models unambiguously identify the following significant determinants of ratings. banks with a greater capitalisation ( equity ), larger assets [ ( )assetsln ], and a higher return on assets ( roae ) have higher bank ratings. further, if a bank’s liquidity ( )liquidity∆ increased two periods ago it ratings will rise. conversely, the greater is a bank’s ratio of operating expenses to total operating income ( oeoi ) and the more recent is the date that the rating is made ( time ) the lower is the rating of the bank. however, we also find that net operating income to total assets ( )noa has an unexpected negative influence on bank ratings. in addition, there is strong evidence that a bank’s country of origin has a significant influence on bank ratings. inclusion of this country effect substantially raises the ability of an ordered choice model to accurately predict international bank ratings relative to models that exclude country effects. the svm results confirm the importance of country effects as significant determinants of ratings: when they are excluded the predictive performance of the model considerably deteriorates relative to a model including country effects. however, although the svm method could be adapted to identify the significant variables it suggested few significant determinants using a % level. using an % confidence interval unsurprisingly raised the number of significant variables. given the arbitrary and unconventional choice of confidence level required to secure this result we considered whether the breadth of the confidence intervals was due to multicollinearity among the lags of the covariates. using an svm that incorporated variables in both difference and level form yielded more significant determinants using a conventional % confidence interval that all had expected signs on the weights. these significant variables are, equity , ( )assetsln , roae , liquidity∆ and oeoi which are the same as for the ordered choice models except noa and time , which were not indicated as significant by the differenced svm. however, the predictive performance of the parsimonious svms were considerably lower than that of the general svms suggesting that the model reduction method applied to svms leads to important determinants being excluded from the model and, therefore, not being identified. for this reason we are cautious to present the variables identified by the svm method as the only significant determinants of ratings. thus, based primarily on the results obtained from the parsimonious ordered choice models, we conclude that ratings reflect a bank’s financial position (as measured by various financial variables), the timing of when the rating was made and a bank’s country of origin. regarding the timing effect, fr may have applied more prudent views and policies as a reaction to critiques of their role during the financial turbulence of the late s. we have therefore identified a set of determinants that are revealing in how ratings agencies determine the risks of inherently opaque banks, especially given the high predictive accuracy achieved by our models. we have also found that svms can produce substantially better in-sample predictions of international bank ratings than the standard method currently used for this purpose, ordered choice models. this appears to be due to the svm’s ability to estimate a large number of country dummies’ coefficients unrestrictedly, which was not possible with the ordered choice models due to the small sample size. given that prediction is the primary purpose of modelling ratings, this is an important result. in this paper we have only considered in-sample predictions due to the fact that using some of the data for an out-of- sample data set would have severely restricted the (training) in-sample to less than the observations. however, consideration of the relative out-of-sample predictive performance of svms and ordered choice models, requiring more observations than were available here, would be a desirable avenue for further research. additionally we deliberately did not use svms with non-linear kernels to model the data in this exercise since we did not have sufficient data for an independent validation set to optimize non-linear model parameter settings. however we expect that using different kernels may improve performance. we intend to pursue these lines of research with a larger data set. references altman, e. i., and saunders, a. ( ). credit risk measurement: developments over the last years. journal of banking and finance, , – . altman, e., and rijken, h., a. ( ). how rating agencies achieve rating stability, journal of banking & finance, , - . altman, e., and rijken, h., a. ( ). a point in time perspective on through-the cycle ratings, financial analysts journal, , - . altman, e., bharath, s., and saunders, a. ( ). credit ratings and the bis capital adequacy reform agenda. journal of banking & finance , - . altman, e., saunders, a., ( ). an analysis and critique of the bis proposal on capital adequacy and ratings. journal of banking & finance , – . amato, j. d. and furfine, c.h. ( ). are credit ratings procyclical?, journal of banking and finance , – . chang c-c and lin c-j ( ) libsvm : a library for support vector machines. software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm greene w. h., ( ). econometric analysis, pearson, prentice hall, th edition. hendry d. f. and santos c. ( ). “regression models with data-based indicator variables” oxford bulletin of economics and statistics, , , – . hendry d. f., ( ). “modelling uk inflation, – ” journal of applied econometrics, , – . huang, zan, hsinchun chen, chia-jung hsu, wun-hwa chen, soushan wu. ( ). credit rating analysis with support vector machines and neural networks: a market comparative study, decision support systems, , pp. - , kamstra, m., kennedy, p., suan, t.k. ( ). combining bond rating forecasts using logit. the financial review, , pp. - . kim, s. k. ( ). predicting bond ratings using publicly available information, expert systems with applications, , pp. - kolari, j. d. glennon, h. shin and m. caputo, ( ). predicting large us commercial bank failures, journal of economics and business, , pp. – . lee, y.c. ( ). application of support vector machines to corporate credit rating prediction, expert systems with applications , pp. – . levich, r., majnoni, g., rinhart, c. ( ). ratings, rating and agencies and the global financial system, kluwer publishing. meyer, p., & pifer, h. ( ). prediction of bank failures. journal of finance, , – . morgan d. p. ( ). rating banks: risk and uncertainty in an opaque industry. the american economic review, ( ), – . pinto, a. r. ( ). control and responsibility of credit rating agencies in the united states, american journal of comparative law, , - . portes, r. ( ). ratings agency reform, vox, january, , vapnik v ( ). the nature of statistical learning theory. springer ny table : ordered choice models’ country index weights country weight country weight andorra . taiwan . canada . colombia . ireland . costa rica . norway . lithuania . sweden . malta . usa . morocco . switzerland . peru . spain . turkey . netherlands . el salvador . saudi arabia . indonesia . austria . india . czech republic . egypt . estonia . hungary . hong kong . bulgaria . iceland . latvia . jordan . argentina . san marino . benin . south africa . iran . korea rep. of . jamaica . france . kenya . united kingdom . lebanon . italy . mongolia . chile . nigeria . germany . tunisia . kuwait . kazakhstan . slovenia . romania . greece . russian federation . bermuda . philippines . qatar . venezuela . united arab emirates . vietnam . bahrain . georgia rep. of . australia . pakistan . macau . china-people's rep. . oman . ukraine . panama . albania . trinidad and tobago . armenia . mexico . azerbaijan . japan . bosnia-herzegovina . israel . macedonia (fyrom) . slovakia . niger . brazil . serbia . cyprus . sri lanka . thailand . dominican republic . poland . belarus . malaysia . bangladesh - . table notes. the country index for the ordered choice models is constructed as the sum of the products of the country weights (given in the column headed weight) and the individual country dummy variables. table : bank ratings ordered choice regressions logit specifications probit specifications variables general model parsimonious model general model parsimonious model country . ( . ) . ( . ) . ( . ) . ( . ) time – . (– . ) – . (– . ) – . (– . ) – . (– . ) −tequity – . (– . ) . ( . ) −tliquidity . ( . ) . ( . ) ( ) ln −tassets . ( . ) . ( . ) . ( . ) . ( . ) ni_margin −t – . (– . ) – . (– . ) −tnoa . ( . ) . ( . ) −toeoi – . (– . ) – . (– . ) – . (– . ) – . (– . ) −troae . ( . ) . ( . ) . ( . ) . ( . ) −tequity . ( . ) . ( . ) −tliquidity . ( . ) . ( . ) ( ) ln −tassets – . (– . ) – . (– . ) ni_margin −t . ( . ) . ( . ) −tnoa – . (– . ) – . (– . ) – . (– . ) – . (– . ) −toeoi – . (– . ) – . (– . ) – . (– . ) – . (– . ) −troae . ( . ) . ( . ) −tequity . ( . ) . ( . ) . ( . ) . ( . ) −tliquidity – . (– . ) – . (– . ) ( ) ln −tassets – . (– . ) – . (– . ) ni_margin −t – . (– . ) – . (– . ) −tnoa . ( . ) . ( . ) −toeoi – . (– . ) – . (– . ) −troae . ( . ) . ( . ) −tequity – . (– . ) – . (– . ) −tliquidity . ( . ) – . (– . ) ( ) ln −tassets . ( . ) . ( . ) ni_margin −t . ( . ) . ( . ) −tnoa . ( . ) – . (– . ) −toeoi – . (– . ) – . (– . ) −troae . ( . ) . ( . ) . ( . ) . ( . ) −∆ tliquidity . ( . ) . ( . ) limit points λ . ( . ) . ( . ) . ( . ) . ( . ) λ . ( . ) . ( . ) . ( . ) . ( . ) λ . ( . ) . ( . ) . ( . ) . ( . ) λ . ( . ) . ( . ) . ( . ) . ( . ) λ . ( . ) . ( . ) . ( . ) . ( . ) λ . ( . ) . ( . ) . ( . ) . ( . ) λ . ( . ) . ( . ) . ( . ) . ( . ) fit measures % correct . . . . pseudo r . . . . sbc . . . . lr statistic . [ . ] . [ . ] . [ . ] . [ . ] lr(general→*) na . [ . ] na . [ . ] observations table notes. the dependent variable is a bank’s rating which has eight categories that correspond to the integer values in the range of to and yields seven limit points, ,..., , , =iiλ (the intercept is not separately identified from the limit points). z-statistics (in parentheses) are based upon huber-white standard errors and the percentage of correct predictions (% correct) use the category with the highest probability to give the predicted rating. also reported are the pseudo r , schwartz’s information criterion, sbc, and likelihood ratio tests for the model’s explanatory power, lr statistic, and the deletion of variables from the general model to obtain the parsimonious model, lr(general→*). probability values are given in square parentheses. all regressions were estimated using e-views . and stata . table : bank ratings svms undifferenced specifications differenced specifications variables general model parsimonious model variables general model parsimonious model time – . [– . , – . ] * – . [– . , . ] time – . [– . , . ] −tequity . [– . , . ] −tequity . [ . , . ] * . [ . , . ] * −tliquidity . [– . , . ] −tliquidity – . [– . , . ] ( ) ln −tassets . [ . , . ] * . [ . , . ] * ( ) ln −tassets . [ . , . ] * . [ . , . ] * ni_margin −t . [– . , . ] ni_margin −t – . [– . , . ] −tnoa – . [– . , . ] −tnoa – . [– . , . ] −toeoi – . [– . , – . ] * – . [– . , – . ] * −toeoi – . [– . , – . ] * – . [– . , – . ] * −troae . [ . , . ] * . [ . , . ] * −troae . [ . , . ] * . [– . , . ] −tequity . [– . , . ] −∆ tequity - . [– . , . ] −tliquidity . [ . , . ] * . [ . , . ] * −∆ tliquidity . [– . , . ] ( ) ln −tassets . [– . , . ] ( ) ln −∆ tassets - . [– . , . ] ni_margin −t . [– . , . ] ∆ ni_margin −t . [– . , . ] −tnoa – . [– . , . ] −∆ tnoa . [– . , . ] −toeoi – . [– . , – . ] * – . [– . , – . ] * −∆ toeoi . [– . , . ] −troae . [– . , . ] −∆ troae - . [– . , . ] −tequity . [– . , . ] −∆ tequity - . [– . , . ] −tliquidity – . [– . , – . ] * – . [– . , – . ] * −∆ tliquidity . [ . , . ] * . [– . , . ] ( ) ln −tassets – . [– . , . ] ( ) ln −∆ tassets . [– . , . ] ni_margin −t – . [– . , . ] ∆ ni_margin −t . [– . , . ] −tnoa – . [– . , – . ] * – . [– . , . ] −∆ tnoa . [– . , . ] −toeoi – . [– . , . ] −∆ toeoi . [– . , . ] −troae . [– . , . ] −∆ troae - . [– . , . ] −tequity . [– . , . ] −∆ tequity - . [– . , . ] −tliquidity – . [– . , . ] −∆ tliquidity - . [– . , . ] ( ) ln −tassets . [– . , . ] ( ) ln −∆ tassets - . [– . , . ] ni_margin −t – . [– . , . ] ∆ ni_margin −t . [– . , . ] −tnoa – . [– . , . ] −∆ tnoa - . [– . , . ] −toeoi . [– . , . ] −∆ toeoi - . [– . , . ] −troae . [– . , . ] −∆ troae - . [– . , . ] table notes. the dependent variable is the bank rating (which is never differenced). confidence intervals are reported from a bootstrap ( , samples) in square brackets []. these are % confidence intervals for the undifferenced specifications while % level intervals are employed for the difference specifications. an asterix, * , indicates a variable that is significant according to the reported confidence intervals. models also include dummy variables for countries but these are not shown to save space. table : percentage correct predictions: ordered choice models logit probit general parsimonious general parsimonious with country . % . % . % . % no country . % . % . % . % with country svm . % . % . % . % table notes. the predicted rating for each observation is chosen upon the basis of the category with the highest probability. the row entitled with country refers to the percentage of correct predictions obtained from the models reported in table . the row entitled no country reports the correct predictions for models developed using the general-to-specific method where the country variable is excluded from the general model. the row entitled with country svm denotes the predictive accuracy of ordered choice models that use the country index constructed with the weights obtained from the svm’s general model (reported in table ) and the parsimonious models are developed applying the general-to-specific method. table : percentage correct predictions from svms undifferenced model with unrestricted dummies general parsimonious ε ε . . . . . . c . . % . % . % . % . % . % . . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % differenced model with unrestricted dummies general parsimonious ε ε . . . . . . c . . % . % . % . % . % . % . . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % undifferenced model with no country dummies general parsimonious ε ε . . . . . . c . . % . % . % . % . % . % . . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % undifferenced model with single country index general parsimonious ε ε c . . . . . . . . % . % . % . % . % . % . . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % . % table notes. the statistics are the percentage of correct predictions obtained from svms. all sections except for the second section are based upon svms using undifferenced variables. the section headed undifferenced model with unrestricted dummies refers to svms that allow all country dummies to be estimated unrestricted – the regressions from which the predictions are made are reported in table under the heading undifferenced specifications. the section headed undifferenced model with no country dummies reports predictive accuracy using svms that exclude all country terms. the section headed undifferenced model with single country index provides results based upon svms including the single country index constructed from the ordered choice regressions (the weights are reported in table ) and no individual country dummy variables. the section headed differenced model with unrestricted dummies refers to svms that allow all country dummies to be estimated unrestricted and include differenced variables – the regressions from which the predictions are made are reported in table under the heading differenced specifications. table : undifferenced general svm model’s country index weights country weight country weight usa . kenya . andorra . latvia . united kingdom . macedonia (fyrom) . switzerland . malta . bermuda . netherlands . canada . norway . korea rep. of . philippines . czech republic . san marino . saudi arabia . serbia . france . south africa . slovenia . spain . panama . sweden . united arab emirates . taiwan . kuwait . tunisia . macau . hungary - . oman . albania - . lithuania . malaysia - . austria . poland - . cyprus . indonesia - . mexico . romania - . slovakia . morocco - . bahrain . israel - . qatar . thailand - . el salvador . niger - . trinidad and tobago . peru - . brazil . turkey - . japan . azerbaijan - . germany . kazakhstan - . jordan . russian federation - . georgia rep. of . india - . mongolia . colombia - . armenia . ukraine - . australia . jamaica - . benin . vietnam - . bosnia-herzegovina . venezuela - . bulgaria . nigeria - . chile . argentina - . costa rica . lebanon - . egypt . belarus - . estonia . pakistan - . greece . bangladesh - . hong kong . sri lanka - . iceland . dominican republic - . ireland . iran - . italy . china-people's rep. - . table notes. the estimated weights for the individual country dummy variables are obtained from the svm model where the variables are not differenced and the model includes all possible variables (general model). the other estimation results for this model are reported in table . [pdf] real-time fault diagnosis using knowledge-based expert system | semantic scholar skip to search formskip to main content> semantic scholar's logo search sign increate free account you are currently offline. some features of the site may not work correctly. doi: . /j.psep. . . corpus id: real-time fault diagnosis using knowledge-based expert system @article{nan realtimefd, title={real-time fault diagnosis using knowledge-based expert system}, author={c. nan and f. khan and m. t. iqbal}, journal={process safety and environmental protection}, year={ }, volume={ }, pages={ - } } c. nan, f. khan, m. t. iqbal published engineering process safety and environmental protection abstract abnormal operating conditions (faults) cost process industry billons of dollars per year and can be prevented if they are predicted and controlled in advance. advanced software applications, based on the expert system, has the potential to assist engineers in monitoring, detecting, and diagnosing abnormal conditions and thus providing safe guards against these unexpected process conditions. abnormal operating conditions (faults) could be modeled and predicted with high confidence using… expand view via publisher lue.ethz.ch save to library create alert cite launch research feed share this paper citationshighly influential citations background citations methods citations view all figures and tables from this paper figure table figure table figure table figure figure figure table figure table figure table figure figure figure figure figure figure figure figure figure figure figure figure figure figure figure figure view all figures & tables citations citation type citation type all types cites results cites methods cites background has pdf publication type author more filters more filters filters sort by relevance sort by most influenced papers sort by citation count sort by recency fault diagnosis and cause analysis using fuzzy evidential reasoning approach and dynamic adaptive fuzzy petri nets hu-chen liu, qing-lian lin, ming-lun ren engineering, computer science comput. ind. eng. save alert research feed risk‐based fault diagnosis and safety management for process systems hui, z. bao, f. khan, t. iqbal, yan-jun chang engineering pdf save alert research feed rdr-based knowledge based system to the failure detection in industrial cyber physical systems dohyeong kim, s. han, y. lin, b. kang, s. lee computer science knowl. based syst. save alert research feed a hybrid fault diagnosis method for mechanic-electronic-hydraulic control system based on simulated knowledge from virtual prototyping deyu he, n. hu, l. hu engineering view excerpt, cites methods save alert research feed design of a real-time fault diagnosis expert system for the east cryoplant zhiwei zhou, m. zhuang, xiaofei lu, l. hu, g. xia computer science save alert research feed fta-based fault diagnose expert system for hydraulic equipments mengmeng bian, j. shi, s. wang engineering proceedings of international conference on fluid power and mechatronics pdf view excerpt, cites methods save alert research feed knowledge acquisition development in failure diagnosis analysis as an interactive approach m. yazdi, hamzeh soltanali computer science save alert research feed process fault diagnosis with model- and knowledge-based approaches: advances and opportunities weijun li, h. li, sai gu, tao chen computer science save alert research feed fault detection and isolation in gas turbine engines zakieh sadough engineering highly influenced pdf view excerpts, cites methods save alert research feed a novel framework for real-time fault diagnosis based on dynamic fault tree analysis r. duan, xiudong ou computer science pdf view excerpts, cites background save alert research feed ... ... references showing - of references sort byrelevance most influenced papers recency intelligent system for process supervision and fault diagnosis in dynamic physical systems c. lo, y. k. wong, a. rad computer science, engineering ieee transactions on industrial electronics view excerpts, references background save alert research feed fuzzy rule-based expert system for power system fault diagnosis h. monsef, a. ranjbar, s. jadid computer science view excerpts, references background and methods save alert research feed fault detection and isolation of industrial processes using optimized fuzzy models l. mendonça, j. sousa, j. m. s. d. costa computer science the th ieee international conference on fuzzy systems, . fuzz ' . view excerpts, references methods save alert research feed a neural network methodology for process fault diagnosis v. venkatasubramanian, k. chan engineering view excerpt, references background save alert research feed a fuzzy expert system for fault detection in statistical process control of industrial processes s. m. el-shal, a. s. morris computer science ieee trans. syst. man cybern. part c highly influential view excerpts, references background save alert research feed model-based fault diagnosis via parameter estimation using knowledge base and fuzzy logic approach l. mohamed, a. s. ibrahim engineering th ieee mediterranean electrotechnical conference (ieee cat. no. ch ) view excerpts, references background save alert research feed fuzzy-logic based trend classification for fault diagnosis of chemical processes sourabh dash, r. rengaswamy, v. venkatasubramanian engineering, computer science comput. chem. eng. pdf view excerpts, references background save alert research feed model-based fault detection in information poor plants j. howell engineering, computer science autom. save alert research feed improved on-line process fault diagnosis through information fusion in multiple neural networks j. zhang engineering, computer science comput. chem. eng. view excerpts, references background save alert research feed component-based multi-model approach for fault detection and diagnosis of a centrifugal pump a. wolfram, d. fussel, t. brune, r. isermann engineering proceedings of the american control conference. (cat. no. ch ) view excerpts, references background save alert research feed ... ... related papers abstract figures and tables citations references related papers stay connected with semantic scholar sign up about semantic scholar semantic scholar is a free, ai-powered research tool for scientific literature, based at the allen institute for ai. learn more → resources datasetssupp.aiapiopen corpus organization about usresearchpublishing partnersdata partners   faqcontact proudly built by ai with the help of our collaborators terms of service•privacy policy the allen institute for ai by clicking accept or continuing to use the site, you agree to the terms outlined in our privacy policy, terms of service, and dataset license accept & continue semantic web service discovery using natural language processing techniques jordy sangersa, flavius frasincara, frederik hogenbooma,∗, vadim chepeginb aerasmus university rotterdam, p.o. box , dr, rotterdam, the netherlands btie kinetix, p.o. box , kb, hoofddorp, the netherlands abstract this paper proposes a semantic web service discovery framework for finding semantic web services by making use of natural language processing techniques. the framework allows searching through a set of semantic web services in order to find a match with a user query consisting of keywords. by specify- ing the search goal using keywords, end-users do not need to have knowledge about semantic languages, which makes it easy to express the desired semantic web services. for matching keywords with semantic web service descriptions given in wsmo, techniques like part-of-speech tagging, lemmatization, and word sense disambiguation are used. after determining the senses of relevant words gathered from web service descriptions and the user query, a matching process takes place. the performance evaluation shows that the three proposed matching algorithms are able to effectively perform matching and approximate matching. keywords: natural language processing, semantic web services, wsmo, web service discovery . introduction with the emergence of web services and the service oriented architecture (soa), business process components are more and more being decoupled. using web services in soa creates a wide network of services that collaborate in order to implement complex tasks. currently, web services are commonly described via narrative web pages containing information about their operations in natural languages. these web pages contain plain text with no machine interpretable structure and therefore cannot be used by machines to automatically process the descriptive information about a web service. ∗corresponding author; tel: + ( ) ; fax: + ( ) email addresses: jordysangers@hotmail.com (jordy sangers), frasincar@ese.eur.nl (flavius frasincar), fhogenboom@ese.eur.nl (frederik hogenboom), vadim.chepegin@tieglobal.com (vadim chepegin) preprint submitted to expert systems with applications january , to promote the automation of web service discovery (keller et al., ) and composition (hikimpour et al., ), a number of different semantic languages (martin et al., ; de bruijn et al., ; vitvar et al., ) have been created that allow describing the functionality of services in a machine interpretable form, while original web service descriptions contained only information about the data types and bindings as a description of a web service functionality. the semantic web service descriptions use ontologies to describe the behavior of a web service by applying reasoning over web service semantics. in this way, the semantics described in ontologies enable systems to interpret what a web service is doing, stimulating automatic web service discovery and compo- sition. the ontologies, however, are created by humans and therefore contain natural language. this allows humans to understand the concepts defined, but a system, in contrary to humans, can only understand on- tology concepts and their relationships to a limited extend. natural language processing (nlp) techniques can therefore help in better defining the context of a web service. when using one holistic ontology, machines can discover and compose web services automatically based on the semantics defined. using one holistic ontology is, however, hardly reachable and therefore it is impossible to reason based only on formal logic. nlp techniques can help overcome the ambiguity problems between different ontologies that are being used by semantic web service descriptions. last, service composition should be driven by people who know business processes and not by techni- cians. thus, end users must be able to discover these web services based on keywords written in human language. therefore, a discovery mechanism must be developed in such a way that a bridge between key- words written in a natural language, on one hand, and web service descriptions provided using semantically enhanced languages, on the other hand, can be created. in this paper, the semantic web service discovery (swsd) framework is proposed, which enables end users to search, using keywords, for existing web services described by means of a semantic web language for service annotation. this process consists of several steps including: • extraction of information from semantic descriptions to create the context of a web service; • nlp for disambiguating words’ meanings and establishing a context for a set of words; • matching the users search context with a web service context by means of a similarity measure. the result of this process will be a ranked list of web services that match the users search criteria. this context-based matchmaking mechanism provides flexibility by not only searching for exact word matches, but also by looking for synonyms found in a popular lexical database as wordnet (miller et al., ) and making use of its extensive network of relationships between different words (e.g., synonyms, hypernyms, etc.). this work is based on our previous efforts (sangers et al., ), in which we propose a linguistic approach for semantic web service discovery, yet in our current endeavours we provide more details on the proposed approach and, additionally, we present a genetic algorithm-based method to learn the feature weights. the remainder of this paper is structured as follows. section discusses related technologies for describing and searching web services. next, section proposes the swsd framework for discovery of web services based on keywords. a possible implementation of the framework is described in section . subsequently, section evaluates different matching algorithms between the user input and a web service description. last, section concludes the paper and discusses future work. . related work this section reviews state-of-the-art languages and tools for service discovery. first, a comparison is made between different types of web service description languages in section . . since this paper covers the discovery of semantic web services, three semantic web service description languages are described and compared with each other in section . . last, several approaches for discovering web services are discussed in section . . . . web service description languages because web service discovery depends to a large extent on how web services are described, an overview of the different types of languages used for describing web services is first given. these lan- guages differ in models and formalisms used for describing web services and which web service properties they cover. this range of languages can go from a simple piece of plain text describing the web service, to large and complex semantic descriptions of the web service’s behavior by means of ontologies. figure shows an overview of the most widely used languages for describing a web service. distinction among four description layers is made in order to demonstrate an increasing role of semantics in such languages. each layer can contribute to the process of matching a service with a given query. thus, it is necessary to be able to make use of the information encoded into each layer in order to have a good overview on the capability of a service. figure : web service description languages. to make machines use web services automatically, their interfaces are commonly defined in languages such as web service description layer (wsdl) (christensen et al., ), web application description layer (wadl) (hadley, ), or hrests (kopecký et al., ). these languages describe the bind- ings, the operations and their associated input and output data types, and the end points of a web service. these descriptions enable humans and applications to understand where, when, and what a web service is expecting from the user on a syntactical level. the top layer depicted in figure consists of semantic web service description languages such as web service modeling ontology (wsmo) (de bruijn et al., ) and owl-s (martin et al., ). these lan- guages employ ontologies for describing the behavior of a web service. concepts, attributes, and relations from existing ontologies and logical expressions can be used to state conditions and effects of a web service. in order to provide a bridge between the syntactical languages such as wsdl and hrests to the semantically enriched languages such as wsmo and owl-s, middle layer languages were defined. mi- crowsmo (kopecký et al., ), sa-rest (lathem et al., ), and semantic annotations for wsdl (sawsdl) (farrell and lausen, ) link the concepts from the semantic descriptions with the data types for the input and output of a web service, or with its operations, and provide methods to transform data types to semantic concepts and the other way around. the latter two layers are not widely used yet. therefore most web service discovery systems use only the descriptions written in languages like wsdl. however, since semantics provide more information about web services, they can easily be used for web service discovery. in our view, both types of languages can contribute to the discovery of web services. . . semantic web service description languages table lists the core differences between three semantic web service description languages, owl-s, wsmo and wsmo-lite. they mainly differ in the syntax they have, the different parts they consist of, and which kind of reasoning they use to describe the logic behavior of a web service. owl-s builds on web ontology language (owl) (bechhofer et al., ) and consists of different ontologies to describe a web service. the language is defined in order to enable three major tasks which could not be fulfilled with older technologies (wsdl). these are automatic web service discovery, auto- matic web service invocation, and automatic web service composition and interoperation. because owl-s uses ontologies to describe web services, these services and their behavior become machine interpretable and thus tasks such as discovery and composition can be automated. owl-s makes use of three different ontologies: a service profile, which states what the web service does, a service model, which describe how the web service performs the tasks, and a service grounding, which describes how to access the web service. in recent work, farrag et al. ( ) introduced an algorithm for mapping wsdl descriptions to owl-s descriptions by making use of an ontology-based approach. wsmo is a framework for describing web services and consists of four top-entities: ontologies, web services, goals, and mediators. ontologies provide the terminology used by other wsmo elements. web services describe the capabilities, interfaces, and internal working of web services. goals represent user desires, and mediators provide bridges between different ontologies, web services, or goals to overcome interoperability problems. wsmo uses a specific designed language called wsml (de bruijn, ) and can contain powerful logical formulae to describe the different wsmo elements. it also contains a ground- ing feature to link concepts with wsdl data types so that automatic invocation can be achieved. table : core differences between owl-s, wsmo and wsmo-lite. owl-s wsmo wsmo-lite syntax owl wsml rdf/xml parts service profile, service model, service grounding ontology, web service, goal, mediator ontology, web service reasoning logic language logic and rule language logic language wsmo-lite (vitvar et al., ) was created because of the need for a simple semantic web service de- scription language. it is therefore a lightweight set of semantic service descriptions written in rdfs (brick- ley and guha, ) that can be used for annotations of various wsdl elements using sawsdl annota- tion mechanism. wsmo-lite only makes use of ontologies and web service descriptions and contains no grounding information, which makes it dependent of sawsdl. the behavior of a web service is only im- plicitly described by defining just preconditions and effects, so no specific how-questions can be answered. although these three languages have different properties, they all describe the semantics of a web service. they provide information about the context domain, behavior and usage of a web service. because they use elements from predefined ontologies, those ontologies have to be used together with the semantic web service descriptions for the discovery and matching of the services. . . web service discovery engines we can distinguish between two types of approaches for web service discovery, i.e., discovery based on clustering operation parameters on the one hand, and rich semantics on the other hand. this section continues by elaborating on existing or proposed web service discovery systems for each of these two different approaches. . . . web service discovery based on clustering operation parameters one approach for web service discovery is by searching for similarities among different wsdl service descriptions, enabling searching for substitutable and composable web services as similar operations and services can be discovered based on operation parameters. web service operation semantics can be extracted and employed for discovery purposes. woogle (dong et al., ) is a web service search engine that employs clustering techniques for group- ing operation parameters, and for a given query, it searches for similar and/or composable web service operations. for this, woogle automatically defines the underlying semantics of wsdl descriptions based on the operation parameters and uses these semantics to match operations. however, if independent ontolo- gies which define the web service semantics exist, the behavior of a web service can be known without investigating parameter names and is therefore preferable to use. operation parameter clustering techniques are also employed in seekda! (semantic technology insti- tute, ), which extracts service semantics from wsdl files, enabling runtime exchange of similar and composable services. seekda! is part of the service-finder (cefriel et al., ) framework, which is a platform for service discovery where information about services is gathered from various sources like web pages and blogs. the information is automatically added to a semantic model using automatic service an- notation, realizing flexible discovery of services. service-finder uses service semantics for discovery and composition, but gathers this information dynamically and hence does not take into account predefined semantics. . . . web service discovery using rich semantics another approach for semantic web service discovery is the use of predefined ontologies. by identify- ing semantic similarities between ontologies, related semantic web services can be discovered. to identify semantic similarities, mediation spaces can be used (dietze et al., ). these mediation spaces me- diate on data-level as well as semantic-level for discovery of related semantic web services according to ontologies, other semantic web services, or wsmo goals. godo (gomez et al., ) does not search for similar web services, but instead employs a goal-driven approach. godo has a repository with wsmo goals and analyzes a user-described goal (in natural lan- guage). the wsmo goal with the highest match will be sent to wsmx (deri galway, ), which is an execution environment for wsmo service discovery and composition. the wsmx environment subse- quently searches for a wsmo web service that is linked to the given wsmo goal via wsmo mediators and returns the wsmo web service. although this approach makes good use of the capabilities of the wsmo framework, it cannot be applied for other semantic languages like owl-s and wsmo-lite, as they do not have goal representation elements. bener et al. ( ) proposed an architecture that performs semantic matching of web services based on both input and output descriptions, as well as preconditions and effects. in the semantic web service matchmaking process, owl-s descriptions are assumed (in contrast to the wsmo descriptions which are the focus of our research), and the matchmaking is done at the conceptual level using swrl rules. similar to the previous approach, yang et al. ( ) make use of rules for matching semantic web services. however, in contrast to the work of bener et al. ( ), the authors take into account the service context, which is described by means of a context profile in owl-s. contexts are gathered through a java expert system shell (jess)-enabled context elicitation system featuring an ontology-based context model formally describing and acquiring contextual information. service matching is done based in inputs and outputs and can result in an exact match, a plug-in match, and a subsumed match. luna et al. ( ) proposed bax-set plus, which is a multi-agent taxonomy-based method for categorization, search, and retrieval, of semantic web services described in owl-s. here, user-selected concepts from a taxonomy are matched against concepts contained in owl-s service descriptions. an important difference is that, in contrast to the work presented by the authors, we extract concepts from a natural language query instead of from an existing taxonomy that is browsed by the user. moreover, our work focuses on wsmo service descriptions in wsml, instead of the owl-s specifications focused on by luna et al. ( ). also, for all aforementioned approaches, no deep linguistic analysis, like proposed in our paper, is performed. recently, paulraj and swamynathan ( ) proposed a method for content-based semantic web ser- vice discovery. in contrast to the general approach where user queries are matched against owl-s inputs, outputs, preconditions, and effects (iope), the framework allows users to submit free text as input. this alleviates the restrictions put on user queries in that they must be of the same format as that of the iopes present in owl-s. in their work, nouns are extracted from text that is initially unstructured. these nouns are subsequently used for service discovery, after a disambiguation process that makes use of the word- net lexical database for determining the meaning of the nouns. although there are similarities with the work presented in our paper, it should be noted that, similar to many semantics-based web service discov- ery frameworks, paulraj and swamynathan ( ) focus specifically on owl-s. in contrast, we focus on wsmo and aim for a more universal approach that can be utilized in various semantic web service de- scription languages (with a few adaptations). although both in (paulraj and swamynathan, ) and our work, inputs are initially unstructured, in our current endeavours we provide a means for ranking the results, while paulraj and swamynathan ( ) do not implement any form of result ranking. . semantic web service discovery framework the swsd framework comprises a keyword-based discovery process for searching web services that are described using a semantic language. the search mechanism incorporates nlp techniques in order to establish a match between a user search query and a semantic web service description. logics-based specifications that are defined in the web service descriptions are not taken into consideration, but the definitions of concepts stated in other imported ontologies are exploited. in this way, the framework is able to establish a broader search field by also using related concepts from the ontologies for identifying the context in which the web service is operating. section . covers the global architecture of the swsd framework and a short description of the main components. section . describes how the swsd framework reads semantic web service descriptions and which elements from those descriptions are being used for the discovery. in section . is presented how the context of a semantic web service is established. last, section . describes two different algorithms to match a user query with a semantic web service using the query and service contexts. . . framework architecture as an input, the swsd framework requires a set of web services that are described in semantic lan- guages (e.g., wsmo (de bruijn et al., ), wsmo-lite (vitvar et al., ), or owl-s (martin et al., )). the descriptions are subsequently analyzed, resulting in the extraction of words that could rep- resent the context of the web services (i.e., the names of the operations, and nouns and verbs stated in non-functional descriptions of concepts or conditions). next, the extracted words are disambiguated, as multiple senses can be assigned to the same words. last, the disambiguated words are matched with the disambiguated words from the search query, resulting in a ranked list of web services. figure describes the architecture of the swsd framework. the process is subdivided into three major tasks, i.e., service reading, word sense disambiguation (wsd), and match making. service reading comprises parsing a semantic web service description, extracting names and non-functional descriptions of used concepts. the wsd task subsequently determines the senses of a set of words. last, the match making step determines the similarity between the different sets of senses, which is ultimately used for ranking the analyzed web services. . . semantic web service reader to enable a search engine to look through web service descriptions written in different languages, it has to have several different web service description readers, one for each language. in the case of semantically described web services, the readers must be able to parse a description and extract concepts, attributes, and relations from wsmo, wsmo-lite, owl-s, etc. therefore the first step in the process of searching for semantically described web services is to implement readers for different languages and formats. a semantic web service reader must be able to extract various elements out of a web service description and its used ontologies. in the case of a wsmo web service, names and non-functional descriptions of elements such as the capabilities and their subelements, conditions and effects, help in understanding the context of the web service. thereby, by extracting concepts out of the definedby statement of those figure : swsd architecture. elements and searching their non-functional descriptions in an ontology, the context of the web service can be determined in a more accurate manner than just by using service names. the non-functional descriptions are written in natural language and thus contain a human description of the specified element. by extracting these descriptions, one can thus establish the context of the web service operations. before extracting words from a web service description, the description has to be parsed. different languages can mean different syntaxes and therefore different parsers are needed. for wsmo a wsml parser like wsmo j (eu ist and fit-it, ) can help in this process and for owl-s and wsmo-lite, which are written using an owl (bechhofer et al., ) or rdf (klyne and carroll, ) language, a parser like sesame (openrdf.org, ) or jena (hp labs semantic web, ) can be used. to find useful words for wsd, the element names and non-functional descriptions must be split into different words. in the case of element names, simply splitting the words when a case transition has occurred is enough, since in most cases they are written in “camel notation” (e.g., hotelbookingwebservice). each word in the sentences found in the non-functional descriptions, must be tagged with the right part-of-speech (pos) in order to be useful. using a pos-tagger, nouns and verbs from sentences are extracted and used for the wsd. . . word sense disambiguation a user can represent its goal by defining a sets of words (nouns and verbs). because many words have associated multiple meanings (e.g., mouse can be used to represent either an animal or a computer device), disambiguation of word senses helps in finding the correct context. the disambiguation will be applied to the set of words gathered from the user input and from a semantic web service description, resulting in two sets of disambiguated senses that can be employed for matching. as non-supervised wsd allows disambiguation of words without user interference, we use a variant of the ssi algorithm (navigli and velardi, ) for retrieving the senses out of a set of words, which is defined as follows: selecteds ense(word) = arg max s j∈senses(word) ∑ sci∈i sim(s j, sci) . ( ) the algorithm disambiguates a word (word) based on a previously disambiguated set of words and their related senses. per word sense s j, a similarity with the senses from the context (sci) is calculated, and the sense with the highest similarity is chosen. subsequently, the word and its chosen sense are added to context i and the process is iterated until there are no ambiguous words left. naturally, at the start of the process, a context is not yet established, and hence i = ∅. hence, we fill context i with all monosemous words, i.e., words that have only one sense, as these do not require any disambiguation. subsequently, the iterative process of disambiguating polysemous words can be initiated, always targeting the least ambiguous words (i.e., with the least amount of senses) first. for each of the available senses, the algorithm is simulated as if the sense was used as the starting context. each time a new sense is added to the context, the similarity between the new sense and the context is stored. the sense which creates the highest sum of similarity measures during its simulation run is used for the context initialization. because previous studies have shown that the method of jiang and conrath (jiang and conrath, ) performs better than other semantic distance measures (budanitsky and hirst, ), we employ this method for computing the sense similarities. the measure makes use of the information content (ic), which depends on the probability of the occurrence of a particular sense in a large corpus: ic(sense) = log p(sense) . ( ) formula ( ) gives the similarity between two senses. besides the ic, also the least common subsumer (lcs ) between two concepts is used, representing the first ancestor concept that subsumes both senses when going in a bottom-up direction through a hypernym tree of senses. the ic of the lcs is computed and compared with the ic of the two senses: sim(si, s j) = ic(si) + ic(s j) − × ic(lcs (si, s j)) . ( ) . . sense matching after the disambiguation of all words gathered from the user input and a semantic web service descrip- tion, one is left with several different sets of senses. for the matching process, each word in the user query is assumed to be equally important and hence the user input contains one set of senses. however, a web service description can contain words that better represent the context of the web service than other words. therefore, after the wsd, several sets of senses, each having a different weight for the matching process, are computed for a web service. this section starts with a high level view of the matching between the user input and the different levels of information extracted from a web service description. next, the jaccard matcher for matching sets of words or senses is described. last, a matching approach based on similarities between words or senses is explained. . . . level matching not only the disambiguated senses are matched, but also words that are not appearing in the used lexicon, as these words can represent important names or concepts for the discovery of web services. hence, for matching user input with a semantic web service description, the user input contains a set of ambiguous words wsu and a set of senses ssu. the web service description on the other hand contains multiple sets of words mwsw and multiple sets of senses mssw. because the web service description, presented in the next section, provides n sets containing words and senses, each having a different importance for the matching process, the final similarity between the user input and the web service input is computed as a weighted average of the similarities between each set of words mwswi ∈ mwsw and senses msswi ∈ mssw from the web service description and the set of words and senses from the user input: f inals im(ssu,mssw,wsu,mwsw) = n∑ i= wi × levels im(ssu,msswi,wsu,mwswi) . ( ) here, the weights wi are established through optimization techniques and sum up to in order to make sure that the final similarity between the user query and a web service description has a range between and . for each set of words and senses from the web service description, the system employs two mea- sures, i.e., one for sense matching (defined in formulas ( ) and ( )), and one for the matching of non- disambiguated words (see formulas ( ) and ( )). these measures range between (no match) and (exact match) and are combined into a single measure using a weighted average: levels im(ssu, ssw,wsu,wsw) = wsense × senses im(ssu, ssw) + wword × words im(wsu,wsw) , ( ) which is used n times in formula ( ), where msswi = ssw and mwswi = wsw. the weights corresponding to the sense similarity and the word similarity are, as with the final similarity, established by means of optimization techniques and must sum up to . . . . jaccard matching the jaccard matcher, which is employed for matching sets of words or senses, makes use of the jaccard index (jaccard, ). this method is often used for computing set similarity and thus compares different sense sets: senses im(ssu, ssw) = |ssu ∩ ssw| |ssu ∪ ssw| . ( ) subsequently, we calculate a similarity coefficient by dividing the number of senses appearing in both sets by the total number of senses in both sets. this results in a percentage of exact matching items and can also be applied for matching the words that could not be disambiguated: words im(wsu,wsw) = |wsu ∩ wsw| |wsu ∪ wsw| . ( ) . . . similarity matching normally, for the calculation of similarity values, only perfect matching items are used. in order to additionally take into consideration partial matches, the similarity matcher makes use of a similarity-based approach for matching different sets of senses or non-disambiguated words, expressing close relatedness with values approaching , and non-relatedness with values close to . sense set similarities are calculated using the same similarity function as in wsd. the similarity be- tween the user set of senses ssu and a web service set of senses ssw is computed as follows: senses im(ssu, ssw) = ∑ su∈ssu senses core(su, ssw) |ssu|+ |ssw| + ∑ sw∈ssw senses core(sw, ssu) |ssu|+ |ssw| , ( ) where the average of the similarity between each sense su from the user set of senses, and the web service set of senses is computed. the average of the similarity between each sense sw from the web service set of senses, and the user set of senses is added to that to provide a symmetric match. the similarity between a sense sa and a set of senses ssb is calculated by taking the maximum similarity between the sense and one of the senses sb from the other set, i.e.: senses core(sa, ssb) = max sb∈ssb sensenorm(sa, sb) . ( ) the similarities calculated with formula ( ) range between and infinity, yet we prefer a range between and for quantifying the matching degree. hence, we employ a logarithmic function in order to transform the values of the similarity: sensenorm(sa, sb) = − e−sim(sa,sb) . ( ) here, exact similar senses will have as resulting similarity and a total mismatch between senses will result in . as we are unable to determine the senses of some words, we fall back to non-semantic measurements of similarities for matching the sets of non-disambiguated words. for this, we employ the levenshtein distance metric (levenshtein, ), which calculates the total number of operations that need to be done in order to transform one word to another. the similarity between two sets of words is subsequently calculated similarly to the comparison of two sense sets, yet now the levenshtein distance is applied instead of the similarity function from wsd. we calculate the similarity between the user set of words wsu and a web service set of words wsw as: words im(wsu,wsw) = ∑ wu∈wsu words core(wu,wsw) |wsu|+ |wsw| + ∑ ww∈wsw words core(ww,wsu) |wsu|+ |wsw| , ( ) where the similarity between a word and a set of words, words core(wa,wsb) is computed as: words core(wa,wsb) = max wb∈wsb wordnorm(wa,wb) , ( ) and the levenshtein distance is used in the following way for comparing two words wa and wb: wordnorm(wa,wb) = max( , − × levenshtein(wa,wb) maxlength(wa,wb) ) . ( ) here, maxlength(wa,wb) denotes the number of tokens of the longest word that is being compared. if this formula returns a negative value, which means that the amount of changes that should be made exceed half of the length of the largest word, a value of will be used to indicate a total mismatch. . semantic web service discovery engine this section describes the swsd engine, which is an implementation of the swsd approach and allows users to search for web services on an existing repository by defining a set of keywords. the steps required for this implementation are closely related to the steps stated for the swsd framework stated in the previous section. however, this is one possible implementation of the framework, using specific languages and external libraries to provide a discovery engine. section . introduces the swsd engine as an implementation of the swsd framework. section . describes how the swsd engine reads wsmo descriptions and which elements from those descriptions are being used for the service discovery. how the context of a semantic web service is established is presented in section . . last, section . describes how matching algorithms are used for computing the similarity between the user input and semantic web service information. . . swsd engine at the moment, the swsd engine can only apply a search on web services which are annotated using the wsmo (de bruijn et al., ) framework as wsmo is a powerful framework that can contain very wide information about web services. so, only a wsmo web service reader and a wsmo ontology reader have been implemented, but based on the modularity of the implementation, the engine can be extended with readers that can parse other semantic web service languages. to read the wsmo files, wsmo j (eu ist and fit-it, ) is used. wsmo j is an api and a reference implementation for building semantic web services and semantic business process applications based on wsmo. by using wsmo j, wsmo files can be parsed, read, and written. for the wsd, word- net (miller et al., ) is used through two wordnet api’s, in order to find senses belonging to words. the java wordnet library (jwnl) (walenz and didion, ) is used for the morphological analysis of each word, and the mit java wordnet interface (jwi) (finlayson, ) is subsequently employed for retrieving wordnet synsets. next, jwordnetsim (the university of sheffield, ) is used to calculate the similarity between two wordnet senses because it contains an implementation of the jiang and conrath formula, which we proposed to use in the swsd framework. for the part-of-speech tagging, the stanford figure : result presentation example. parser (the stanford natural language processing group, ) is used. the overall implementation is made in java, due to the availability of external packages written in this language for wsd and wsmo parsing and reading. figure shows the user interface of the swsd engine. the user can fill in a comma-separated list of words representing his goal and click the search button. the system will then check if each word given by the user can be a noun, verb, both noun or verb, or something else. this is needed since the wsd process can only be used for a set of words containing the same pos tag. for the words that do not have a clear pos, the user will be asked to select the appropriate pos to use for the search. after that, the system will do the search and propose a list of web services, which are ranked by their similarity with the user input. each item in the list contains the name of the web service, its non-functional description, its similarity score and a normalized similarity score that has a range from to (computed by dividing the similarity score with the highest similarity score). namespace { _"http://api.google.com/googlesearch#", dc _"http://purl.org/dc/elements/ . #" } webservice googlesearchwebservice nonfunctionalproperties dc#description hasvalue "web service that searches for web sites containing words that matches given search words." endnonfunctionalproperties importsontology _"http://api.google.com/http_api.google.com_googlesearch#googleontology.wsml" capability dosearch postcondition nonfunctionalproperties dc#description hasvalue "if there is an input with at least a query, the output must be a googlesearchresult. some values in the input, must also be returned in the output." endnonfunctionalproperties definedby ?input[hasquery hasvalue ?query] memberof googlesearchparameterset and ?query memberof query implies exists ?output (?output memberof googlesearchresult and ?output[hassearchquery hasvalue ?query]) . name and non-functional description of the web service are extracted. nouns and verbs are used for wsd. identifier of the ontology where concepts are defined. non-functional descriptions of properties of the capabilities are extracted. nouns and verbs are used for wsd. concepts uses from ontologies are extracted and searched in their ontologies. figure : a wsmo web service information extraction example. namespace { _"http://api.google.com/googlesearch#", dc _"http://purl.org/dc/elements/ . #" } ontology simplegoogleontology concept googlesearchresult nonfunctionalproperties dc#description hasvalue "result of a google search method." endnonfunctionalproperties hassearchquery oftype ( ) query nfp dc#description hasvalue "the query for the search request." endnfp concept query hastext oftype ( ) _string nfp dc#description hasvalue "string containing keywords representing the search query." endnfp name and non-functional description of the concept are extracted. nouns and verbs are used for wsd. name and non-functional description of the attribute are extracted. nouns and verbs are used for wsd. concepts related via a property are also searched in the ontology. figure : a wsmo ontology information extraction example. . . semantic web service reader for reading wsmo files, an implementation of the wsmo j api is used, resulting in an engine that can extract different types of information from wsmo web services and wsmo ontologies. because wsmo web services and ontologies use different structures, two different readers must be used. figure shows an example wsmo web service describing the google search web service. the parts surrounded by rectangles are extracted by the system to establish the context of the web service. first the name and non-functional description of the web service are analyzed for extracting relevant words (nouns/verbs). then the words from the non-functional descriptions of the properties of the capabilities are extracted and the reader will search in the logical formulae, written after each definedby statement, for concepts used from external ontologies, which are stated after the importsontology statement. after reading the wsmo web service file, the found concepts are used to search in ontologies to find their description. figure shows an example wsmo ontology containing some concepts used in the web service description from figure . based on a full identifier, the reader can search for a concept. if a concept is found, the non-functional definition, attributes and related concepts can be used for wsd. in total the engine creates seven different levels of information about the web service. each of these levels has a different amount of importance to the matching process. those amounts are expressed in weights, summing up to . the different levels and their associated weights are: • non-functional description and name of the web service, . , (direct relation, from web service); • non-functional descriptions of properties of capabilities of the web service, . , (direct relation, from web service); • non-functional descriptions and names of concepts used by web service, . , (direct relation, from ontology); • non-functional descriptions and names of attributes of concepts used by the web service, . , (indirect relation, from ontology); • non-functional descriptions and names of concepts related via attributes with concepts used by the web service, . , (indirect relation, from ontology); • non-functional descriptions and names of superconcepts of the concepts used by the web service, . , (semi-direct relation, from ontology); • non-functional descriptions and names of subconcepts of the concepts used by the web service, . , (semi-direct relation, from ontology). the weights are established by making use of a genetic algorithm, which is a specific technique to find approximate solutions for optimization problems. in our case, we want to find a set of weights which opti- mize the search results. genetic algorithms mimic an evolution of individuals in a certain population. each individual is represented with a chromosome consisting of data that can be recombined and mutated during next generations. we choose to represent the individuals using a chromosome consisting of nine variables, the seven weights used for stating the importance of the information gathered from a web service, and two weights for setting the trade-off between the sense similarity and the lexical representation similarity. from each generation, the best individuals will survive and create an offspring. to define which indi- viduals are good enough to survive, a fitness function is used. in our case, this is a function that takes the average precision values for each query used in our two data sets described in section . last, the genetic algorithm has been executed with a population size of and the number of generations set to . as can be seen from the weights, the information directly related to the web service is the most im- portant information for the matching process. this information consists of the name and non-functional descriptions of the web service, its capabilities, and concepts used. less important for the matching are the non-functional descriptions and names of the super- and subconcepts of the used concepts by the web service. least important and thus given a low weight is the information that is indirectly related to the web service description. this information is represented by attributes and concepts related to the concepts used by the web service. using these weights, the information that is semantically closest to the core web service description will have the most impact for the matching process. the names and non-functional descriptions of the entities returned from these two readers will then go through a nlp step. nouns and verbs are extracted from the non-functional descriptions using the stanford pos-tagger and words are split if they consist of case-transitions. a sentence like the non-functional description of the google search web service presented in figure (“web service that searches for web sites containing words that matches given search words.”) will generate a set of nouns {web, service, sites, words} and verbs {searches, containing, matches, given}. the name of the google search web service, googlesearchwebservice, will be split into the set of nouns {google, web, service} and verbs {search}. so instead of using compound words and whole sentences, the system only uses nouns and verbs extracted from them (as only these can be found in wordnet). . . word sense disambiguation for wsd the system makes use of two different api’s of wordnet. by using them, words can be found in wordnet by giving a lexical representation of the word and its pos. these words will have different senses and each sense has its own identifier, a wordnet synset. the system has to find the synsets based on a set of lexical representations of words employed by users or a web service description. for establishing a starting context, wordnet will be used to find monosemous words. if no monosemous word is found, the word with the least synsets will be used to simulate the best starting context. for each synset, the similarity of the other words will be computed as if this synset was the starting context. each time a new synset is added to the context, the similarity between the new synset and the context is stored. the synset which creates the highest sum of similarity measures during its simulation is used as the real starting context. once having a context, the ssi algorithm will be used to find the synsets of the other words. the similarity is measured using an implementation of the jiang and conrath method (greenwood, ). it can handle pairs of wordnet synsets as input and will return the similarity between them. for each word, the synset with the highest similarity to the context, will be added to the context. . . sense matching the synsets resulting from the disambiguation process must be matched in order to get a final similarity measure. because the information in a web service description has different levels of importance in the matching process, several sets of synsets belonging to a web service will come out after the disambiguation phase. thus one set of synsets coming from the user input must be matched with several sets of synsets coming from a web service. each set from the web service will have a weight representing the value of its information for the matching process. for example, the synsets found from the non-functional description of the web service are more important during the matching, than the names of related concepts. these weights are determined using a genetic algorithm and must sum up to in order to get a final match with the range [ . . . ]. to overcome the fact that words that are not present in wordnet are not used in the matching process, every lexical representation of the extracted words from the user input, which are not in wordnet, is being compared to the lexical representations of the extracted words from a web service description that are not in wordnet. to provide a flexible matching, the words that are not in wordnet are first being lemmatized before they are matched. this means that for example the word searching will also match the word searched. by combining the synset similarity value and the lexical representation similarity, using a weighted average, we determine a final similarity value. the weights have been optimized by the genetic algorithm described in section . and set to / for the synset similarity and / for the lexical representation similarity. the value of the synset similarity is higher than the lexical representation similarity as they provide for a richer information comparison. . evaluation this section covers the evaluation of different matching algorithms that can be used for semantic web service discovery. the algorithms described in section . are implemented in the swsd engine and are evaluated by using a set of predefined queries and sets of preferred web services related to one of the queries. section . explains how the testing is done. in section . the results of those tests are presented. section . concludes which algorithms provides the best matching for discovery of semantic web services. . . experimental setup for testing the algorithms provided in section . , we make use of a repository of wsmo web ser- vice descriptions stored in wsml format. as for wsmo web service descriptions there are no existing data sets readily available such as for instance the owls-tc . owl-s service retrieval test collection (klusch et al., ), we have manually built wsmo service descriptions. for evaluation purposes, we have defined test queries in order to analyze the outcomes of each of the algorithms. these queries represent possible sets of keywords a user could use when searching for a web service. for each query, we have defined a list of preferred web services (present in the repository of the swsd engine) that should be returned with a high ranking, as they all are significantly related to the queries. besides the algorithms described in section . , a simple matching algorithm that does not make use of natural language processing is added for the evaluation. the algorithm employs the jaccard matching algorithm only for lexical representations of the query keywords and the extracted words. the algorithm makes use of the different information levels of a web service and can therefore be used for testing whether there is an added value in using nlp in the process of semantic web service discovery. the testing is done with usage of precision and recall metrics. for every query, the precision and recall values for each of the algorithms are computed according to the list of web services they provide using that particularly query as an input. this list is a ranked list, based on the similarity values calculated by the used matching algorithm, of all the web services that are in the repository and will be compared with the list of preferred web services stated for the query. the precision and recall values are computed by traversing the provided list of web services, according to a query and matching algorithm, from top to bottom. if a web service y j is stated as preferred, then the web service is classified as . if it is not preferred, the web service is classified as . using this classification, the precision pi and recall ri for a position i in the list of provided web services can be calculated as follows, where k is the number of preferred web services: pi = ∑i j= y j i , ( ) ri = ∑i j= y j k . ( ) figure shows the calculation of precision and recall for a given list of classified web services. for each preferred web service – classified as and visualized with a blue dot – the precision and recall values are calculated. each time a web service that is not stated as preferred (classified as and visualized with a yellow dot) is found, the precision will drop. the pr-graph can show when it takes a long time before another preferred web service is found in the list. this is the case when the precision drops heavily. . . experimental results to test the performances of the three matching algorithms, we divide our predefined queries into two types. in total, queries are used for measuring the matching performance of the algorithms for searches for web services that are present in the repository. the remaining queries are used for measuring the performance of queries for web services that are not present in the repository. in the latter set of queries, a number of similar web services from the repository are defined in order to test performance of similar web service discovery when a specified service does not exist in the repository. both sets of queries are divided into two parts, a training set and a test set. the training set, consisting of queries, is used for training the weights that are used by the matching algorithms, while the test set, having queries, is used for testing the performance of the matching algorithms. testing with queries and three matching algorithms generates pr-graphs. visualizing this amount of graphs is not very insightful and hence we create pr-graphs consisting of average precision values for the recall points. this enables us to relate all the different algorithms at once. however, the testing is done with figure : precision and recall calculation example. lists of preferred web services that can vary in the number of web services they consists of. for testing, lists that contain two to five preferred web services have been used. because these variations in number of web services cause different recall values, average precision values was only be calculated for queries that have the same amount of preferred web services. hence, we analyze eight different pr-graphs that visualize the performances of the different match- ing algorithms. the four pr-graphs that are depicted in figure , show the average results for the exact matching tests. for each of the four different numbers of preferred returned web services (n) a pr-graph is created. the four pr-graphs that are shown in figure , show the average results for the approximate matching tests. from the different pr-graphs that are shown in figure , we can make two observations. first, in most cases, the jaccard algorithm shows a higher precision for most recall values than the simple and the similarity algorithm. second, all algorithms require about the same precision for providing a full recall. this means that in order to provide all the preferred web services to the user, the algorithms are required to display about the same amount of web services. however, according to the fact that the jaccard algorithm provides a higher precision for a lower recall, the jaccard algorithm provides at least some of the preferred web services in an earlier stage to the user than the others. it can therefore be seen as the best algorithm to discover exact matching web services. from the different pr-graphs that are shown in figure , we can make the observation that the similarity algorithm performs overall better for discovery of similar web services than the jaccard and the simple matching algorithm, as in most of the cases the pr-graph lines of the similarity algorithm are above the lines of the jaccard algorithm and the simple matching algorithm. . . summary in order to test the performance of the three matching algorithms explained in section . , we have performed tests, of which tests were done to measure the performance of the algorithms according to discovery of exact matching web services, and tests were done to measure the performance of the algorithms according to discovery of similar web services. each test consisted of a set of keywords used as a query and a set of web services that were preferred to be provided to the user as early as possible. based on our evaluation, we conclude that the jaccard matching algorithm is the best method for discovering exact matching web services, whereas the similarity matching algorithm is the best method for discovering similar web services. n= n= n= n= figure : pr-graphs for discovery of exact matching services. n= n= n= n= figure : pr-graphs for discovery of similar services. . conclusions and future work the swsd framework proposes a keyword-based discovery process for searching web services that are described using semantically enriched annotations done by means of semantic languages for service description. it makes an intensive use of natural language processing techniques and a wordnet-based similarity measure for matching keywords. by using an approach with two different similarity functions, one for lexical similarities and one for semantic similarities, standard lexical matching algorithms as well as semantic matching algorithms, using ontologies, can be applied for discovery of semantic web services. the swsd engine can search for wsmo web services based on user search keywords. a matching score is computed based on the similarity between the words in the user query and a web service description. experiments have been done to test the performance of three different matching algorithms. the jaccard matching algorithm performs best for discovering exact matching web services, while matching using a similarity approach gives the best results for finding similar web services. as a future work, the swsd engine could be extended in such a way that it has the ability to read more annotation formats, e.g., wsmo-lite. in this case, not only wsmo web services can be discovered, but also web services described using other semantic languages than wsmo. also, additional information about a web service (e.g., signature of its operations, messages being exchanged, etc.) could be retrieved from its wsml specification. with this information, the context of the web service can be described in more detail. references bechhofer, s., van harmelen, f., hendler, j., horrocks, i., mcguinness, d. l., patel-schneider, p. f., stein, l. a., . owl web ontology language reference – w c recommendation february . from: http://www.w .org/tr/owl-ref/. bener, a. b., ozadali, v., ilhan, e. s., . semantic matchmaker with precondition and effect matching using swrl. expert systems with applications ( ), – . brickley, d., guha, r., . rdf vocabulary description language . : rdf schema – w c recommendation february . from: http://www.w .org/tr/rdf-schema/. budanitsky, a., hirst, g., . semantic distance in wordnet: an experimental, application-oriented evaluation of five mea- sures. in: workshop on wordnet and other lexical resources at nd meeting of the north american chapter of the association for computational linguistics (naacl ). association for computational linguistics, pp. – . cefriel, seekda!, ontoprise, university of sheffield, . service-finder. from: http://www.service-finder.eu/. christensen, e., curbera, f., meredith, g., weerawarana, s., . web services description language (wsdl) . – w c note march . from: http://www.w .org/tr/wsdl. de bruijn, j., . d the wsml specification – wsml working draft - - . from: http://www.wsmo.org/tr/ d /. de bruijn, j., bussler, c., domingue, j., fensel, d., hepp, m., keller, u., kifer, m., konig-ries, b., kopecky, j., lara, r., lausen, h., oren, e., polleres, a., roman, d., scicluna, j., stollberg, m., . web service modeling ontology (wsmo) – w c member submission june . from: http://www.w .org/submission/wsmo/. deri galway, . web service execution environment. from: http://www.wsmx.org/. dietze, s., gugliotta, a., domingue, j., . exploiting metrics for similarity-based semantic web service discovery. in: ieee th international conference on web services (icws ). ieee computer society, pp. – . dong, x., halevy, a., madhavan, j., nemes, e., zhang, j., . similarity search for web services. in: th international conference on very large data bases (vldb ). vol. . pp. – . eu ist, fit-it, . wsmo j api. from: http://wsmo j.sourceforge.net/. farrag, t. a., saleh, a. i., ali, h. a., . towards swss discovery: mapping from wsdl to owl-s based on on- tology search and standardization engine. ieee transactions on knowledge and data engineering. to appear (doi: . /tkde. . ). farrell, j., lausen, h., . semantic annotations for wsdl and xml schema – w c recommendation august . from: http://www.w .org/tr/sawsdl. finlayson, m., . jwi: the mit java wordnet interface. from: http://projects.csail.mit.edu/jwi/. gomez, j. m., rico, m., garcia-sanchez, f., bejar, r. m., bussler, c., . godo: goal driven orchestration for semantic web services. in: st workshop on web services modeling ontology implementations (wiw ). vol. . ceur workshop proceedings. greenwood, m., . jwordnetsim. from: http://nlp.shef.ac.uk/result/software.html. hadley, m. j., . web application description language – w c member submission august . from: http://www. w .org/submission/wadl/. hikimpour, f., sell, d., cabral, l., domingue, j., motta, e., . semantic web service composition in irs-iii: the structured approach. in: th ieee international conference on e-commerce technology (cec ). ieee computer society, pp. – . hp labs semantic web, . jena. from: http://jena.sourceforge.net/. jaccard, p., . étude comparative de la distribution florale dans une portion des alpes et des jura. bulletin de la société vaudoise des sciences naturelles , – . jiang, j., conrath, d., . semantic similarity based on corpus statistics and lexical taxonomy. in: international conference research on computational linguistics (rocling x). pp. – . keller, u., lara, r., lause, h., polleres, a., predoiu, l., toma, i., . semantic web service discovery – wsmx working draft - october , . from: http://www.wsmo.org/tr/d /v . /d .pdf. klusch, m., kapahnke, p., fries, b., khalid, m. a., vasileski, m., . owls-tc . – release september . from: http://semwebcentral.org/projects/owls-tc/. klyne, g., carroll, j. j., . resource description framework (rdf): concepts and abstract syntax – w c recommendation february . from: http://www.w .org/tr/rdf-concepts/. kopecký, j., vitvar, t., fensel, d., gomadam, k., . d v . hrests & microwsmo – cms wg working draft march . from: http://cms-wg.sti .org/tr/d /v . /. lathem, j., gomadam, k., sheth, a. p., . sa-rest and (s)mashups: adding semantics to restful services. in: st interna- tional conference on semantic computing (icsc ). ieee computer society, pp. – . levenshtein, v. i., . binary codes capable of correction deletions, insertions, and reversals. soviet physics doklady ( ), – . luna, j. a. g., pardo, i. d. t., builes, j. a. j., . recent progress in data engineering and internet technology. vol. of lecture notes in electrical engineering. springer, ch. bax-set plus: a taxonomic navigation model to categorize, search and retrieve semantic web services, pp. – . martin, d., burstein, m., hobbs, j., lassila, o., mcdermott, d., mcilraith, s., narayanan, s., paolucci, m., parsia, b., payne, t., sirin, e., srinivasan, n., sycara, k., . owl-s: semantic markup for web services – w c member submission november . from: http://www.w .org/submission/owl-s/. miller, g. a., beckwith, r., fellbaum, c., gross, d., miller, k., . introduction to wordnet: an on-line lexical database. international journal of lexicography ( ), – . navigli, r., velardi, p., . structural semantic interconnections: a knowledge-based approach to word sense disambiguation. in: ieee transactions on pattern analysis and machine intelligence. vol. . ieee computer society, pp. – . openrdf.org, . sesame. from: http://www.openrdf.org/. paulraj, d., swamynathan, s., . advanced computing, networking and security. vol. of lecture notes in computer science. springer, ch. content based service discovery in semantic web services using wordnet, pp. – . sangers, j., frasincar, f., hogenboom, f., hogenboom, a., chepegin, v., . a linguistic approach for semantic web service discovery. in: casillas, j., martı́nez-lópez, f. j., corchado, j. m. (eds.), first international symposium on management intelligent systems (is-mis ). vol. of advances in intelligent systems and computing. springer, pp. – . semantic technology institute, . seekda! from: http://seekda.com/. the stanford natural language processing group, . stanford log-linear part-of-speech tagger. from: http://nlp. stanford.edu/software/tagger.shtml. the university of sheffield, . pure java wordnet similarity library. from: http://nlp.shef.ac.uk/result/software. html. vitvar, t., kopecky, j., fensel, d., . wsmo-lite: lightweight semantic descriptions for services on the web. in: th ieee european conference on web services (ecows ). ieee computer society, pp. – . walenz, b., didion, j., . jwnl: java wordnet library. from: http://sourceforge.net/projects/jwordnet/. yang, s. j. h., zhang, j., chen, i. y. l., . a jess-enabled context elicitation system for providing context-aware web services. expert systems with applications ( ), – . knowledge acquisition for fuzzy expert systems knowledge acquisition for fuzzy expert systems gwo-jen hwang computer center, national chiao tung university, hsinchu, , taiwan, republic of china expert systems have been successfully applied to a wide variety of application domains. to achieve better performance, researchers have tried to employ fuzzy logic to the development of expert systems. however, as fuzzy rules and membership functions are difficult to define, most of the existing tools and environments for expert systems do not support fuzzy representation and reasoning. thus, it is time-consuming to develop fuzzy expert systems. in this article we propose a new approach to elicit expertise and to generate knowledge bases for fuzzy expert systems. a knowledge acquisition system based upon the approach is also presented, which can help knowledge engineers to create, adjust, debug, and execute fuzzy expert systems. some control techniques are employed in the knowledge acquisition system so that the concepts of fuzzy logic could be directly applied to conventional expert system shells; moreover, a graphic user interface is provided to facilitate the adjustment of membership functions and the display of outputs. the knowledge acquisition system has been integrated with a popular expert system shell, clips, to offer a complete development environment for knowledge engineers. with the help of this environment, the development of fuzzy expert systems becomes much more convenient and efficient. john wiley & sons, inc. i. introduction in recent years, expert systems have been applied to a wide variety of application domains, which not only showed the utility of this approach, but also revealed some problems of applying it. part of the problems in employing conventional expert systems may be due to the use of binary logic to express domain knowledge, whose expressions may not be natural or straightforward and, hence, the expertise is not completely represented. for example, there are only two possible sides (true or false) in binary logic and a clear boundary must be indicated for making choices, while in the real world, it is difficult to judge things to be absolutely true or false. the problem becomes more compli- cated if linguistic hedges (e.g., very, gradually, moderately, m o r e or less, etc.) are used in expressing expertise.’ a good solution to this problem is to employ the concept of f u z z y s e t and fuzzy l o g i c in expert systems; that is, fuzzy expert systems. for example, let e represent the set of “the people who are tall.” in conventional binary logic, international journal of intelligent systems, vol. , - ( ) john wiley & sons, inc. ccc - /y / - hwang we must clearly specify whether an element belongs to e or not. however, in fuzzy logic, each element belongs to the set with a degree of membership. a person of -cm height belongs to e with the degree of membership . ; while the persons of -cm, -cm, and -cm heights may have degrees of membership . , . and .o, respectively. moreover, when considering the linguistic hedges (e.g., very, gradually, moderately, inore or l e s s , etc.), all we need to do is to compute the degrees of memberships according to some hedge functions. for example, let the fu,,(x) = x', where x is degree of memberships without qualifier. if a person is said t o be t a l l with degree . , according to f v e v ( x ) , we conclude that the person is very t a l l with degree . . however, the development of a fuzzy expert system is time-consuming. the difficulty may be due to the lack of methodologies and tools to help in eliciting knowledge for constructing fuzzy rules and in defining the membership functions for fuzzy values. moreover, most existing expert system shells do not support fuzzy reasoning, which increases the cost needed for developing fuzzy expert systems. t o cope with these problems, we propose a knowledge acquisition system that not only elicits knowledge from domain experts, but also generates knowledge bases for fuzzy expert systems. we have integrated the environment with a popular expert system shell, clips, to enable fuzzy reasoning. some experimental results showed that the approach is desirable and is worth further study. in this article, we propose a knowledge acquisition environment that con- sists of a knowledge acquisition unit, a graphic membership function developer, a fuzzy interface for expert systems shells, and a set of control rules for fuzzy reasoning. the environment has been integrated with a well-known expert system shell to perform fuzzy reasoning. finally, some test results of an experi- ment on a practical medical diagnosis problem are shown to evaluate the new approach, and a conclusion is given. . knowledge acquisition for fuzzy expert systems one critical bottleneck to building expert systems is the elicitation of knowl- edge from domain experts. many knowledge engineering tools (e.g., ets,? n ~ o e t s , ~ aquinas, mole,' nicod,' iconkat,' k s s , ask, smore,'o alt," rulecon,'* kitten,i etc.) and interviewing technique^,'^,'^ have been developed to cope with this problem. in this article, we propose a knowledge acquisition environment, knowledge acquisition for fuzzy expert systems (kafes), which is capable of eliciting expertise from domain experts and generating knowledge bases for fuzzy expert systems. the environment is integrated with a well-known expert system shell, clips, to enable fuzzy reasoning. it consists of four components: ( ) knowledge acquisition unit: a revised repertory grid mechanismi is used t o ( ) graphic membership function developer: a tool to help knowledge engineers enable the elicitation of fuzzy knowledge. adjust membership functions of fuzzy values. knowledge acquisition domain knowledge engineers figure . the structure of kafes. ( ) fuzzy interface for expert system shells: the interface performs the fuzzification and defuzzification operations, which facilitate fuzzy reasoning on conventional expert system shells. currently, it is available for clips, a shell for the develop- ment and delivery of expert systems developed by the software technology branch of the information system directorate at nasa/johnson space center. ( ) a set of control rules for clips, which enables fuzzy reasoning on clips. the whole structure of the environment is depicted in figure . a. knowledge acquisition unit many existing systems employ the repertory grid test, which was originally invented and widely used by george kelly in clinical psychology when inter- viewing experts. in a repertory grid, a set of elements is listed across the top of the grid. various combinations of three elements are then presented t o the expert, and the question, “think of an important trait or characteristic, which makes two of the elements different from the third,” is asked. the trait given by the expert is listed on the left-hand side of the grid, and its opposite is listed on the right-hand side. each pair of a trait and its opposite is called a construct. after the set of constructs is ready, the expert is asked t o fill the grid with ratings. a -+ rating mechanism is often used in filling the grid; i.e., each rating is an integer ranging from - . a is at the left-hand pole, a is a t the right-hand pole, while the rest fall between the poles. an example of a repertory c j c ; c c ; h w a n g figure . an example of a repertory grid. grid with five elements (el, e , e,, e , and e,) and four constructs ( c l , c , c , and c,) is given in figure . each rating expresses the relationship between a construct and an element, which does not provide sufficient information for building fuzzy expert systems. consider the following repertory grid: dangue measles . . . fever low . . . high . . . . . . . . . from the ratings of the grid, we are only aware of the information, “for dangue fever, a patient is likely to have a high temperature,” and, “for measles, a patient’s temperature is usually not high.” it is quite possible that the expert wants to emphasize the expressions by adding some linguistic hedges: “ f o r dangue fever, a patient’s temperature should be very high.” “ f o r measles, a patient’s temperature is usually m o r e or less low.” from the examples, we find it difficult t o apply conventional repertory grids to the development of fuzzy expert systems, which is d u e to the following reasons: ( ) repertory grids do not explicitly hold information for linguistic hedges. ( ) repertory grids do not hold information for helping the determination of mem- bership functions. thus, it would be more desirable to adapt the grid so that it is appreciate for fuzzy logic instead of using fuzzy logic with repertory grids. to cope with these problems, we propose thefuzzy table to hold the information for generating fuzzy rules and membership functions. the procedure for constructing a fuzzy table is similar to that of a repertory grid, which consists of three steps shown as follows: step : elicit all the elements (candidate solutions o r actions) from the knowledge acquisition expert. the elements (e.g., g , , gz, g,, g , and g,) provided by the expert are placed across the top of the table: step : elicit attributes (fuzzy variables) from the expert. each time, the expert is asked for an attribute so that three elements can be partially or fully distin- guished. the fuzzy variables are listed on the left-hand side of the table. for each fuzzy variable, three fuzzy values are specified and listed on the right- hand side according to the following procedure: kafes: please select a set of fuzzy values for fuzzy variable v,: . low/ middle/high. . short/middle/tall. . lightinormali heavy, . small/middle/big. . other (user-defined). expert: . kafes: please select a set of fuzzy values for fuzzy variable v,: . low/ middleihigh. . shortimiddleitall. . light/normal/ heavy. . small/middle/big. . other (user-defined). expert: . kafes: please select a set of fuzzy values for fuzzy variable v,: . low/ middle/high. . shortimiddleitall. . light/normal/ heavy. . small/middle/big. . other (user-defined). expert: . kafes: please indicate the lower bound of the fuzzy values. expert: slow. kafes: please indicate the middle of the fuzzy values. expert: normal. kafes: please indicate the upper bound of the fuzzy values. expert: fast. an example of a fuzzy table with fuzzy variable and values is given as follows: ( , ( , g g g. vl v v v low ; middle ; high short ; middle ; tall slow ; normal ; fast low ; middle ; high step : fill all of the [element, attribute] entries of the grid. each entry consists of two parts: a rating to indicate the most desirable fuzzy value and the degree of certainty for giving that rating. it is possible that the expert selects the most desirable fuzzy value, but still lacks confidence about his (or her) choice. a -scale rating mechanism is employed in the fuzzy table. the ratings are integers ranging from - to + , which represent the possible combinations hwang of linguistic hedges and fuzzy values. consider the ratings of fuzzy variable v, in the table; means v e r y high, means high, means more or less high, means middle, - i means more or less low, - means low, and - means very low. the degree of certainty for each rating is described by “s”or “ n ” , which represent verysureand notverysure, respectively. for example: gi g g g gs s oin - s s s low ; middle ; high s /s s s o / s short ;middle ;tall v s s s ln s slow ; normal ; fast v, - l l n ‘ s - s - s s low ; middle ; high vl v each column of the fuzzy table will be translated to a fuzzy rule. the truth value for the output of i - th rule is computed by x . + . . # of “s” truth = ( # of “s,! + # of $“,i) for example, the first column of the above fuzzy table is translated to the following rule: i f v, i s v e r y h i g h , and v i s tall, and v, i s fast, and v, is more or less low then i n s e r t n e w f a c t g , truth= . . initially, experts are asked to provide numerical information for fuzzy val- ues, and some default functions are employed. let fuzzy values of afuzzy variable be represented as ls (left side), ms (middle side), and rs (right side). assuming that the numerical values with membership degrees .o for ls, ms, and rs are a, p , and y, respectively, we have the following membership function^:'^ knowledge acquisition , low middle high v figure . graphic representation of fuzzy values for v , . for v, in the above table, if the expert gives numerical value for fuzzy value low (ls), for middle (ms), and for high (rs), membership functions given in figure are derived. the degrees of memberships of linguistic hedges are computed by the follow- ing hedge functions: . f(x),or = - x . f(x)very = x - xo. . f(x)more-or-less - where x is the degree of memberships without considering linguistic hedges. b. graphic membership function developer graphic membership function developer is used to help knowledge engi- neers to create and adjust membership functions for input facts. it provides a graphic view, so that knowledge engineers can adjust the shape of membership function curves directly to get the desirable results. hwang defuuificanon figure . fuzzy reasoning by applying the fuzzy interface. c. fuzzy interface the knowledge representation and inference process in a fuzzy expert system is quite different from those in a conventional expert system. however, as there already exist many expert system shells and knowledge engineering tools for conventional expert systems, it would be very convenient if we could build fuzzy expert systems based upon the existing shells or tools. we selected one of the most popular expert system shells, c language integrated production system (clips) to work with the knowledge acquisition system. clips is a tool for the development and delivery of expert systems developed by the software technology branch of the information system directorate at nasa/ johnson space center. the fuzzy interface is capable of accepting input values from users, per- forming fuzzification operations, invoking the inference of clips, and perform- ing defuzzification operations on the outputs from clips. if the knowledge engineer or the expert is not satisfied with the system output, the graphic developer is helpful in adjusting the membership functions to achieve better performance. the process of fuzzy reasoning by applying the fuzzy interface is shown in figure . after performing the inference process, all the system outputs are collected by the fuzzy interface to make the final conclusion. the centroid defuzzification is employed to derive a more reasonable result, which will be shown by graphic representation as well as text display to facilitate the modification and debugging of the fuzzy expert systems. d. control rules for fuzzy reasoning a significant difference between fuzzy rules and conventional rules is the use of linguistic hedges, such as not, v e r y , not v e r y , m o r e o r l e s s , not m o r e o r l e s s , etc. for example, two rules with truth value . are given as follows: knowledge acquisition ( d e f r u l e domain-rule, (level ? gg) ? f f l < - (temp i s h i g h ? d d l & : (> ? d d l ?gg) ) ? f f < - (fan i s not fast ?dd &: (> ?dd ?gg) ) => ( a s s e r t (del-fact fan i s fast) ) ( b i n d ?dd (min ? d d l ?dd ) ) ( p r i n t o u t t 'fan i s fast ' ? d d c r l f ) ( a s s e r t (new-fact fan i s fast ? d d ) ) ) ( d e f r u l e domain-rule, (level ?gg) ? f f l < - (temp i s very h i g h ? d d l & : (> ? d d l ?gg) ) ? f f < - (fan i s not fast ?ddz&: (> ?dd ?gg) ) => ( a s s e r t (del-fact fan i s fast) ) ( b i n d ?dd (min ? d d l ?ddz) ) ( p r i n t o u t t 'fan i s very fast ' ?dd c r l f ) ( a s s e r t (new-factfan isveryfast ? d d ) ) ) for an input value (e.g., "c), the fuzzy interface first performs fuzzifica- tion operations to produce the corresponding fuzzy fact (e.g., temperature is normal . ), and then transfers it to clips for inference. the problem is, how could those fuzzy facts match the premise patterns of rules if linguistic hedges exist. for example, the fact (temperature is high . ) should be able to match the patterns (temperature is very high) with degree . , (temperature is high) with degree . , (temperature is low) with degree . , (temperature is not very high) with degree . , etc. to cope with this problem, some control rules are used to generate a set of facts with linguistic hedges to replace the original fuzzy fact. one such control rule is shown as follows: ( d e f r u l e g e n e r a t e - l i n g u i s t i c - h e d g e - f a c t s ( d e c l a r e ( s a l i e n c e ) ) ? f f < - (new-fact ? o o i s $?ccl ? c c ~ ?dd) => ( r e t r a c t ? f f ) ( i f (member not $ ? c c l ) t h e n ( b i n d ? d d ( - ? d d ) ) ) ( i f (member very $ ? c c l ) t h e n ( b i n d ? d d ( * * ? d d o . ) ) ) ( i f (member more-or-less $?ccl) t h e n ( b i n d ? d d ( * * ?dd ) ) ) (assert (?oo is?ccz?dd) ( ? o o i s n o t ? c c = ( - l ? d d ) ) (?ooisvery?cc = ( * * ? d d z ) ) ( ? o o i s notvery ? c c z = ( - ( * * ?dd ) ) ) (?ooismore-or-less?ccz=(** ?ddo. ) ) ( ? o o isnotmore-or-less ?ccz = ( - ( * * ? d d o . ) hwang for example, let the input fact be (temperature is "c), the following fuzzy facts will be produced by fuzzification operations: (temperature is high . ) (temperature is middle . ) (temperature is low . ) the fact (temperature is low . ) is not removed immediately, since some of its corresponding facts with linguistic hedges will have nonzero degrees. to enable fuzzy inference with linguistic hedges, the following fuzzy facts are generated by the control rules to replace the three original facts: (temperature is high . ) (temperature is not high . ) (temperature is very high . ) (temperature is not very high . ) (temperature is more-or-less high . ) (temperature is not more-or-less high . ) (temperature is middle . ) (temperature is not middle . ) (temperature is very middle . ) (temperature is not very middle . ) (temperature is more-or-less middle . ) (temperature is not more-or-less middle . ) (temperature is low . ) (temperature is not low . ) (temperature is very low . ) (temperature is not very low .o) (temperature is more-or-less-low . ) (temperature is not more-or-less low .o) . evaluation of the new approach some experiments concerning the diagnosis of acute exanthema were made to evaluate the new approach. the earlier knowledge base was constructed by applying the emcud method," and rules are generated for diseases. i n appendix a, it can be seen that there are lots of redundant rules although the test results (in table i) seem to be desirable. with the new approach, fuzzy rules are generated (as given in appendix b) with the same test results. iv. conclusions in this article, we propose a knowledge acquisition system, kafes, for constructing fuzzy expert systems. we have applied kafes to a medical application. according to our experience, it can be seen that the use of fuzzy concepts to represent expertise significantly reduces the redundant rules in the knowledge bases while achieving a promising performance. moreover, it seems knowledge acquisition table i. test results of the knowledge base with redundant rules.a case number physician without redundant rules x x x x x x x x with redundant rules case number physician without redundant rules x x x with redundant rules athe codes of diseases and their translations are: i-measles, -german measles, -chickenpox, -smallpox, -scarlet fever, -exanthem subitum, -fifth disease, -meningococcemia, -rocky mi. spotted fever, -typhus fevers, i i-infectious mononucleosis, -enterovirus infections, -drug eruptions, and -eczema herpeticum. to be more natural and straightforward to represent knowledge of experts by fuzzy rules. currently, we are trying to enhance the system by applying multime- dia techniques. some related studies, including the learning algorithm for the adjustment of membership functions, the improvement of inference methodolo- gies, and the representation of multimedia knowledge bases, are in progress. appendix a: the repertory grid for acute exanthema d, - measles d - german measles d - chickenpox d - smallpox d , - scarlet d - exanthem subitum d , - fifth disease d - meningococcemia d - rocky mt. spotted fever d l - typhus fevers d,, - infectious mononucleosis d , , - enterovirus infections d , , - drug eruptions d , , - eczema herpeticum koplik’s spots - days of fever - days of fever headache flushed cheeks vomiting lymphadenopathy red maculopapular deep lesion centripetal circumoral pallor petechiae splenomegaly chills sore throat desauamation d l d d d d , d ol n o koplik’s spots not - days not - days no headache no flushedcheeks novomiting n o lymphadenopathy n o t r e d superficial centrifugal no circumoral pallor n o petechiae n o splenomegaly n o chills no sore throat no desauamation hwang ds d d i o d i l d d d koplik's spots - days of fever i - days of fever headache flushed cheeks vomiting lymphadenopathy red maculopapular deep lesion centripetal circum. pallor petechiae splenomegal y chills sore throat desquamation no koplik's spots not - days not - days no headache no flushed cheeks no vomiting no lymphadenopathy not red superficial centrifugal n o circumoral pallor no petechiae no splenomegaly n o chills no sore throat no desquamation rules generated from the repertory grids by emcud: rule,: i f then rule,, ,: i f then rule,,,: i f then rule,,,: i f then rule,: i f (koplik'sspots = t) a (daysoffever ) a (daysoffever ) a (vomiting = fj a (maculopapularcolor = brickred) a (circumoralpallor = fj a (petechiae = f) a (desquamation = t ) disease = measles cf = . (koplik'sspots = tj ((daysoffever > ) v (daysoffever < ) ) a (vomiting = fj a (maculopapularcolor = brickredj a (circumoralpallor = fj a (petechiae = fj a (desquamation = t ) disease = measles cf = . (koplik'sspots = t ) a (daysoffever ) a (daysoffever ) a (vomiting = f) a (maculopularcolor = brickred) a (circumoralpallor = f ) a (petechiae = t) a (desquamation = t ) disease = measles cf = . (koplik'sspots = t) a (daysoffever ) a (daysoffever ) a (vomiting = f) a (macu opapularcolor = bri ckred) a (circumoralpallor = f) a (petechiae = f) a (desquamation = f) disease = measles cf = . (daysoffever = ) a (vomi ting = f) a (lymphadenopathy = t ) a (maculopapularcofor = pink) a (circumoralpal- l o r = f ) a knowledge acquisition then rule , : if then rule , : if then rule,,,: if then rule,,,: if then rule,: if then rule,,,: if then rule,.,: if then (petechiae = f ) a (chills = f ) a (desquamation = f ) disease = german measles cf = . (daysoffever > ) a (vomiting = f ) a (lymphade- nopathy = t ) a (maculopapularcolor =pink) a (circumoralpal- lor = f ) a (petechiae = f ) a (chills = f ) a (desquama- tion = f ) disease = german measles cf = . (daysoffever = ) a (vomiting = f) a (lymphade- nopathy = t ) a (maculopapularcolor = pink) a (circumoralpal- lor = f ) a (petechiae = t ) a (chills = f ) a (desquama- tion = f) disease = german measles cf = . (daysoffever = ) a (vomiting = f) a (lymphade- nopathy = t ) a (maculopapularcolor = pink) a (circumoralpal- lor = f ) a (petechiae = f) a (chills = t ) a (desquama- tion = f ) disease = german measles cf = . (daysoffever = ) a (vomiting = f ) a (lymphade- nopathy = t ) a (maculopapularcofor = pink) a (circumoralpal- l o r = f ) a (petechiae = t ) a (chills = f ) a (desquama- tion = t ) disease = german measles cf = . (daysoffever ) a (daysoffever > ) a (head- ache = t ) a (vomiting = f ) a (maculopapufarcolor e {pink, red}) a (lesiondegree = superficial) a (distribution = centripetal) a (circumoralpallor = f ) a (petechiae = f ) disease = chickenpox cf = . ((daysoffever > ) v (daysoffever = )) a (headache = t ) a (vomiting = f ) a (maculopapularcolor e {pink, red}) a (lesiondegree = superficial) a (distribution = centripetal) a (circumoralpallor = f ) a (petechiae = f ) disease = chickenpox cf = . (daysoffever ) a (daysoffever > ) a (head- ache = f ) a (vomiting = f ) a (maculopapularcolor e {pink, red}) a (lesiondegree = superficial) a (distribution = centripetal) a (circumoralpallor = f ) a (petechiae = f ) disease = chickenpox cf = . hwang rule,,,: if then rule,: if then rule,,,: i f then rule,,,: if then rule,,,: if then rule,: if then rule,,,: if (daysoffever i ) a (daysoffever > ) a (head- ache = t ) a (vomiting = f) (maculopapularcolor e {pink, red}) a (lesiondegree = superficial) a (distribution = centripetal) a (circumoralpallor = f ) r\ *petechiae = t) disease = chickenpox cf = . (daysoffever < ) a (daysoffever ) a (head- ache = t ) a (vomiting = f ) a (maculopapularcolor e {pink, red}) a (lesiondegree = deepseated) a (distribution = centrifugal) a (circumoralapplor = f ) a (chills = t ) a (sore- throat = t ) disease = smallpox cf = . ( (daysoffever ) v (daysoffever < ) ) a (headache = t ) a (vomiting = f ) a (maculopapularcolor e {pink, red}) a (lesiondegree = deepseated) a (distribution = centrifugal) a (circumoralpallor = f ) a (chills = t ) a (sore- throat = t ) disease = smallpox cf = . (daysoffever < ) a (daysoffever ) a (head- ache = f ) a (vomiting = f) a maculopapularcolor e {pink, red}) a (lesiondegree = deepseated) a (distribution = centrifugal) a (circumoralpallor = f ) a (chills = t ) a (sore- throat = t ) disease = smallpox c f = . (daysoffever < ) a (daysoffever ) a (head- ache = t ) a (vomiting = f ) a (maculopapularcolor e {pink, red}) a (lesiondegree = deepseated) a (distribution = centrifugal a (circumoralpallor = f ) (chills = t ) a (sore- throat = f ) disease = smallpox cf = . (daysoffever ) a (daysoffever > . ) a (vom- iting = t ) a (maculopapularcolor = red) a (circumoralpallor = t ) a (petechiae = f ) a (desquamation = t ) disease = scarlet fever cf = . ((daysoffever > ) v (daysoffever . ) ) a (vomiting = t ) a (maculopapularcolor = red) a (circumoralpall or = t ) a (petechiae = f ) a (desquamation = t) knowledge acquisition rule,, : rule, : rule,, : rule,, : rule, : rule,, : rule, : rule,, : rule, : then if then if then if then if then if then if then if then if then if disease = scarlet fever cf = . (daysoffever ) a (daysoffever > . ) a (vom- iting = t ) a (maculopapularcolor = red) a (circumoralpallor = t ) a (petechiae = t ) a (desquamation = t ) disease = scarlet fever cf = . (daysoffever c: ) a (daysoffever ) a (vomit- ing = f ) a (maculopapularcolor = pink) a (circumoralpal- lor = f ) a (petechiae = f ) disease = exanthem subitum cf = . ( (daysoffever > ) v (daysoffever < ) ) a (vom- iting = f ) a (maculopapularcolor = pink) a (circumoralpal- l o r = f ) a (petechiae = f ) disease = exanthem subitum cf = . (daysoffever ) a (daysoffever ) a (vomit- ing = f ) a (maculopapularcolor = pink) a (circumoralpal- lor = f ) a (petechiae = t ) disease = exanthem subitum cf = . (daysoffever = ) a (flushedcheeks = t ) a (vom- iting = f ) a (macul opapul arcol o r = red) a (circumoralpall or = t ) a (petechiae = f) a (chills = f) disease = fifth disease cf = . (daysoffever = ) a (flushedcheeks = t ) a (vom- iting = f ) a (macul opapul arcol or = red) a (circumoralpall or = t ) a (petechiae = t ) a (chills = f) disease = fifth disease cf = . (daysoffever < ) a (daysoffever > ) a (vomit- ing = f ) a (maculopapularcolor e {pink, red}) a (circumoral- pallor = f ) a (petechiae = t ) disease = meningococcemia cf = . ( (daysoffever > ) v (daysoffever = ) ) a (vom- iting = t ) a (maculopapularcolor e {pink, red}) a (circumoral- pallor = f ) a (petechiae = t ) disease = meningococcemia cf = . (daysoffever ) a (daysoffever ) a (head- ache = t)a (vomiting = fj a (maculopapularcolor e {pink, red} j a (distribution = centrifugal) a (circumoralpal- lor = f j a hwang then rule,, ,: if then rule,,: if then rule,,,,: if then rule,,: if then rule,,,,: if then rule,,,,: if then rule,,,,: if then rule,,: if (petechiae = t) a (chills = t) disease = rocky mt. spots fever cf = . (daysoffever ) a (daysoffever ) a (head- ache = f ) a (vomiting = f ) a (maculopapularcolor e {pink, red}) a (distribution = centrifugal) a (circumoralpal- lor = f ) a (petechiae = t) a (chills = t ) disease = rocky mt. spots fever cf = . (daysoffever ) a (daysoffever ) a (head- ache = t ) a (vomiting = f ) a (maculopapularcolor e {pink, red}) a (distribution = centripetal) a (circumoralpal- lor = f ) a (petechiae = t ) a (chills = t) disease = typhus fevers cf = . (daysoffever ) a (daysoffever ) a (head- ache = f ) a (vomiting = f) a (maculopapularcolor e {pink, red}) a (distribution = centripetal) a (circumoralpal- l o r = f ) a (petechiae = t) a (chills = t) disease = typhus fevers cf = . (daysoffever i ) a (daysoffever > ) a (vomit- ing = f ) a (maculopapularcolor e {pink, red}) (circumoral- pallor = f ) a (splenomegaly = t ) a (sorethroat = t) disease = infectious mononucleosis cf = . ((daysoffever > ) v (daysoffever = )) a (vom- iting = f ) a (maculopapularcolor e {pink, red}) a (circumoral- pallor = f ) a (splenomegaly = t) a (sorethroat = t ) disease = infectious mononucleosis cf = . (daysoffever ) a (daysoffever > ) a (vomit- ing = f ) a (maculopapularcolor e {pink, red}) a (circumoral- pallor = t ) a (splenomegaly = t ) a (sorethroat = t) disease = infectious mononucleosis cf = . (daysoffever ) a (daysoffever > ) a (vomit- ing = f ) a (maculopapularcolor e {pink, red}) a (circumoral- pallor = f ) a (splenomegaly = t) a (sorethroat = f ) disease = infectious mononucleosis cf = . (daysoffever ) a (daysoffever ) a (vomit- ing = f ) a (maculopapularcolor = pink) a (circumoralpal- lor = f ) a (petechiae = f ) k n o w l e d g e acquisition then rule,,,,: i f rule,, : rule,, , rule,, : rule,, , : rule, : rule, : rule, : rule, : rule, : rule, : rule, : rule, : then i f then i f then i f then i f then i f then i f then i f then i f then i f then i f then i f then i f disease = enterovirus infections cf = . (daysoffever ) a (daysoffever ) a (vomit- ing = f ) a (macul opapul arcol o r = pink) a (circumoralpal- l o r = f ) a (petechiae = t) disease = enterovirus infections cf = . (daysoffever ) a (daysoffever > ) a (vomit- ing = f)a (maculopapularcolor = pink) (circumoralpall or = f ) a (petechiae = f ) disease = drug eruptions cf = . (daysoffever ) a (daysoffever > ) a (vomit- ing = f ) a (maculopapularcol or = pink) (circumoralpall or = f ) a (petechiae = t) disease = drug eruptions cf = . (daysoffever = ) a (vomiting = f ) a (maculopapularcolor = pink) a (circumoralpal- l o r = f ) a (chills = f) a (sorethroat = f) disease = eczema herpeticum cf = . (daysoffever = ) a (vomiting = f ) a (maculopapularcolor = pink) a (circumoralpal- l o r = f ) a (chills = f) a (sorethroat = t) disease = eczema herpeticum cf = . (wbc f low) disease = measles cf = . (wbc # low) disease = exanthem subitum cf = . ( wbc # low) a ( wbc f normal ) disease = german measles cf = . ( whil eballcel lamount i ) wbc = low cf = . ( whi ebal cell amount ) a (whileballcellamount ) wbc = normal cf = . (whileballcellamount > ) wc = highcf = . appendix b: fuzzy rules (koplik’s spots is definite) a (fever is seri- ous) a (vomiting is not apparent)a (maculopapular color is brick red) a (circu- moral pallor is not apparent) a (petechiae is not apparent) a (desquamation is apparen t disease = measles cf = . (fever is none) a (petechiae is not apparent)a (vomiting is not apparent) a (lymphadenopathy is definite) a hwang then rule,: i f then rule,: i f then rule,: i f then rule,: i f then rule,: i f then rule,: i f then rule,: i f then ( m a c u l o p a p u l a r c o l o r i s p i n k ) a ( c i r c u m o r a l p a l - l o r i s n o t a p p a r e n t ) a ( c h i l l s i s n o t a p p a r e n t ) a ( d e s q u m a t i o n i s n o t a p p a r e n t ) d i s e a s e = german m e a s l e s c f = . ( f e v e r i s l i g h t ) a ( h e a d a c h e i s a p p a r e n t ) a ( v o m i t i n g i s n o t a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r i s more or l e s s r e d ) a ( l e s i o n d e g r e e i s s u p e r f i c i a l ) a ( d i s t r i b u t i o n i s c e n t r i p e t a l ) a ( c i r c u m o r a l p a l l o r i s n o t a p p a r e n t ) a ( p e t e - c h i a e i s n o t a p p a r e n t ) d i s e a s e = c h i c k e n p o x c f = . ( f e v e r i s s e r i o u s ) a ( h e a d a c h e i s a p p a r e n t ) a ( s o r e t h r o a t i s a p p a r e n t ) ( v o m i t i n g i s n o t a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) a ( l e s i o n d e g r e e i s d e e p s e a t e d ) a [ d i s t r i b u t i o n i s c e n t r i f u g a l ) a ( c i r c u m o r a l p a l l o r i s not a p p a r e n t ) a ( c h i l l s i s a p p a r e n t ) a d i s e a s e = s m a l l p o x c f = . ( f e v e r i s m o r e o r l e s s s e r i o u s ) a ( v o m i t i n g i s a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r i s r e d ) a ( c i r c u m o r a l p a l - l o r i s a p p a r e n t ) a ( p e t e c h i a e i s n o t a p p a r e n t ) a ( d e s q u a m a t i o n i s a p p a r e n t ) d i s e a s e = s c a r l e t f e v e r c f = . ( f e v e r i s s e r i o u s ) a ( v o m i t i n g i s n o t a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r is p i n k ) ( c i r c u m o r a l p a l l o r i s n o t a p p a r e n t ) a ( p e t e - c h i a e i s not a p p a r e n t ) d i s e a s e = e x a n t h e m subitum c f = . ( f e v e r i s n o n e ) a ( f l u s h e d cheeks i s a p p a r e n t ) a ( v o m i t i n g i s not a p p a r e n t ) a [ m a c u l o p a p u l a r color i s r e d ) a ( c i r c u m o r a l p a l - l o r i s a p p a r e n t ) a ( p e t e c h i a e i s n o t a p p a r e n t ) a [ c h i l l s i s not a p - p a r e n t ) d i s e a s e = f i f t h d i s e a s e c f = . ( f e v e r i s l i g h t ) a ( v o m i t i n g i s a p p a r e n t ) a ( p e t e c h i a e i s a p p a r e n t ) ( m a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) a ( c i r c u - m o r a l p a l l o r i s n o t a p p a r e n t ) a d i s e a s e = m e n i n g o c o c c e m i a c f = . ( f e v e r i s s e r i o u s ) a ( h e a d a c h e i s a p p a r e n t ) a ( v o m i t i n g i s not a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) a ( d i s t r i b u t i o n i s c e n t r i f u g a l ) a ( c i r c u m o r a l p a l - l o r i s not a p p a r e n t ) a ( p e t e c h i a e i s a p p a r e n t ) a ( c h i l l s i s a p p a r e n t ) d i s e a s e = r o c k y mt. s p o t s f e v e r c f = . knowledge acquisition rule,,: rule,, : rule,, : rule,,: rule,, : i f then i f then i f then i f then if then ( f e v e r i s s e r i o u s ) a ( h e a d a h c e i s a p p a r e n t ) a ( v o m i t i n g i s n o t a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) a ( d i s t r i b u t i o n i s c e n t r i p e t a l ) a ( c i r c u m o r a l p a l - l o r i s n o t a p p a r e n t ) a ( p e t e c h i a e i s a p p a r e n t ) a ( c h i l l s i s a p p a r e n t ) d i s e a s e = t y p h u s f e v e r s c f = . ( f e v e r i s l i g h t ) a ( v o m i t i n g i s n o t a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) a ( c i r - c u m o r a l p a l l o r i s not a p p a r e n t ) a ( s p l e n o m e g a l y i s a p p a r e n t ) a ( s o r e t h r o a t i s a p - p a r e n t ) d i s e a s e = i n f e c t i o u s m o n o n u c l e o s i s c f = . ( f e v e r i s m o r e o r l e s s s e r i o u s ) a ( v o m i t i n g i s not a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r is more or l e s s p i n k ) a ( c i r c u m o r a l p a l l o r i s n o t a p p a r e n t ) ( p e t e c h i a e i s n o t a p p a r e n t ) d i s e a s e = enterovirus i n f e c t i o n s c f = . ( f e v e r i s l i g h t ) a ( v o m i t i n g i s n o t a p p a r e n t ) a ( p e t e c h i a e i s not a p p a r e n t ) ( m a c u l o p a p u l a r color i s more or l e s s p i n k ) ( c i r - c u m o r a l p a l l o r i s not a p p a r e n t ) / d i s e a s e = drug e r u p t i o n s c f = . [fever i s none) a ( v o m i t i n g i s not a p p a r e n t ) a ( m a c u l o p a p u l a r c o l o r i s more o r l e s s p i n k ) a ( c i r c u m o r a l p a l l o r i s not a p p a r e n t ) ( c h i l l s i s not a p p a r e n t ) a ( s o r e t h r o a t i s n o t a p p a r e n t ) d i s e a s e = eczema h e r p e t i c u m c f = . this study was supported by the national science council, republic of china, under the contract number: nsc - -e- - . references . g.j. klir and t.a. folger, fuzzy sets, uncertainty, andinformation, prentice-hall, new jersey . . j . boose, “a knowledge acquisition program for expert systems based on personal construct psychology,” int. j . man-mach. s t u . , , - . . j. boose and j.m. bradshaw, “neoets: capturing expert system knowledge in hierarchical rating grids,” ieee expert s y s t e m in government symposium, . . j. boose and j.m. bradshaw, “expertise transfer and complex problems: using aquinas as a knowledge acquisition workbench for knowledge-based systems,” int. j . man-mach. s t u . , , - ( ). . l. esheman, d. ehret, j . mcdermott, and m. tan, “mole: a tenacious knowledge acquisition tool,” int. j . man-mach. s t u . , , - ( ). . k.m. ford, f.e. petry, j.r. adams-webber, and p.j. chang, “an approach to knowledge acquisition based on the personal construct systems,” ieee trans. k n o w / . and d a t a e n g . , ( ), - ( ). . k . m . ford, a. canas, j. jones, h. stahl, j. novak, and j.r. adams-webber, hwang “iconkat: an integrated constructivist knowledge acquisition tool,” int. j . knowl. a c q . , , - ( ). . b.r. gaines, “an overview of knowledge acquisition and transfer,” int. j . . s.t. gruber, the acquisition of strategic knowledge, academic press inc., san diego, . . g. kahn, s . nowlan, and j . mcdermott, “more: an intelligent knowledge acquisi- tion tool,” international joint conference of art$cial intelligence, , - ( ). . s. marcus, “taking backtracking with a grain of salt,” int. j . man-mach. s t u . , . r.m. o’bannon, “an intelligent aid to assist knowledge engineers with interviewing experts,” ieee western conference on expert s y s t e m , . . m.l.g. shaw and b.r. gaines, “kitten: knowledge initiation and transfer tools for experts and novices,” int. j . man-mach. s t u d . , , - ( ). . j.m. bradshaw, k.m. ford, j.r. adams-webber, and j.h. boose, “beyond the repertory grid: new approaches to constructivist knowledge acquisition tool develop- ment,” int. j . intell. sys., , - ( ). . l . e . wood and j.m. ford, “structuring interviewings with experts during knowl- edge elicitation,” int. j . intel. sys., , - ( ). . g.a. kelly, the psychology of personal constructs, norton, new york, . . j . giarratano and g . riley, expert systems: principles and programming, pws- kent publishing company, boston, . . g.j. hwang and s.s. tseng, “emcud: a knowledge acquisition method which captures embedded meanings under uncertainty,” int. j . man-mach. stu., , man-mach. stti., , - ( ). , - ( ). - ( ). a neural-adaboost based facial expression recognition system expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a a neural-adaboost based facial expression recognition system - � the authors. published by elsevier ltd. http://dx.doi.org/ . /j.eswa. . . ⇑ corresponding authors. tel.: + (e. owusu), tel.: + (y. zhan). e-mail addresses: kayowusueb@yahoo.com (e. owusu), yzzhan@ujs.edu.cn (y. zhan), mao_qr@ujs.edu.cn (q.r. mao). open access under cc by-nc-nd license. ebenezer owusu ⇑, yongzhao zhan ⇑, qi rong mao school of computer science and communication engineering, jiangsu university, xuefu road, zhenjiang, jiangsu, china a r t i c l e i n f o a b s t r a c t keywords: facial expression recognition bessel transform gabor feature adaboost mffnn this study improves the recognition accuracy and execution time of facial expression recognition system. various techniques were utilized to achieve this. the face detection component is implemented by the adoption of viola–jones descriptor. the detected face is down-sampled by bessel transform to reduce the feature extraction space to improve processing time then. gabor feature extraction techniques were employed to extract thousands of facial features which represent various facial deformation patterns. an adaboost-based hypothesis is formulated to select a few hundreds of the numerous extracted features to speed up classification. the selected features were fed into a well designed -layer neural network clas- sifier that is trained by a back-propagation algorithm. the system is trained and tested with datasets from jaffe and yale facial expression databases. an average recognition rate of . % and . % are regis- tered in jaffe and yale databases, respectively. the execution time for a � pixel size is . ms. the general results of the proposed techniques are very encouraging when compared with others. � the authors. published by elsevier ltd. open access under cc by-nc-nd license. . introduction facial expression is the explicit transformation of the human face due to the automatic responses to the emotional instability. in most situations it is spontaneous and uncontrollable. the auto- matic facial expression involves the application of an artificial intelligent system to recognize the expressions of the face under any circumstance. today, the studies of facial expressions have gained keen interest in pattern recognition, computer vision and its related fields. mainly, such facial expressions are the seven pro- totypical ones, namely; anger, fear, surprise, sad, disgust, happy, and neutral. research into automatic facial expression recognition is very important in this modern society of technological age. for instance, the technology is applied in a wide variety of contexts, including robotics, digital signs, mobile applications, and medicine. it is re- ported that ‘‘some robots can operate by first recognizing expres- sions’’ of humans (bruce, ). the aibo robot for instance is a biologically-inspired robot that can show its emotions via an array of leds located in the frontal part of the head (breazeal & scassel- lati, ). in addition to this, the robot can also display ‘happiness’ feeling when it detects a face. in behavioral sciences and medicine for instance, expression recognition is effectively applied for inten- sive care monitoring (morik, brockhausen, & joachims, ). cur- rently, there are developing systems that are capable of making routine examinations of facial behavior during pain in clinical set- tings. in infants the neonatal facial coding system (nfcs) has been employed for real-time assessment within to week post-con- ceptional age infants who are undergoing a heel lance. the technol- ogy is being used in more advanced settings to reduce accidents through the implementation of automated detection of driver drowsiness in public transports. this system relays information about the drivers’ emotional states to observers for effective sur- veillance leading to necessary awareness. the hallmark of every facial expression system is accuracy and to some extent the speed of execution. however most of the exist- ing systems produce poor performances in terms of accuracy; as for execution speed, most of the systems are even silent to give a hint. some few examples; franco and treves ( ) proposed a neural based facial expression recognition system that used princi- pal component analysis (pca) to reduce the feature vectors. the features were fed into a feed-forward neural network that was trained by a back-propagation network. in this system an average recognition of . % was reported on the yale facial expression database – an achievement which is not very encouraging. kumb- har, jadhav, and patil ( ) described a neural network classifica- tion facial expression recognition system that employs gabor feature extraction and feature reduction by pca to distinguish - class facial expression recognition on the jaffe database. in this system they specified inputs, to hidden layers and seven output feed-forward neural networks. again, the – % recogni- tion accuracy they obtained by their procedure is not encouraging http://crossmark.crossref.org/dialog/?doi= . /j.eswa. . . &domain=pdf http://dx.doi.org/ . /j.eswa. . . mailto:kayowusueb@yahoo.com mailto:yzzhan@ujs.edu.cn mailto:mao_qr@ujs.edu.cn http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa http://creativecommons.org/licenses/by-nc-nd/ . / http://creativecommons.org/licenses/by-nc-nd/ . / fig. . sample face detection images (left), cropped detected face (right). e. owusu et al. / expert systems with applications ( ) – to befit the expectations of a real-time system. recently, londhe and pawar ( ) extracted features of the face using affine mo- ment invariants and performed the classification using feed-for- ward neural network. the expression recognition obtained was . % on the jaffe database. tai and chung ( ) extracted the facial features using a sobel filter. in their experiment they re- served the maximum connected component to reduce the wrinkles and noises and conducted -class classification on jaffe facial expression database through the application of elman network with two hidden layers, each layer containing fifteen neurons. with this approach the average accuracy of automatic facial expression recognition is . %. zhang and tjondronegoro ( ) extracted the expressive face by using gabor filters, feature reduc- tion by pca and expression classification by neural network. in this method an average facial recognition of . % was recorded in the jaffe facial expression database. dailey and cottrell ( ) also extracted facial features by gabor techniques and reduced the fea- tures by pca. the expression classifier was neural network and the average expression recognition was . % ± . on the seven proto- typical facial expressions, however the facial expression database was not mentioned. most of these studies advocate the use of neural network as the expression classifier and extracted the facial features by gabor fil- ters and reduced the features via pca. the displeasing thing is that all the results were not very encouraging. this study persists in exploring the potentials of neural net- works to execute this kind of assignment, trying to esteem some biological constraints, utilizing the capabilities of modular systems. though many techniques have been used to extract the facial features, gabor feature extraction remains a high-quality choice; there are other alternatives but they are not very promising. just a few examples: satiyan and nagarajan ( ) utilized the haar technique to extract facial features which were used as input to the neural network for classifying facial expressions. the haar wavelet extraction is very fast (satiyan & nagarajan, ; van, ), however the wavelets are too huge to result to effective classification when used as input to classifiers in facial expression recognition (cemre, ); in other words it is a potential cause to misclassifications and poor performance. distance-based feature extraction methods are also one of the largely applied techniques used for feature extraction in both d and d static faces. the idea behind these procedures is that the muscle deformations which are the major causes of changes in facial expression from normal expression results in variations of the euclidean distances between facial landmarks or points. these points, as well as their distances, have been widely employed for static facial expression analysis (sha, song, bu, chen, & tao, ; soyel & demirel, ; tang & huang, ). among the most successful ones is feature extrac- tion based on the bhattacharyya distance (choi & lee, ). how- ever, despite some advantages of this method, the degree of computational complexity is unacceptably high. the matching of even a small model shape with a normal image can take half an hour on an eight-processor sun sparcserver (rucklidge, ; zhang & lu, ). the patch based feature extraction meth- od is another alternative widely exploited for facial expression bio- metrics. maalej, amor, daoudi, srivastava, and berretti ( ) for instance represented extracted patches from facial surfaces by sets of closed curves and then applied a riemannian framework to ob- tain the shape analysis of the extracted patches. however, the patch-based features also have numerous drawbacks. first, partic- ular representations cannot be applied to other solutions without major modifications: the majority of the techniques have only been utilized to a single class. also, most methods do not exploit the large amounts of available training data (aghajanian et al., ). thus on the basis of these we still considered gabor features as the best approach, not because it does not have drawbacks, but the drawbacks can be easily managed. the gabor filter is a superior model of simple cell receptive fields in cat striate cortex (jones & palmer, ), and it grants exceptional basis for object recogni- tion and face recognition (lades et al., ; wiskott, fellous, kru- ger, & vondermalsburg, ). again, the gabor methods are superior to all the above-mentioned methods because it extracts the maximum information from local image regions (deng, jin, zhen, & huang, ), and it is invariant against, translation and rotations (al daoud, ). in this work, the data were reduced in dimensions by bessel transform (ganga, prakash, & gangashetty, ) and then after extraction of the face by gabor methods, the features were further reduced via an adaboost-based (freund & schapire, ) feature reduction technique. the selected features which represented the facial deformation patterns were then fed into a -layer feed-for- ward neural network that is trained by a back-propagation algo- rithm. it is interesting to note that bessel down-sampling techniques have never been adopted for facial expression recogni- tions. again, the combinations of bessel down-sampling and the formulated adaboost-based algorithm is an innovation that re- duces the expression dataset to enhance accuracy and speed. final- ly, the construction of the feed-forward neural network is influential to bring about successful results. the rest of the work is arranged as follows. section discusses face detection and image down-sampling. section discusses ga- bor feature extraction. section discusses feature selection. sec- tion discusses the multilayer feed-forward neural network (mffnn). results and analysis are presented in section . the final conclusions of the work are drawn in section . . face detection and down-sampling the face detection component was implemented by the adop- tion of viola and jones system (viola & jones, ). fig. shows sample face detection by the viola–jones classifier. the size of the image is rescaled to a window of size � pixels by the use of bessel down-sampling. methods like bilinear interpolations have been utilized by several authors for this task in particular but interpolations are prone to aliasing problems (munoz, blu, & unser, ). bessel down-sampling reduces the size of the image and preserves the details and perceptual quality of the original image (ganga et al., ). the down-sampling im- age signal xd (t , t ) is expressed as: xdðt ; t Þ¼ xp n ¼ xq n ¼ cðn ; n Þj an p � r t � � j an q � s t � � ð Þ where p and q refer to the respective image size, p � r and q � s are the required reduced size of the image; r and s are positive integers that represent the reduction values, n is the number of low-fre- quency dct coefficients, j ðan Þ and j ðan Þ are zero order bessel e. owusu et al. / expert systems with applications ( ) – functions, c(n , n ) are bessel coefficients computed from the first order bessel function, t and t are chosen such that t p � r and t q � s. interested readers are referred to (al daoud, ). . gabor feature extraction the -d gabor filters are spatial sinusoids localized by gaussian window, and they can be created to be selective for orientation, localization, and frequency as well. it is very flexible to demon- strate images by gabor wavelets because the details about their spatial relations are preserved in the process. gðx; y; h; u; rÞ¼ pr exp � x þ y r � � expf piðr þ r g ð Þ where i is a complex number representing the square root of � . r = ux cos h and r = uy sin h, u is the spatial frequency of the band pass, h is the spatial orientation of the function g, (x, y) spec- ify the position of light impulse in the visual field, r is the standard deviation of -d gaussian envelop. in this gabor family, we chose eight orientations ; p ; p ; . . . ; p � � and five scales ; ffiffiffi p ; ; n ffiffiffi p ; g: in order to give added robustness to illumination we turned the gabor filter to zero dc (direct current) by the expression ~gðx; y; h; u; rÞ¼ gðx; y; h; u; rÞ� q xn i¼�n xn j¼�n gðx; y; h; u; rÞ " # ð Þ where, q is the size of the filter, given by q = ( n + ) . fig. shows the gabor filter image. the sample points of the filtered image is coded to two bits, real bit x and imaginary bit x such that, g ¼ x ¼ ; if r ~gðx; y; h; u; rÞ h i � i n o p x ¼ ; if r ~gðx; y; h; u; rÞ h i � i n o < >< >: ð Þ g ¼ x ¼ ; if i ~gðx; y; h; u; rÞ h i � i n o p x ¼ ; if i ~gðx; y; h; u; rÞ h i � i n o < >< >: ð Þ where i is subimage of the expressional face, r and i are the real and the imaginary parts of each gabor kernel, � is the convo- lution operator. with this coding, only the phase information in the facial expressions image is stored in the feature vector of size bytes. the final magnitude response which is used to represent the feature vectors is computed by g ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi g þ g q ð Þ fig. shows the magnitude response of a template image. fig. . gabor filtered image: real (left) and imaginary (middle) parts of gabor filter in . feature selection due to the large size of the gabor wavelets, it is not practically possible to use all the wavelets as input to our classifier, for fear of misclassification and possible system crash. the adaboost feature reduction algorithms have special speed advantage in increasing classification process (shen & bai, ). thus we formulated an adaboost-based algorithm to select a few deserving portions of the wavelets. assuming the extracted gabor features are represented by a total of i e ( , ,..., n) appearance features. then the image i is represented as ui ¼fðxn; ynÞg n n¼� configured by the parameters z, l, v. the positive sets /(+) and the negative sets /(�) are denoted by /ðþÞ ¼ fðxn; ynÞg n n¼ � r j �ð� Þ and /ð�Þ ¼ fðxn; ynÞg n n¼� � r j� ð� Þ respectively, where xn is the nth data sample containing j fea- tures, and yn is its corresponding class label. to train the vectors ||g||, which is denoted by /(u,v,z) over a distribution d, we simply determined the weights of all the feature vectors ui ¼ fðxn; ynÞg n n¼� ¼ / ðþÞ þ /ð�Þ: this gives us a threshold k which indicated the decision hyperplane. k is computed as: k ¼ p i /ðþÞ dðiÞ:/ðl;v;zÞ jj p i /ðþÞ dðiÞ:/ðl;m;zÞjj þ p i /ð�Þ dðiÞ:/ðl;v;zÞ jj p i /ð�Þ dðiÞ:/ðl;v;zÞjj ð Þ a sample is positive or client if it is located at the positive half of k (which is the majority decision), otherwise it is a negative or an imposter. the status is reversed if the minority of the positive in- stances is rather located in the positive half space. let c be denoted by clients and p be the imposters. for a given training dataset con- taining both positive and negative samples, where each sample is (si, yi); y e { ± } represents the corresponding class label, the fea- ture selection algorithm is formulated as follows: � initialize sample distribution d by weighting every training sample equally such that the initial weight w ,i = / c, / p for y = and � , respectively. � for the iteration t = , ,..., t, where t is the final iteration, do: (i) normalize the weights, wt;i wt;ipn i¼ wt;i , where wt is a probabil- ity distribution and n is the total number of features. (ii) train a weak classifier ht for feature j, which uses a single feature. the training error nt is estimated with respect to wt such that: d, and t nt ¼ x t wt;ijhtðxiÞ� yij ð Þ (iii) select the hypothesis h t with the most discriminating infor- mation, that is to say, the hypothesis with the least classifi- cation error n t ; on the weighted samples. (iv) compute the weight xt that weights h t by its classification performance as: xt ¼ ln n t � " # ð Þ he real part of the gabor kernels ( d) in the spatial and frequency domain. fig. . a gabor magnitude response of face image: a sample image (left), the magnitude response image of the whole gabor filter bank of gabor filters (right). e. owusu et al. / expert systems with applications ( ) – (v) the weight distribution is then updated and normalized by: wtþ ;i � wt;i:e�xt yi h t ðs t Þ ð Þ � the final feature selection hypothesis h(s) which is a function of the selected features is denoted by: hðsÞ¼ sgn xt t¼ xt h t ðs t Þ " # ð Þ the selected features represent samples of the facial deforma- tion patterns of the expressive face. the datasets which were images from the jaffe and yale databases were partitioned into training and testing by leave-one-out cross validation (wu, bru- baker, mullin, & rehg, ). fig. . a -layer feed-forward neural network. . multilayer feed-forward neural network (mffnn) classifier the selected features are fed into the constructed neural net- work to train it to identify the seven universal facial expressions. the architecture is a -layer feed-forward neural network and trained by a back-propagation algorithm (bouzalmat, belghini, zarghili, kharroubi, & majda, ; londhe & pawar, ). the back propagation algorithm basically replicates its input to its out- put via a narrow conduit of hidden units. the hidden units extract regularities from the inputs because they are completely connected to the inputs. every network was trained to give the maximum va- lue of for exact facial expression and for all other expressions. the construction is shown in fig. . the input layer has nodes, each for each facial expression whiles the hidden layer had neu- rons; each expression for neurons. we chose neurons to com- pensate for the target output of seven facial expressions. this was the case for the seven prototypical facial expressions, which was validated by the use of the jaffe facial expression database. since the experiment was also validated using the yale database, where four expressions were used there was a slight modification in the construction of the network for this application. here, the hidden layer neurons were settled at each facial expression dedicated for neurons and neurons in the output layer. the input vectors of the network represented by x = [x , x ,..., xl] t. the output layers are denoted by y = [y , y ,..., yk] t. the optimi- zation model is formulated as x : h ? y. the output dataset of each layer of the network is denoted by yj ; :::; y j k� ; y j k ; j ¼ ; ; :::; k � ; k; where k � corresponds to the total hidden layers and k represents the total output layers. we denote the tar- get datasets and its additive white noise by (t , t ,..., tk) and g = (e , e ,..., ek), respectively. the variable k represents the total patterns of the network. the corresponding vectors of the hidden units are denoted by v = (v , v ,..., vk� ). the sigmoid activation function h = (h , h ,..., hk) of each layer is h , h ,..., hk� . the weights of the network are updated by w , w ,..., wk. the training epochs are and the target of error is . . the training algorithm is modeled as: min h ;h ;v ;w ;w xk j¼ ðtj � yj Þ ð Þ subject to the constraints yj ¼ h ðw x jÞ; w rv �m; yj r v ; xj rm yj ¼ h ðw ; y j Þ; w r �v ; yj r >= >; ð Þ the process of training involves weight initialization, calcula- tion of the activation unit, adjustment, weight adaptation, and test- ing for convergence of the network. assuming vji represents the weight between the jth hidden unit and ith input unit; and wkj rep- resents the weight between the kth output and the jth hidden unit. the activation unit is then calculated sequentially, starting from the input layer. the activation of hidden and output unit is calcu- lated as: table confusion matrix of -class facial expression recognition on yale. neural (%) sad (%) happy (%) surprise (%) neutral . . . sad . . . happy . . . surprise . . . average recognition = . % e. owusu et al. / expert systems with applications ( ) – yðpÞj ¼ h ðpÞ yj xi i¼ v ji zi � v jo ! ð Þ oðpÞk ¼ h ðpÞ ok xj j¼ wkjyj � wko ! ð Þ where yðpÞj is the activation of the jth hidden unit and o ðpÞ k is the acti- vation of the kth output unit for the pattern, p. h is a sigmoid func- tion. k is the total number of output units, i is the total number of input units and j is the total number of hidden units. vjo is the weight connected to the bias unit in the hidden layer, zo = � and yo = � . we adjusted the weights, starting at the output units and recursively propagated error signals to the input layer. the detected output oðpÞk is compared with the corresponding target value t ðpÞ k which is a facial image, over the entire training set using the sig- moid function to express the approximation error in the network’s target functions. eðpÞ ¼ xk k¼ tðpÞk � o ðpÞ k ð Þ the minimization of the error e(p), requires the partial deriva- tive of e(p) with respect to each weight in the network to be com- puted. the change in weight is proportional to the corresponding derivative. dv jiðt þ Þ¼�g @eðpÞ @v ji þ adv jiðtÞ ð Þ dwkjðt þ Þ¼�g @eðpÞ @wkj þ adwkjðtÞ ð Þ where, g is the learning rate, normally between and , we set it to . . the function a is also set to . . the last term is a function of the previous weight change. @e @v ji ¼ @yj @v ji xk k¼ �ðtk � okÞokð � okÞyj wkj ð Þ @yj @v ji ¼ yjð � yjÞzi: therefore, dv ji ¼ g xk k¼ ðtk � okÞokð � okÞyjwkj yjð � yjÞzi ð Þ the weights are updated by, wkjðt þ Þ¼ wkjðtÞþ dwkjðt þ Þ ð Þ v jiðt þ Þ¼ v jiðtÞþ dv jiðt þ Þ ð Þ where, t is equal to the current time step. dvji and dwkj are the weight adjustments. we repeated the process once from the equa- tion ( ) in order to achieve the desired output. table confusion matrix of -class facial expression recognition on jaffe. neutral (%) sad (%) fear (%) neutral . . . sad . . . fear . . anger . . disgust happy . . surprise . average recognition = . % . results and analysis the facial expression recognition was validated with the jaffe and yale facial expression databases. the jaffe database contains images of female japanese persons. each respondent in the database posed three or four examples of each of the seven facial expression prototypes (happy (ha), sad (sa), anger (an), disgust (di), fear (fe), surprise (su), and neutral (ne)). images of each individual from each class of expres- sion are randomly selected for training, leading to a total of images corresponding to . %, the rest was preserved for testing. the trial was performed using tenfold cross-validation to obtain the average recognition rate. in order to create distinct datasets for cross-validation, none of the sets in the training folder appear in any of the remaining folders. the yale facial expression database contains grayscale images in gif format of individuals. there are images per subject. each subject exhibited one of the six facial expressions; ha, ne, sa, sleepy (sl), surprise (su) and wink (wi). in this database we manually extracted images corresponding to ha, ne, sa, and su. the datasets in this database were also partitioned into training and testing by using the same method described for the jaffe. in due course about % of the images were used for training and the remaining, for testing. we recorded an average recognition rate of . % in jaffe and . % in yale on intel(r) core(tm) duo cpu p @ . ghz ( cpus) – . ghz and . ghz ram computer running on win- dows ultimate -bit ( . , build ) (see tables and for the confusion matrices). the best recognitions were detected in ha, su and di, where we obtained almost % in the jaffe database. we saw that facial images with extreme exhibited expressions recorded the best re- sults. generally, the performance of ne was the weakest; about . % in jaffe and . in yale. the results show that the defor- mations of the muscles around the mouth and the eyes are the most reliable determinants for automatic facial expressions. this accounts why recognition in the neutral face is poor. thus the in- crease in these muscle deformations increases the accuracies of automatic recognitions. the execution time for a pixel of size � is . ms. fig. shows the comparative performance of execution time with other neural network classifiers. anger (%) disgust (%) happy (%) surprise (%) . . . . . . . . . . . . . . . . x image pixels p ro ce ss in g t im e (m s) proposed method ma & khorasani classifier kharat & dudul classifier kumbhar et al. classifier fig. . comparing execution time of different classifiers. fig. . comparing recognition rates of different methods in jaffe database. fig. . comparing recognition rates of different methods in jaffe database. fig. . comparing the average recognition rates of different methods in jaffe and yale databases. e. owusu et al. / expert systems with applications ( ) – the proposed method was compared with three different clas- sifiers to access its performance in terms of recognition accuracy and execution speed. the system was also tested with some real life images from the world wide web. the results indicate that the proposed method is statistically better (p < . ) both in accu- racy and speed. the three methods are described as follows: method i (same as ma and khorasani ( ) method): the fea- ture detector is by a discrete cosine transform (dct), pruning tech- nique is used to reduce the input size of the network, the training algorithm is back-propagation procedure, the expression classifier is a feed-forward neural network with one hidden layer. method ii (same as kumbhar et al. ( ) method): the feature detector is gabor method, the feature dimensionality is principal component analysis (pca) and the expression classifier is a feed- forward neural network. method iii (same as kharat and dudul ( )): the feature detector is discrete cosine transform, the feature reduction is by principal component analysis (pca), and the classifier is a feed-for- ward neural network. all the classifiers were trained and tested with the same data- sets used for the proposed method. figs. and shows the average recognition rates of various expressions in jaffe and yale respec- tively; the intent is not to compare the performances of the two databases but to investigate for the robustness of the system in di- verse databases. fig. also shows the comparison of the overall average recognition rates of various descriptors in jaffe and yale. fig. shows sample real-time expression recognitions by the sys- tem. the average recognition rates are also compared with other methods that employed the same datasets in their experiment to give a general idea of the performance of the proposed method (see tables and for details). however this does not signify a di- rect comparison because the experiments were not conducted un- der the same environment. the results show that the proposed method is very encouraging. though performance in the yale facial expression database is reduced as compared to that in the jaffe fa- cial expression database, it is far better than all the performances in yale we compared with. fig. . sample facial expression recognitions: happy (left), anger (middle), disgust (right). table comparative performance of recognition rates in different methods on jaffe database. author classifier/ method database rate (%) lekshmi and sasikumar ( ) svm jaffe . kumbhar et al. ( ) image feature jaffe – zhi and ruan ( ) d-dlpp jaffe . zhao, zhuang, and xu ( ) pca and neural network jaffe . lee, huang, and shih ( ) rda jaffe . proposed method mffnn jaffe . table comparitive performance of recognition rates in different methods on the yale database. author classifier/method database rate (%) lekshmi and sasikumar ( ) svm yale . franco and treves ( ) neural network yale . proposed method mffnn yale . e. owusu et al. / expert systems with applications ( ) – . conclusions this study employs many advanced techniques to improve the recognition rate and execution time of facial expression recogni- tion system. face detection was carried out by the application of viola–jones descriptor. detected faces were down-sampled by the bessel transform. this approach reduced the image dimensions and preserved the perceptual quality of the original image. an ada- boost based algorithm was formulated to select a few hundreds of gabor wavelets from the several thousands of the extracted fea- tures to reduce the computational cost and to avoid misclassifi- cation as well. the selected features were fed into a well- designed multilayer feed-forward neural network classifier. the network is thus trained with sample datasets from both jaffe and yale facial expression databases. the remaining datasets from the two databases and some images from the world wide web were used to test for the system. the execution time for a pixel of size � is . ms; the average recognition rate in jaffe database is . % and that in yale is . %. the proposed method is compared with several methods and the performance is out- standing. the results of the study also show that automatic expres- sion recognitions are very accurate in surprise, disgusts and happy; about %. mild expressions like sad, fear and neutral have lower accuracies. however fear can be very accurate when it is at the peak because accuracies in recognitions largely depend on the magnitude of facial deformations around the mouth and eyes. to advance towards % efficiency we believe the development of natural databases would be of more help since many artificial dat- abases have many confused scenarios among facial expressions in sad, neutral and mild anger. again future improvements of recog- nition accuracies will look at the possibility of increasing the num- ber of hidden neurons to expressions that recorded lower values. acknowledgments this paper is supported by the national nature science founda- tion of china (nos. and ), the natural science foundation of jiangsu province (no. bk ), and the research foundation for talented scholars of jiangsu university (no. jdg ) references aghajanian, j., warrell, j., prince, s. j., li, p., rohn, j. l., & baum, b. ( ). patch- based within-object classification. in: ieee th international conference on computer vision, (pp. – ). al daoud, j. e. ( ). enhancement of the face recognition using a modified fourier–gabor filter. international journal of advanced software computer applications, . bouzalmat, a., belghini, n., zarghili, a., kharroubi, j., & majda, a. ( ). face recognition using neural network based fourier gabor filters & random projection. international journal of computer science and security, , – . breazeal, c., & scassellati, b. ( ). robots that imitate humans. trends in cognitive sciences, , – . bruce, v. ( ). what the human face tells the human mind: some challenges for the robot–human interface. advanced robotics, , – . cemre, z. ( ) facial expression recognition. msc thesis, university of surrey. choi, e., & lee, c. ( ). feature extraction based on the bhattacharyya distance. pattern recognition, , – . dailey, m. n., & cottrell, g. w. ( ). pca = gabor for expression recognition. ucsd cse tr cs- . deng, h. b., jin, l. w., zhen, l. x., & huang, j. c. ( ). a new facial expression recognition method based on local gabor filter bank and pca plus lda. international journal of information technology, , – . franco, l., & treves, a. ( ). a neural network facial expression recognition system using unsupervised local processing. in: ieee proceedings of the nd international symposium on image and signal processing and analysis (pp. – ). freund, y., & schapire, r. e. ( ). a desicion-theoretic generalization of on-line learning and an application to boosting. in computational learning theory (pp. – ). berlin, heidelberg: springer. ganga, m. p., prakash, c., gangashetty, s. v. ( ). bessel transform for image resizing. in: ieee th international conference on systems, signals and image processing, (pp. – ). jones, j. p., & palmer, l. a. ( ). an evaluation of the two-dimensional gabor filter model of simple receptive fields in cat striate cortex. journal of neurophysiology, , – . http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h e. owusu et al. / expert systems with applications ( ) – kharat, g. u., & dudul, sv. ( ). emotion recognition from facial expression using neural networks. human–computer systems interaction. berlin, heidelberg: springer (pp. – ). berlin, heidelberg: springer. kumbhar, m., jadhav, a., & patil, m. ( ). facial expression recognition based on image feature. international journal of computer and communication engineering, , – . lades, m., vorbruggen, j. c., buhmann, j., lange, j., von der malsburg, c., wurtz, r. p., et al. ( ). distortion invariant object recognition in the dynamic link architecture. ieee transactions on computers, , – . lee, c. c., huang, s. s., & shih, c. y. ( ). facial affect recognition using regularized discriminant analysis based algorithms. eurasip journal on advances in signal processing, . lekshmi, v. p., & sasikumar, m. ( ). analysis of facial expression using gabor and svm. international journal of recent trends in engineering, . londhe, r., & pawar, v. ( ). facial expression recognition based on affine moment invariants. international journal of computer science issues, . ma, l., & khorasani, k. ( ). facial expression recognition using constructive feed- forward neural networks. ieee transactions on systems, man, and cybernetics, part b: cybernetics, , – . maalej, a., amor, b., daoudi, m., srivastava, a., berretti, s. ( ). local d shape analysis for facial expression recognition. in: ieee th international conference on pattern recognition (pp. – ). morik, k., brockhausen, p., & joachims, t. ( ). combining statistical learning with a knowledge-based approach – a case study in intensive care monitoring. in: icml (pp. – ). munoz, a., blu, t., & unser, m. ( ). least-squares image resizing using finite differences. ieee transactions on image processing, , – . rucklidge, w. j. ( ). efficient locating objects using hausdorff distance. international journal of computer vision, , – . satiyan, m., & nagarajan, r. ( ). recognition of facial expression using haar-like feature extraction method. in: ieee international conference on intelligent and advanced systems (pp. – ). sha, t., song, m., bu, j., chen, c., & tao, d. ( ). feature level analysis for d facial expression recognition. neurocomputing, , – . shen, l., & bai, l. ( ). adaboost gabor feature selection for classification. in: proceeding of image and vision computing, (pp. – ), new zealand. soyel, h., & demirel, h. ( ). facial expression recognition using d facial feature distances. image analysis and recognition. berlin, heidelberg: springer (pp. – ). berlin, heidelberg: springer. tai, s. c., & chung, k. c. ( ). automatic facial expression recognition system using neural networks. in: tencon ieee region conference (pp. – ). tang, h., & huang, t. ( ). d facial expression recognition based on properties of line segments connecting facial feature points. in: th ieee international conference on automatic face & gesture recognition (pp. – ). van, p. f. ( ). discrete wavelet transformations: an elementary approach with applications. hoboken, nj: wiley-interscience. viola, p., & jones, m. j. ( ). robust real-time face detection. international journal of computer vision, , – . wiskott, l., fellous, j. m., kruger, n., & vondermalsburg, c. ( ). face recognition by elastic bunch graph matching. ieee transactions on pattern analysis and machine intelligence, , – . wu, j. x., brubaker, s. c., mullin, m. d., & rehg, j. m. ( ). fast asymmetric learning for cascade face detection. ieee transactions on pattern analysis and machine intelligence, , – . zhang, d., & lu, g. ( ). review of shape representation and description techniques. pattern recognition, , – . zhang, l., & tjondronegoro, d. ( ). selecting, optimizing, and fusing ‘salient’ gabor features for facial expression recognition. neural information processing. berlin, heidelberg: springer (pp. – ). berlin, heidelberg: springer. zhao, l., zhuang, g., & xu, x. ( ). facial expression recognition based on pca and nmf. in: ieee th world congress on intelligent control and automation (pp. – ). zhi, r., & ruan, q. ( ). facial expression recognition based on two- dimensional discriminant locality preserving projections. neurocomputing, , – . http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h a neural-adaboost based facial expression recognition system introduction face detection and down-sampling gabor feature extraction feature selection multilayer feed-forward neural network (mffnn) classifier results and analysis conclusions acknowledgments references pii: - ( ) - pergamon expert systems with applications, vol. , no. / , pp. -- , copyright © elsevier science lid printed in great britain. all rights reserved - / $ . + . s - ( ) - interactive induction of expert knowledge bingchiang jeng, ting-peng liang and minyang hong department of information management, national sun yat-sen university, kaohsiung, taiwan , republic of china abstract--the process o f extracting, structuring and organizing elicited knowledge (called knowledge acquisition) is a bottleneck in developing knowledge-based systems. a manual approach that elicits domain knowledge by interviewing human experts typically has problems, because the experts are often unable to articulate their reasoning rules. an automatic approach that induces knowledge from a set o f training cases also suffers from the unavailability o f sufficient training cases. we present an integrated approach that combines the strengths o f both methods to compensate f o r their weaknesses. in this approach, human experts are responsible f o r solving problems, whereas an inductive learning algorithm is responsible f o r reasoning and consistency checking. copyright © elsevier science ltd . i n t r o d u c t i o n k n o w l e d g e acquisition is a process o f extracting, structuring and organizing elicited knowledge from domain experts or other knowledge sources and convert- ing the knowledge into rules or other forms o f representation accessible with a computer program [ , ]. as the power o f an expert system comes from the knowledge that it possesses [ ], the acquisition o f knowledge is a major stage in the development of an experts system. the transfer of expertise from various sources to a knowledge base is difficult and challenging. a manual process that encodes domain knowledge by interviewing human experts suffers from several difficulties [ , ]. for example, human experts are typically not also expert in articulating their rules of reasoning because they are not explicitly aware o f the structure o f their knowledge. a paradox o f expertise [ ] claims that the better one is an expert, the worst one is to tell the details. knowledge acquired during interview may be unreliable, as verbal reports and mental behavior are not necessarily corre- lated [ ]. incongruities may occur among what an expert says that he does, what he actually does and what he should have done [ ]. communication between the knowledge engineer and the domain expert is another difficult problem. the knowledge engineer may know little about the problem domain, and may not understand clearly the jargon used by the domain expert. on the other hand, the domain expert may not understand the process o f knowledge acquisition and what the knowledge engineer needs. even the cultural differences between the two parties o f knowledge acquisition may undermine the development o f a knowledge-based system. when historical examples are available, an automatic process can be used to induce rules from fragments o f knowledge [ , ]. the unique power o f this approach is the "automatic" learning capability o f its inductive algorithm. many successful applications are reported, such as disease diagnoses [ , , ], business applica- tions [ , , , ] and others [ , , ]. a major constraint o f automatic learning is that it requires many training cases from which to induce knowledge. the quality o f the training result depends on the quality o f the data. another drawback is that induction may generate knowledge that uses paths of reasoning different from those used by the expert. this may make human experts reluctant to accept the induced knowledge base regardless o f its performance. given that each approach alone has limitations and their respective strengths and weaknesses seem com- plementary, we seek to integrate the two approaches to gather the advantages and to avoid the deficiencies. the idea is that the weaknesses o f one approach be compensated by the strengths o f the other. for example, the task o f articulating a reasoning process for a human expert can be left to an inductive method to solve, or the restriction o f inductive learning on training examples can be given to an experienced human expert to provide help. this approach is useful as human experts are good at solving problems whereas inductive learning algorithms are good at extracting rules o f reasoning. previous research that aimed at a similar goal exists. for example, parsaye [ ] proposed an approach that combines interactive knowledge acquisition with rule induction. based on the theory o f repertory grids [ , ], the system can interactively interview an expert by bingchiang jeng et al asking questions to capture the expert's knowledge. the captured knowledge is then generalized using rule induction. buntine and stirring [ ] presented another approach that allows a human expert to interact with a set o f automatically induced rules so that each other's knowledge is cross-checked. recently evans and fisher [ ] reported a similar approach that is successfully applied to a problem domain where the experts have only weak causal knowledge. a limitation o f buntine and stirling's approach is that it still depends on a set o f pre-existing training examples. in some cases, these examples are hard to come by. furthermore, the way an expert interacts with the preliminary rules is informal. the approach presented in this paper extends their ideas to allow interactive knowledge acquisition from (nearly) scratch. a modified algorithm for inductive learning is designed so that an expert communicates with the system in an interrogative style during the inductive process. in addition to inducing a decision tree from existing training cases, the system also identifies near- miss cases falling on the classification borders for clarification and verification by human experts. these cases, after classification, become new training cases for accurate learning by the modified inductive algorithm. this way, the expert's knowledge is accumulated incre- mentally and the learning process can be repeated until the induced knowledge is satisfied by the human expert. the remainder o f the paper is organized as follows. the process of automatic knowledge acquisition and its related techniques is reviewed in section . an inter- active process for inductive learning that shows how to elicit knowledge from a domain expert is described in section . experimental results from the evaluation of the performance o f the proposed approach is presented in section . section summarizes the research. . a u t o m a t i n g k n o w l e d g e a c q u i s i t i o n there are many different methods for knowledge acquisition, which can be classified as: manual, semi- automatic and automatic [ , ]. eliciting knowledge by manual methods is highly labor-intensive. methods belonging to this category include structured/unstruc- tured interviews, analysis o f protocol, observations and so forth [ ]. semi-automatic methods are primarily designed to support either the expert or the knowledge engineer to perform the necessary task more effectively. tools to elicit domain knowledge include ets, kriton, aqui- nas, more, m o l e etc. [ , ]. others include teiresias [ ] that assists the knowledge engineer to update a knowledge base, and kads [ ] that provides a collection o f tools to support a knowledge engineer to extract, to structure, to analyze and to document expert knowledge. although semi-automatic methods expedite the work o f knowledge engineers and/or domain experts, they still share problems o f manual methods: the acquired knowl- edge is difficult to validate; correlation between verbal reports and mental behavior may be weak. in addition, michie [ ] stated that knowledge o f certain types could not be elicited using manual methods, simply because the domain was so large and complicated that the expert would be unable to explain how it operates. automatic methods that use machine-learning tech- niques to extract knowledge require less or no participation by either knowledge engineers or domain experts. therefore, they do not have the difficulties associated with human experts. currently, the most popular automated method is rule induction. induction is a process o f general inference from particular instances. rule induction (or inductive learn- ing) refers to a concept learning process by which a set o f rules is created from training cases to explain or to predict a problem-solving behavior. when the induced structure o f knowledge is represented in the form o f a decision tree, it is also called (decision) tree induction [ , ]. a decision tree is considered as a set o f rules in a compact form; it can be transformed into rules easily. early work on rule induction is traced to when hunt, martin and stone developed a method for induction. the method was later implemented and expanded b y paterson and niblett [ ] to create acls (a concept-learning system) and by quinlan [ , ] to develop the popular id . in the following, we shall use the id algorithm to illustrate how rule induction works. input to id is a collection o f training cases. each is described by a set o f attributes associated with a class name (or outcome). an attribute can be either categorical (e.g. color) or numerical (e.g. age). a numerical attribute can adopt discrete or continuous values. the basic procedure o f id applies a divide-and-conquer approach to partition recursively the data set, based on a test on selected attribute values, into mutually exclusive subsets. the procedure continues until each subset contains cases o f the same class (to avoid over-fitting the training data, termination conditions may be defined) or no attribute is available for further decomposition. if s= {(ai~ . . . . . ain; ci)laij~a j, ci~c)} is a set o f training cases, where aj denotes the domain o f attribute j and c for the class, the algorithm is shown in fig. . an example shown in fig. is the decision tree created by id from financial data to analyze the risk of bankruptcy. a new case is classified by examining which leaf it reaches when traveling down through the tree. the traversed path is determined b y a sequence of tests to the new case (i.e. a branch is selected depending on the outcome o f a test that the new case generates). when it reaches a leaf in this way, it is considered similar to other existing cases contained in the leaf as they have all passed the same tests. hence, the outcome o f the new case is the same as its neighbors and is assigned with the i n t e r a c t i v e i n d u c t i o n o f e x p e r t k n o w l e d g e procedure i n d u c t i o n (s: trainset, t: tree) begin if s contains cases of the same class then label t with the class name and exit else begin for each numerical attribute a i, do find a value v/to decompose the training set into two subsets, calculate the entropy of the decomposition, choose the decomposition whose entropy value is the smallest; for each categorical attribute, decompose s by its classes and calculate its entropy; partition s into two or more mutually exclusive subsets: s i, i = ,..., k, based on the attribute a s whose entropy value is the smallest after decomposition. create a node t i for each subset s i, and link it to the parent node as its child. for i = to k, call i n d u c t i o n (s i, t i) end end, figure . t h e id algorithm. class n a m e labeled at the leaf. f o r e x a m p l e the prediction o f b a n k r u p t c y for case ( , , , . , . , . , . ) is yes, as it follows the path at the far left d o w n to a leaf labeled " y e s " . . i n t e r a c t i v e i n d u c t i o n t h e fundamental basis for methods o f decision tree induction (or rule induction) is that t w o cases with similar features are classified into the same class. the success o f such an a p p r o a c h relies on sufficient training cases to c o v e r every aspect o f the p r o b l e m d o m a i n w i t h o u t inconsistency. otherwise, the i n d u c e d k n o w - ledge m a y be unreliable. in practical applications, however, this a s s u m p t i o n is c o m m o n l y violated. t h e sources o f training cases are, in general, f r o m a domain expert o r f r o m historical data in a maintained database. difficulties arise typically because the domain expert can provide only a few selected examples, whereas historical data, if they exist, m a y be obsolete o r contain errors and missing values. a further difficulty with inductive learning is in its explanation capability. explaining the reasoning process b y which a conclusion is r e a c h e d is a key requirement for an expert system. rules i n d u c e d f r o m the training set may, however, differ f r o m those used b y the expert. this makes its explanation s o m e t i m e s less acceptable to h u m a n beings because the u n d e r l y i n g process o f reason- ing used b y the expert s y s t e m m a y be incomprehensible to users. it is thus useful to have interactions between h u m a n experts and inductive learning algorithms. this allows h u m a n experts to provide useful knowledge, such as hand-crafted tutorial examples, rules o f thumb, general hints and problem-solving strategies, to help an inductive algorithm. the inductive algorithm presented b e l o w is an example that supports interaction with experts during its learning process. after a k n o w l e d g e structure is induced f r o m its training cases, the a l g o r i t h m identifies cases that cannot be correctly classified. t h e s e cases are b r o u g h t to the expert in the f o r m o f questions to be solved. o n c e they are solved, they b e c o m e n e w training cases to the inductive algorithm for further learning. in this way, expert k n o w l e d g e is incrementally elicited and incorpor- ated into the induced k n o w l e d g e structure. a typical scenario o f the interactive inductive process is as follows. initially, the k n o w l e d g e e n g i n e e r collects k n o w l e d g e f r o m all possible sources. different forms o f k n o w l e d g e are stored differently. rules (mostly f r o m domain experts) are stored in the rule set and tutorial cases (mostly f r o m the historical database) are stored in the training set. then, the algorithm tests any inconsistencies l e a f < ' ~ leaf leaf leaf l e a f l e a f b a n k r u p t c y : y e s , n o c i : c o n s i s t e n c y o p i n i o n c : subject to o p i n i o n c : g o i n g c o n c e r n qi : net income/total assets q : current assets/total assets ( ~ : current r a t i o : cash/total assets figure . a n e x a m p l e of a decision tree for bankruptcy prediction. bingchiang jeng et al. y xc a l a a a x < . x / • • • " b ~ - ~ y < ~ / ~y>ye a a • b sb sc xc a a b b sc y c sb b x (a) a decision tree (b) partition of feature space figure . s p a c e partition by a decision tree. in the knowledge structure represented by these two sets. as both sets are incomplete at the beginning, contra- dictory cases (i.e. cases classified differently by those two knowledge sources) are identified. these cases are then presented to the domain expert for further review. prompting experts to review contradictory cases has two purposes. first, certain knowledge ignored by the expert may be use to provoke him to describe the inference rules in more detail. then the contradictory cases, after correction, become new examples to the inductive algorithm for further learning. revision o f the two sets triggers the cross-validation process again and the above procedure is repeated. a sketch of this approach is shown in fig. . an important feature of the above algorithm is its ability to detect inconsistencies during the process o f interactive induction. this is important when two different sources o f knowledge are merged because sa x inconsistencies may exist. in the following we shall present an algorithm for detecting inconsistencies. `% sb g i v e n b o rd er • • m e m o . • a e o y con-ect border . . s t r a t e g y f o r b o r d e r v a l i d a t i o n a decision tree created from training cases can be represented as a partitioned feature space. the induction o f decision trees (or rules) in some sense is a process to maximize the internal similarity within the partitioned subspaces. each leaf o f the tree corresponds to a subspace, and each case in the problem domain corre- sponds to a point in it. for example, a decision tree shown in fig. partitions a two-dimensional feature space with attribute (x, y) into three subspaces. sub- spaces s, and st represent class a and subspace sb represents class b. a new case is evaluated based on the subspace into which it falls (e.g. a case with x - and y~< and y ~ > - then class = else class= ; the concept learned after the first iteration o f the induction is given in fig. (b), from which the contra- dictory cases identified are shown in fig. (c). the final concept induced after three iterations o f the learning process is given in fig. (d). one can see that the result approaches the actual circle pattern. in addition to the nine given training cases, the system generates a total o f questions for the human expert to answer in the learning process. figure shows how other geometric patterns can be learned in experiments. they all achieve satisfactory results. the learned patterns are near the actual patterns. interactive induction of expert knowledge y x x (a) initial training cases ~ . ~ l a s s (b) first inductive result y ~ . . . . , , o . . - _x :::cia s i : : : . : . ; . ; . ; . : ( c ) identified conflicting (d) final inductive result cases figure . a learning process for the circle experiment. the only limitation is that the nature o f the orthogonal partition with id does not allow a real curvilinear draw. . . learning blackjack games another experiment shows how a system can learn the rules that a human expert uses in playing blackjack. the game is played between a dealer and players. each player receives two reversed cards, while the dealer receives one reversed and one obverse cards. an ace can count as either point or points. during the game, a player m a y request for more cards (to "hit") or to stay with the current hand (to "stand"). a player wins if the total points o f cards in hand is below points but greater than the dealer's. a human expert decides whether to hit or to stand based on his current count in hand, with or without an ace, and the dealer's obverse card. these three attributes determine a decision. the expert's rules can be drawn in a decision diagram as shown in fig. (a). initially five randomly chosen cases were given for training. the rules learned after three iterations are shown in fig. (b), which are pretty close to the actual rules used b y the expert. . related work the subject discussed in this article can also be seen as a problem o f concept learning in the machine learning field. in particular, the way a domain expert interacts with a learning system, can be formulated as a model o f learning with membership queries [ ], in which an instructor or oracle exists to answer any queries placed by a learner. in this model, the learner has control o v e r what part o f information it receives from a problem domain next. although our work starts initially from a different perspective, its result happens to be in accordance with most recent work in machine learning. cohn et al. [ ] presented an approach to learn a concept by generating queries from a region o f uncertainty, an area in the domain where misclassification is still possible. an interesting coincidence in their neural network's imple- mentation is that it uses two networks s and g to identify an uncertain case, when outcomes o f the case are inconsistent between them. they demonstrated in several domains that the approach gives better learning results for a fixed number o f training cases than simply learning from examples alone. this is encouraging since the rationale o f identifying regions o f uncertainty is similar to our identification o f contradictory cases. another learning model similar to the above is the on- line (or incremental) learning model in which the learner answers a sequence o f yes/no questions with immediate feedback provided after each question. a variant o f the on-line learning model, called self-directed learning, has been recently proposed b y goldman and sloan [ ] which allows the learner to select the presentation order for the instance. thus it can be seen as a variation o f learning with membership queries in which the learner is only "charged" for queries whose outcomes are unpre- y . : . : . . . : - : . --i.',';:i:i:i;:i:::i.;.,~'~" ' • .~ ~ (a) ~ class x ~,:'j ..v.k-.~ t ; [ . " x . ' . ~.::;: !!!i (b) figure . l e a r n i n g other geometric patterns. class (c) bingchiang jeng et al. p l a y e r ' s current count player's current count i d e a l e r ' s o b v e r s e c a r d g a - - s t a n d . . . . - - - h i t . . . . a - - - s t a n d _ - - with ace without ace ( a ) e x p e r t s ' r u l e s to p l a y b l a c k j a c k dealer's obverse card a g a o - - s t a n d - - - ~ o _ _ _ stand - h i t - - - h i t - - o lo hit w i t h a c e w i t h o u t a c e ( b ) r u l e s l e a r n e d a c c o r d i n g t o t h e i n d u c t i v e a l g o r i t h m figure . learning the blackjack playing rules. • s t a n d . . _ h i t _ dictable. theoretical results given by goldman and sloan show that the performance of self-directed learning is the best among all other commonly studied on-line and query learning models. our process of learning by detecting contradictory cases near the classification border is also in accordance with the main results discovered from an autonomous exploratory learning system (i.e. without an external teacher). winston [ ] first drew attention to the critical role of near-miss training examples. recently the protos system, developed by porter et al. [ ], also used near- miss cases to learn an examplar difference before matching a new case to an existing examplar. . d i s c u s s i o n a n d c o n c l u s i o n s in this article, we have presented an algorithm for interactive induction and evaluated the performance of the implemented system in learning goemetric patterns and blackjack game strategies. the algorithm is simple but efficient. the contribution of this work is two-fold. first, it suggests an effective way o f integrating manual knowl- edge acquisition and inductive learning to achieve a more accurate knowledge base. second, it provides a new way of creating near-miss training examples and learning from these examples. this allows critical knowledge to be learned without a large number o f training cases. further research following this includes replacing id with other learning methods, testing in other domains and exploring the handling o f certainty factors in interactive induction. the current approach can also be extended to acquire knowledge from multiple experts. the main difficulty for knowledge acquisition from multiple experts is how to combine their expertise and find out any inconsistencies systematically. due to different subjective opinions, two competitive domain experts may disagree with each other occasionally. our approach should be helpful in identifying conflicting rules or cases during the knowledge acquisition process. r e f e r e n c e s f i. angluin, d. ( ). queries and concept learning. machine learning, , - . . bainbridge, l. ( ). asking questions and accessing knowledge. in future computing systems. new york: elsevier. . boose, j. ( ). personal construct theory and the transfer of human expertise. proc. of the nat'l conf on artificial intelligence, austin, tx. . boose, j. & gaines, b. r. ( ). knowledge acquisition for knowledge-based system: notes on the state-of-the-art. machine learning, , - . . braun, h. & chandler, j. s. ( ). predicting stock market behavior through rule induction: an application of the learning- from-example approach. decision sciences, , - . . buchanan, b. g. et al. ( ). constructing an expert system. in e hayes-roth, d. waterman & d. lenat (eds), building expert systems. reading, addison-wesley. . buntine, w. & stirling, d. ( ). interactive induction. in j. e. hayes-michie, d. michie & e. tyugu (eds), machine intelligence, vol. , pp. - , oxford: oxford university press. . carter, c. & catlett, j. ( ). assessing credit card applications using machine learning. ieee expert, , - . . cohn, d., atlas, l. & ladner, r. ( ). improving generalization with active learning. machine learning, , - . . chung, h. m. & silver, m. s. ( ). rule-based expert systems and linear models: a n empirical comparison of learning-by- example methods. decision science, , - . . eriksson, h. ( ). a survey of knowledge acquisition techniques and tools and their relationship to software engineering. journal of system and software, , - . . evans, b. & fisher, d. ( ). overcoming process delays with decision tree induction. ieee expert, -- . . feigenbaum, e. a. ( ). the art of artificial intelligence: themes and case studies of knowledge engineering. international joint conference on artificial intelligence, vol. , pp. - . . gaines, b. r. ( ). knowledge acquisition: development and advances.. in m. d. oliff, (ed.) expert system and intelligent programming. new york: elsevier. . goldman, s. a. & solan, r. h. ( ). the power o f self-directed learning. machine learning, , - . . hart, a. ( ). knowledge acquisition for expert systems, nd edn. new york: mcgraw-hill. . hayes-roth, e , waterman, d. a. & lenat, d. ( ). building expert systems. reading ma: addison-wesley. . jeng, b. & weyuker, e. ( ). a simplified approach to domain testing. acm transactions on software engineering and method- ology, , - . . jeng, b., liang, t. p, & jeng, y. m. film: a fuzzy inductive learning method for automatic knowledge acquisition. forthcoming in decision support systems. . johnson, p. e. ( ). w h a t kind of expert should a system be? journal of medicine and philosophy, g, - . . kelly, g. a. ( ). the psychology of personal constructs., new york: norton. . liang, ". p. ( ). a composite approach to inducing knowledge for expert systems design. management science, g, - . interactive induction o f expert knowledge . marchand, a., vanlente, e & galen, r. ( ). the assessment of laboratory tests in the diagnosis of acute appendicitis. american journal of clinical pathology, : , - . . marcus, s. ( ). automating knowledge acquisition for expert systems. boston, ma: kluwer. . mcgraw, k. l. & harbison-briggs, k. ( ). knowledge acquisition: principles and guidelines. englewood cliffs, nj: prentice-hall. . messier, w. e & hansen, j. v. ( ). inducing rules for expert system development: an example using default and bankruptcy data. management science, , - . . michalski, r. s. & chilausky, r. l. ( ). knowledge acquisition by encoding expert rules versus computer induction from examples: a case study involving soybean pathology. int. j. man-machine study, , - . . michie, d. (ed.) ( ). introductory readings in expert systems. new york: breach. . michalski, r., mozetic, i., hong, j. & lavrac, n. ( ). the multi-purpose incremental learning system aq and its testing application to three medical domains. in proc. of the th annual nat'l conference on artificial intelligence, pp. - , phil- adelphia, pa. . mitchell, t. m. ). generalization as search. artificial intelligence, lg, - . . parsaye, k. ( ). acquiring and verifying knowledge automat- ically. a expert, pp. - . . paterson, a. & niblett, t. ). acls user manual. scotland: intelligence terminal ltd. . porter, b.w., bareiss, r. & holm, r.c. ( ). concept learning and heuristic classification in weak-theory domains. artificial intelligence, , - . . quinlan, j. r. ( ). discovering rules from large collections of examples: a case study. in d. michie (ed.), expert systems in the micro electronic age. scotland: edinburgh university press. . quinlan, j. r. ( ). induction of decision trees. machine learning, , - . . rendell, l. a. ( ). a general framework for induction and a study of selective induction. machine learning, , - . . ruff, r. a. & dietterich, t. g. ( ). what good are experiments. proc. of the sixth international workshop on machine learning, pp. - , ithaca, new york. . scott, p. d. & markovitch, s. ( ). t~xperience selection and problem choice in an exploratory learning system. machine learning, , - . . shaw, m. j. & gentry, j. a. ( ). using an expert system with inductive learning to evaluate business loans. financial manage- ment, , - . . turban, e. ( ). expert systems and applied artificial intelli- gence. new york: macmillan. . utgoff, p. e. ( ). incremental induction of decision trees. machine learning, , - . . wardle, a. & wardle, l. ( ). computer aided diagnosis--a review of research. meth. inform. med., , - . . winston, p. h. ( ). learning structural descriptions from examples. in p. h. winston (ed.), the psychology o f computer vision. new york: mcgraw-hill. digital signal processor control algorithm for pwm inverter a. a. heggo abstract recent developments in power electronic device technology promises faster switching capability at high power. the hybrid pulse width modulation method which requires two of the four switches in a full bridge inverter are used, and enable pulse pattern operation at high frequency. the use of two level switching instead of three level switching enables the use of higher frequency for given computation time delay. the proposed control scheme is implemented using bi-polar junction transistors (bjt) controlling an inverter t o produce a very low frequency (thd) sinusoidal output voltage. simulation and experimental results are presented to verify the performance. introduction the bridge inverter in fig. ( ) is widely employed in various applications such as motor drives and active filter [ ][ ]. the inverter comprises switching poles s , s , s , s , s and s . the switching poles are commonly controlled by a variety of pwm techniques , and by pulse- shified square-wave drives [ ]. application of switching devices such as igbt achieve very high switching frequency pwm inverters with improved performance [ ]. with the availability of high frequency switching devices the instantaneous feedback control (ifc) was presented [ ]. the advantages of this technique are high transient response and the disadvantage is that relatively large harmonic amplitudes occur for frequencies near the average switching frequency. by using a microprocessor, a digital feedback approach such as' a microprocessor deadbeat control was proposed [ ][ ]. the pwm inverter system is converted into a discrete time system and a state feedback output deadbeat control is applied. manuscript received from dr; a. a. heggo on : accepted on: / / engineering research bulletin, vol ,no , minufiya university, faculty of engineering, shebin el-kom , egypt, issn - digital signal processor @sps) are now applied for the control of power electionics and drive systems. dsps are much faster (ten or one hundred times) than a microprocessor [ ]. simplification of control hardware and corresponding reduction of cost are the principal advantages ofdsp control [lo]. a digital controller was used to implement the control algorithm and provide switching signal t o the power circuit. the proposed control model is implemented using dsps controlling an inverter to produce a very low'total harmonic distortion (thd) in sinusoidal output voltage. fig. - system description control algorithm a modified algorithm of the deadbeat controlled pwm inverter is suitable for stabilised power supply systems. two levels are used in the pulse pattern. given a computation time delay, twice the switching fi-equency can be adapted resulting in lower thd sinusoidal output. this work presents the deduction of a new discrete time state equation and the proposed deadbeat control algorithm for pwm inverters. simulation results are obtained and presented. , . deadbeat control this technique depends upon the digital feedback closed loop. this means measuring of output, and controls the inverter switches to generate the required pwm pattern to produce a low thd sinusoidal output voltage. the digital control algorithm is designed to control pulse width such that the output voltage equals the sinusoidal reference at every sampling instant, fig. ( ). so the output voltage will be in phase and very close to sinusoidal reference. any deviation of the output voltage from the reference due to a load disturbance or non- linear circuits is controlled within one sampling interval ts. (a) sampling instant fig. - sampling time the pwm pattern is determined at every sampling instant digitally by dsps based on the output measurements and the references. . state model reference i error vin vout + inverter filter load digital aid controller converter fig. - block diagram of digital control inverter as shown in fig. ( ) the deadbeat controller for pwm inverter is considered. the inverter, l c filter and resistive load represent the plant of closed loop digital feedback system. l '~ load r v o u t fig. - state model the circuit shown in fig. ( ) is modelled as a second-order system with state vector fli], where vis load voltage and i is capacitor current. the state equation becomes: where: . two level scheme the basic circuit for the deadbeat controlled pwm inverter with a two level switching pattern is shown in fig. ( ). positive half cycle negative half cycle fig. - two level deadbeat controlled pwm inverter pattern the continuous time domain state ( i ) can be written as: i= a x + b u where: x = state vector u = scalar input a = non-singular matrix then the closed form circuit is: where x(fj is the initial state vector at t = to ifthe input u is constant for f , i t i t then eq.( ) becomes: using eq.( ), the discrete-time system equation with the input is derived as follows: and by expansion the terms: aat/ aat a ~ ( a t / ~ ) ~ e x ] + - + e a t ~ l + a t + (at ) then eq.( ) becomes: x/(k + i ) ti = eat x(ki] - eati beat + ( + a'z + (i i ) where (k + i ) = t;, and kt = , this is a discrete-time system of eq.( ) rewriting eq.(l ) gives: where: f, is corresponding to element of eat gi is corresponding to element of e a t n ~ ~ hi is corresponding to element of (t + a f / + a ~ / ~ ) b e v(k), i(x) and at@) represent the values of voltage, the current and the sampling time: t = k t therefore, by analysis of eq.(l ) gives: computer simulation the following circuit parameters were used in computing the pulse width at(k), and to obtain waveforms of voltage and current. line load two level scheme thdifimdamental . l . the simulation results are shown in figs. ( ), ( ) and ( ) conclusion a two deadbeat controller is used to minimise the total harmonic distribution of the digital inverter. this scheme will lead to the following advantages: - provide more computation time for the same interval switching number - provide a maximum voltage equal to the dc busbar voltage during each switching interval, ei + e the obtained results show the waveforms of output voltages during various types of modulation. references hue, c., two-level switching pattern in deadbeat dsp controlled pwm inverter, ieee translation of power electronic, vol. , no. , may . goodnough, f., mos controlled thyristor turns off imw in ms, electron design, november . gokhale, k. p., kawamers, a,, and hor, r. g., deadbeat microprocessor control of pwm for sinusoidal output waveform systems, ieee trans. ind. appliket, vol. a , no. , pp - , september/october . enjeti, p. n., programmed pwm technique to eliminate harmonics - a critical evaluation, ieee ias c o g rec., pp - , . patel, h. and hoft, r. g., generalised technique of harmonics elimination and voltage control in thyristor, part i - harmonic elimination, ieee trans. ind. appli., vol. ia- , no. , pp - , mayijune . kawamura, a., haneyoshi, t. and hoft, r. g., deadbeat controlled pwm inverter with parameter estimation using only voltage sensor, ieee, power electron, vol. , no. , pp - . . hue, c. and hofi, r. g., deadbeat controlled pwm inverter with two level pulse pattern for ups, by msee project - university of massouri, columbia, november . . truxal, feedback control system and synthesis. . heuldewoth, j. and grant, d. a,, the use of harmonic distortion to increase the output voltage of three-phase pwm inverter, ieee proc. vol. , plb no. , july , pp - . . heumann, k., rapp, g. and jung, m., comparative study of new power transistor with respect to high frequency inverter application, epe , , pp - . nomenclature at sampling time a, b matrix of state variable equation aid analogue / digital converter c filter capacitance dsp digital signal processor f output frequency i capacitor current k integer, positive number , , , l filter inductive n no. of sampling pwm pulse width modulation r load resistance t interval time v load output voltage v;, input voltage ( - b) current wavici;orhl fig. the voltage and current waverorms under the " pulse modulation mode" - pul :e acsc fig. the voltage and current waveforms under the " pulse modulation mode" , , , . ( - b ' c u r i < e n t \vavi follh] fig. the voltage and current waveforms under the " pulse modulation mode" d a t a a n a l y ti c s e n h a n c e d c o m p o n e n t v ol a t ili t y m o d e l yao, y, z h a i, j, c a o , y, di n g , x, li u , j a n d l u o , y h t t p : // d x. d oi. o r g / . /j. e s w a . . . t i t l e d a t a a n a l y ti c s e n h a n c e d c o m p o n e n t v ol a t ili ty m o d e l a u t h o r s yao, y, z h a i , j, c a o , y, di n g , x, li u , j a n d l u o , y ty p e ar ti cl e u r l t hi s v e r s i o n i s a v a il a b l e a t : h t t p : // u s ir. s a lf o r d . a c . u k /i d / e p r i n t / / p u b l i s h e d d a t e u s i r i s a d i gi t a l c oll e c t i o n of t h e r e s e a r c h o u t p u t of t h e u n iv e r s i t y of s a lf o r d . w h e r e c o p y r i g h t p e r m i t s , f ull t e x t m a t e r i a l h e l d i n t h e r e p o s i t o r y i s m a d e f r e e l y a v a il a b l e o n li n e a n d c a n b e r e a d , d o w n l o a d e d a n d c o p i e d fo r n o n- c o m m e r c i a l p r i v a t e s t u d y o r r e s e a r c h p u r p o s e s . p l e a s e c h e c k t h e m a n u s c r i p t fo r a n y f u r t h e r c o p y r i g h t r e s t r i c ti o n s . f o r m o r e i nf o r m a t i o n , i n cl u d i n g o u r p o li c y a n d s u b m i s s i o n p r o c e d u r e , p l e a s e c o n t a c t t h e r e p o s i t o r y te a m a t : u s i r @ s a lf o r d . a c . u k . mailto:usir@salford.ac.uk data analytics enhanced component volatility model author: yuan yao institute of management science and engineering, business school, henan university, , jinming district, kaifeng, henan province, china prof.yuanyao@gmail.com jia zhai salford business school, university of salford, crescent, salford m wt, uk j.zhai@salford.ac.uk jia.zhai @gmail.com yi cao department of business transformation and sustainable enterprise, surrey business school, university of surrey, guildford, surrey, gu xh, united kingdom jason.caoyi@gmail.com xuemei ding school of computing and intelligent systems, ulster university, magee campus, northland rd, londonderry northern ireland, uk, bt jl x.ding@ulster.ac.uk faculty of software, fujian normal university, upper rd rd, cangshan, fuzhou, fujian province, , china xuemeid@fjnu.edu.cn mailto:prof.yuanyao@gmail.com mailto:j.zhai@salford.ac.uk mailto:jia.zhai @gmail.com mailto:jason.caoyi@gmail.com mailto:x.ding@ulster.ac.uk mailto:xuemeid@fjnu.edu.cn junxiu liu, school of computing and intelligent systems, ulster university, magee campus, northland rd, londonderry northern ireland, uk, bt jl j.liu @ulster.ac.uk yuling luo guangxi key lab of multi-source information mining & security, faculty of electronic engineering, guangxi normal university, diecai, guilin, guangxi, china, yuling @mailbox.gxnu.edu.cn corresponding author: jia zhai salford business school, university of salford, crescent, salford m wt, uk + ( ) j.zhai@salford.ac.uk jia.zhai @gmail.com mailto:j.zhai@salford.ac.uk mailto:jia.zhai @gmail.com data analytics enhanced component volatility model abstract volatility modelling and forecasting have attracted many attentions in both finance and computation areas. recent advances in machine learning allow us to construct complex models on volatility forecasting. however, the machine learning algorithms have been used merely as additional tools to the existing econometrics models. the hybrid models that specifically capture the characteristics of the volatility data have not been developed yet. we propose a new hybrid model, which is constructed by a low-pass filter, the autoregressive neural network and an autoregressive model. the volatility data is decomposed by the low- pass filter into long and short term components, which are then modelled by the autoregressive neural network and an autoregressive model respectively. the total forecasting result is aggregated by the outputs of two models. the experimental evaluations using one-hour and one-day realized volatility across four major foreign exchanges showed that the proposed model significantly outperforms the component garch, egarch and neural network only models in all forecasting horizons. . introduction volatility is considered the “barometer for the vulnerability of financial markets and the economy” (jiang, ahmed, & liu, ) (poon & granger, ) and is crucial for asset pricing, derivative valuation and risk management. volatility modelling and forecasting is a much devoted area of research and have attracted many attentions of researches. a number of econometrics and machine learning models have been developed in past years. conventional models include generalized autoregressive conditional heteroskedastic (garch) model, developed by (tim, ) and (bollerslev, ), was accepted as one of the most popular volatility models for years. at the same time, autoregressive fractionally integrated moving average (arfima) model has been proposed to capture the long memory property in realized volatility (granger c. w., ) (granger & joyeux, ). in , (lee & engle, ) introduced a component garch model, which decomposes volatility into two components: a permanent long run trend component and a transitory short-run one that is mean-reverted to the long-run trend (harris, stoja, & yilmaz, ). after that, more empirical evidences suggested that two-component model can capture the volatility structure much better than the one-component models. for example, (brandt & jones, ) showed that daily volatility can be well characterized by two- component model structure with one highly persistent component and one strongly stationary component. a number of literatures have found that two-component volatility models perform better than the one- component models in explaining stock as well as exchange rate volatilities. started from component garch model, introduced by (lee & engle, ), those two-component models usually consider the volatilities as a composition of a permanent long-run trend component and a transitory short-run component, both of which follow a mean-reverting process with a slow reversion speed in long-run trend and quick reversion speed in the short-run one. additional to the traditional econometrics models, machine-learning algorithms have been widely used in financial modelling (jia, yi, yuan, xuemei, & yuhua, ) (zhai, cao, yao, ding, & li, ) (yi, yuhua, sonya, ammar, & martin, ) in recent years. (boyacioglu & avci, ) proposed an adaptive neural fuzzy inference system (anfis) for predicting earning per shares on istanbul stock market and concluded that their method performed well on monthly forecasting. this anfis model was also effectively applied in predicting closing price of zagreb stock market index (svalina, galzina, lujić, & Šimunović, ). two hybrid models were proposed in (hajizadeh, seifi, zarandi, & turksen, ), where explanatory input variables were selected based on garch, egarch and gjr-garch models and were fed into neural network model for volatility forecasting. the experimental evaluations showed that the hybrid models outperforming traditional models effectively. a hybrid model of self-organized fuzzy neural network and arima model has been proposed in (mcdonald s. , coleman, mcginnity, & li, ) and applied on financial markets. this hybrid model was thoroughly compared with other traditional forecasting models in (mcdonald s. , coleman, mcginnity, li, & belatreche, ). the results showed that hybrid model achieved better forecasting results in average. (kristjanpoller, fadic, & minutolo, ) developed a hybrid model composed of neural network and garch model, where the volatility is modelled and forecasted by garch model at the first step and the output of the garch as well as the original volatility data were then fed into a neural network model for forecasting the volatility. it showed that the hybrid model significantly outperformed the traditional arfima and garch model. the hybrid model was extended in (kristjanpoller & minutolo, ) for spot and future oil price and showed % increase on precision over previous models. among the structures of most hybrid models, artificial neural network is widely used for modelling the non-linear part of the underlying variable. artificial neural network is also widely used in time series forecasting in recent studies (rubio, elias, cruz, & pacheco, ) (aljarah, faris, mirjalili, & al- madi, ) (rubio, ) (restrepo, manotas, & lozano, ) (jesús, ) (liu, et al., ). those studies show that careful feature selection and certain neural network structure with tailor-made algorithms can reach a stable, convergent, and better accuracy when compared with the traditional neural network for the prediction of time-series in different areas, i.e., mechatronic processes, brain signal, and some other biomedical data. (kristjanpoller, fadic, & minutolo, ) and (kristjanpoller & minutolo, ) further show that particularly designed hybrid model composed of the artificial neural network and traditional garch models achieves significantly better performance than the traditional models, i.e., garch, egarch, and arfima in forecasting the volatility of foreign exchange and commodity future. the hybrid models are trained by carefully selected features that are related to the volatility. in addition, the work in (sharda & patil, oct ) (gorr, june ) (zhang & qi, january ) show that autoregressive neural network (arnn) outperforms other structure of artificial neural network in trend and seasonality forecasting of a non-linear time series, where the trend is defined as removing the short-term components from the raw time series and the seasonality is defined as removing the long-term components. (zhang & qi, january ) and (nelson, hill, remus, & o'connor, ) show that prior data processing, removing either the trend or the seasonality, can dramatically reduce forecasting errors and is critical to build an adequate arnn based forecasting model. inspired by previous work of two-component models in econometrics area and hybrid models in machine learning area, we propose an alternative, very simple-structured model to capturing and forecasting volatility across short and long forecasting horizons. our measure of volatility is based on the realized volatility. we decompose the volatility to long-term and short-term components using a low-pass hodrick-prescott filter, and then model the long-term component using an autoregressive neural network and the short-term component as a stationary autoregressive process around the long-term component. the structure of a low- pass filter, a first order autoregressive process and an autoregressive neural network is simple but effective to capture the dynamics of two components of the realized volatility. since neural network is one of the most popular data analytics algorithms, we name it as “data analytics enhanced component volatility model”. we evaluate the model’s out-of-sample forecasting performance based on the one-hour and one-day realized volatility constructed using eur/usd, gbp/eur, gbp/jpy and gbp/usd high frequency exchange rates over the period september to august , which includes around . billion observations across days. as a benchmark, we compare the forecasting accuracy of the proposed model with those of the two-component garch model (lee & engle, ) and the traditional egarch model as well as the neural network only model. we evaluate the performances up to time points, which are days for one- hour volatility and roughly two years for one-day volatility. consistent with the findings of (lee & engle, ) and (brandt & jones, ), our experimental evaluations also show that in almost all the cases, our proposed model provides a significant improvement in forecasting performance over component garch and egarch as well as the neural network only model. the improvement is stable across short and long horizons. in particular, the forecasting accuracy of our proposed model is roughly consistent across all horizons while the performance of the component garch and egarch model deteriorate significantly in long horizon forecasting. overall, the experimental evaluations show that our proposed model achieves a stable and much better forecasting performance than most of the traditional volatility models. the outline of the remainder of this paper is as follows. in section , we present the structure and details of the proposed model. section describes the data used in the empirical analysis and the forecasting evaluation criteria. section presents the empirical results and section provides a summary, concluding remarks and some future research directions. . methodology . model structure in this paper, we follow this idea and the model format in (lee & engle, ), we assume the realized volatility follows a two-component process given by σt = 𝐿𝑡 + 𝑆𝑡 ( ) 𝑆𝑡 = 𝛼𝑆𝑡− + εt ( ) where 𝐿𝑡 and 𝑆𝑡 are the long and short term component of the realized volatility respectively and εt is the random error term with zero mean and constant variance. the short term component 𝑆𝑡 is an autoregressive, ar( ), process with the parameter 𝛼, which measures the speed of the short term revision to long term trend. this structure of the volatility is following the two-component characteristics in previous literatures, for example (lee & engle, ), (alizadeh, brandt, & diebold, ), and (brandt & jones, ). however, the implementation of our two-component model is different from traditional component garch model. we implement the two-component model given in equation ( ) and ( ) in three steps. firstly, we decompose and extract the long-term component 𝐿𝑡 from the realized volatility using low-pass hodrick-prescott filter (hodrick & prescott, ). after getting 𝐿𝑡, the short term component of the realized volatility can be obtained by 𝑆𝑡 = σt − 𝐿𝑡. in the second step, we train an artificial neural network (ann) using the long-term component 𝐿𝑡 to obtain a forecasting model. a future value of 𝐿𝑡+𝑛 at time 𝑡 + 𝑛 can be forecasted by the trained ann. in the last step, the ar( ) process model 𝑆𝑡 = 𝛼𝑆𝑡− + εt is estimated using the short term component 𝑆𝑡 , which is obtained from the first step. a future value of 𝑆𝑡+𝑛 at time 𝑡 + 𝑛 can be then forecasted by the estimated ar( ) model. therefore, we can calculate the future value of the realized volatility at time 𝑡 + 𝑛 by 𝜎𝑡+𝑛 = 𝐿𝑡+𝑛 + 𝑆𝑡+𝑛. . volatility decomposition in the first step, the long-term component 𝐿𝑡 is extracted from the realized volatility. to do this, the low-pass hodrick-prescott filter (hodrick & prescott, ) is applied to extract a low frequency non-linear component from a time-series. that low frequency component represents the trend of the long-term component of the realized volatility. hodrick-prescott filter (hodrick-prescott filter) is widely used in applied macroeconomics (stock & watson, ) (mcelroy, ) (stock & watson, ) for removing the short-term cyclical component of a time series from raw data. given the value of the smoothing parameter 𝜆, the long term trend component shall solve min 𝜏 (∑(𝑦𝑡 − 𝜏𝑡) + 𝜆 ∑[(𝜏𝑡+ − 𝜏𝑡) − (𝜏𝑡 − 𝜏𝑡− )] 𝑇− 𝑡= 𝑇 𝑡= ) where the smoothing parameter 𝜆 penalizes the variations in the growth rate of the trend component. the larger the value 𝜆, the higher is the penalty. we follow the studies in (baxter & king, ) (ravn & uhlig, ) (harris, stoja, & yilmaz, ) to set the 𝜆 as the widely used empirical value of multiplied by the squared frequency of the data, which for one hour realized volatility (assuming trading days per year and hours per day) is 𝜆 = ∗ ( ∗ ) = . . autoregressive neural network . . network structure in the second step of implementing our two-component model, we use an autoregressive neural network (arnn), a special format of artificial neural network, to model and forecasting the long-term component 𝐿𝑡 of realized volatility. as the discussion in section , the work in (sharda & patil, oct ) (gorr, june ) (zhang & qi, january ) show that arnn outperforms other structure of artificial neural network in trend and seasonality forecasting of a non-linear time series. (zhang & qi, january ) and (nelson, hill, remus, & o'connor, ) further show that prior data processing, removing either the trend or the seasonality, can dramatically reduce forecasting errors and is critical to build an adequate arnn based forecasting model. therefore, in our paper, we follow the existing study and employ the arnn to model and forecasting the trend, the long-term component 𝐿𝑡 of realized volatility. the reason of selecting arnn is twofold. first, main strand finance literature usually assumes the volatility is composed of two autoregressive process, long term and short term. we follow the traditional assumption with augmented adjustment: arnn, which models the lags as a nonlinear function. secondly, compared with other artificial neural network, arnn can achieve even better accuracy in “deseasonalized” financial time-series forecasting by a relatively simple structure, where appropriate lags instead of a number of additional features is crucial for the forecasting performance. arnn is also proved to have advantages over recurrent feed-forward neural network and is less sensitive to the problem of long-term dependence (mustafaraj, lowry, & chen, ). compared with traditional feed-forward artificial neural network, the arnn has interconnection from the lagged input to the output layer, which enhances its capability of forecasting more than one-step as a time- series predictor. we follow the traditional arnn structure: 𝑌𝑡 = 𝛼 + ∑𝛼𝑖𝑌𝑡−𝑖 𝑛 𝑖= + ∑Ψ(𝛾 𝑗 + ∑𝛾𝑖𝑗𝑌𝑡−𝑖 𝑛 𝑖= )𝛽𝑗 ℎ 𝑗= + 𝑡 where 𝑛 is the number of lags, ℎ is the number of hidden neurons, Ψ(.) is the activation function, 𝛾 is the weights between input and hidden neurons, 𝛽 is the weights between hidden and output neurons. the non- linear part contains ℎ hidden neurons transforms the input variables, defined as the 𝑛-lagged long-term component 𝐿𝑡 of realized volatility, weighted by parameters 𝛾𝑖𝑗 plus a bias 𝛾 𝑗, via a non-linear activation function Ψ(.). if representing the number of lags and hidden neurons as 𝑖 and 𝑗, a hidden neuron can be denoted by Ψ(𝛾 𝑗 + ∑𝛾𝑖𝑗𝑌𝑡−𝑖 𝑛 𝑖= ) each of which is weighted by a parameter 𝛽𝑗 before it produces the output layer. since we want to model and forecasting the long-term component of volatility, arnn operates as a non-linear regression function at the end as 𝑌𝑡 = 𝐺(𝑌𝑡− ,𝑌𝑡− ,…,𝑌𝑡−𝑛) + 𝑡,𝐴𝑅𝑋𝑁𝑁 which maps the unknown relation, 𝐺(∙), between the input variable and the target function 𝑌𝑡 and the error term 𝑡,𝐴𝑅𝑋𝑁𝑁. for the purpose of this study, the hidden layer uses the hyperbolic tangent sigmoid transfer function, while the output layer uses a linear transfer function. the arnn structure with only one hidden layer is considered since it operates as a non-linear regression function and can be trained to approximate most non-linear function arbitrarily well (siegelmann, horne, & giles, ), (andreou, charalambous, & martzoukos, ) (mustafaraj, lowry, & chen, ). following this structure, in this paper, to forecasting long-term component 𝐿𝑡 of the realized volatility, an autoregressive neural network (arnn) with three layers structure is used: an input layer that includes lagged 𝐿𝑡 inputs to the network; a hidden layer with hyperbolic tangential activation function, and an output layers with a linear activation function. in previous studies, the estimation of the model parameters usually follows the back-propagation algorithm. it is shown in (charalambous, june ) that the back-propagation algorithm is often unable to converge rapidly to the optimal solution. therefore, we utilize the modified levenberg-marquardt (lm) algorithm, which is much more sophisticated and efficient in terms of time capacity and accuracy (hagan & menhaj, ). when training the arnn, we divide the training dataset into three subsets, % of the data for training the arnn, % for validation, and the last % for testing. during the training process, the errors on the training dataset and validation dataset are monitored at the same time. when the validation error rises while the training error maintains, the arnn begins to overfit the data. the weights and bias at the minimum of the validation error are saved as the trained arnn . . network parameters under this structure of arnn, the number of lags 𝑛 and the number of hidden neurons ℎ are the crucial parameters for constructing the arnn model. we follow the widely used configuration for the arnn in (siegelmann, horne, & giles, ) and (mustafaraj, lowry, & chen, ), and investigate the performance using the number of lags from to , which is 𝑛 = , , , . we firstly use ℎ = as the preliminary configuration for the number of hidden neurons and finds the appropriate lags. we use the long- term component 𝐿𝑡 of one hour realized eurusd rate from sep to dec to train the arnn model under 𝑛 = , , , and use long-term component values of eurusd rate from dec to jan to test the trained model. the forecasting error is defined as �̂�𝑡 − 𝐿𝑡, the difference between the original value 𝐿𝑡 and the forecasted one. the average errors across the forecasting horizons is listed in table , where we can observe that the best average error is at 𝑛 = , although the errors do not show a significant difference among different configuration of lag numbers. four lags also conform to the configuration suggested in (siegelmann, horne, & giles, ) and (mustafaraj, lowry, & chen, ). therefore, in this paper, we use four lags autoregressive neural network (arnn) following the suggested configuration and our empirical investigation. table this table contains the average forecasted error across the horizon from dec to jan under configurations of lag number from , , and 𝑛 = 𝑛 = 𝑛 = 𝑛 = average error . e- . e- . e- . e- the following error! reference source not found.figure shows an example of using arnn model to forecasting the long-term component 𝐿𝑡 of one hour realized eurusd rate from dec to jan . the arnn model was constructed using one hidden layer with neurons and was trained using the 𝐿𝑡 of eurusd rate from sep to dec . the forecasting results �̂�𝑡 include long-term component values of eurusd rate from dec to jan . the forecasting error, defined as �̂�𝑡 − 𝐿𝑡, between the original value 𝐿𝑡 and the forecasted one �̂�𝑡 are shown in the bottom sub-figure, from which we can observe that out-of-sample forecasting errors are around − . figure this figure shows an example of out-of-sample forecasting long term component 𝑳𝒕 of eurusd using arnn. the top figure shows the long term component of eurusd from dec to feb decomposed by hodrick-prescott filter. the middle figure shows the forecasted value �̂�𝒕 of the long term component of eurusd from dec to jan using arnn model. the bottom figure shows the differences �̂�𝒕 − 𝑳𝒕 between the original long term component of eurusd and the forecasted one. . autoregressive model in the third step, we estimate an autoregressive model for the short-term component 𝑆𝑡: 𝑆𝑡 = 𝛼𝑆𝑡− + εt. the n-step ahead forecasting of the short-term component 𝑆𝑡 is therefore given by �̂�𝑡+𝑛 = 𝛼 𝑛𝑆𝑡 + ∑ 𝛼 𝑖𝑛− 𝑖 εt−i ( ) we can obtain the forecasted the long-term component �̂�𝑡+𝑛 of the realized volatility through arnn model. therefore, the n-step ahead forecasting of the one hour realized volatility is given by equation ( ) as σ̂t+n = �̂�𝑡+𝑛 + �̂�𝑡+𝑛 . we name the proposed model as autoregressive neural network enhanced two-component model (shorted as: nne c) due to the implementation of the model structure. to estimate the autoregressive model for the short-term component 𝑆𝑡, we utilize the method of moment through yule–walker equations, named for udny yule and gilbert walker (yule, ) (walker, ): 𝛾𝑚 = ∑ 𝛼𝑘𝛾𝑚−𝑘 𝑝 𝑘= + 𝜎𝜖 𝛿𝑚, where 𝛾𝑚 is the autocovariance function of 𝑆𝑡, 𝜎𝜖 is the variance of the input noise process and the 𝛿𝑚, is the kronecker delta function. for the case 𝑝 = , one lag autoregressive process ar( ), the 𝛼 can be obtained by 𝛾 /𝛾 . . model flow we summarize the nne c model workflow when applied in realized volatility forecasting. different from other hybrid models, which usually use the neural network as the dominate one to model the non-linear part of the underline financial variable, i.e. volatility or foreign exchange and use the arima (mcdonald s. , coleman, mcginnity, li, & belatreche, ) or garch model (kristjanpoller, fadic, & minutolo, ) as the pre-processing part for modelling the linear part of the underline financial variable, the nne c model follows the traditional financial theory to consider the volatility as a combination of the long and short term while producing an enhanced mechanism through explicitly decomposing the two components and modelling them separately. the simple nne c model follows and enhances the financial theory by capturing the volatility structure explicitly and effectively. algorithm workflow of proposed nne c model given the realized volatility σt of selected financial security, i.e. fx, we assume it is composed of long and short term components: σt = 𝐿𝑡 + 𝑆𝑡 we use hodrick-prescott filter as a low-pass filter to explicitly extract the long term component 𝐿𝑡. the smoothing parameter 𝜆 of hodrick-prescott filter is selected according to the empirical study: 𝜆 = ∗ ( ∗ ) = . min 𝜏 (∑(σt − 𝜏𝑡) + 𝜆 ∑[(𝜏𝑡+ − 𝜏𝑡) − (𝜏𝑡 − 𝜏𝑡− )] 𝑇− 𝑡= 𝑇 𝑡= ) the short term component 𝑆𝑡 is then obtained by 𝑆𝑡 = σt − 𝐿𝑡 modelling 𝐿𝑡 and 𝑆𝑡 simultaneously: using 𝐿𝑡 to train the arnn with lags and hidden neurons, 𝑛 step ahead 𝐿𝑡+𝑛 is forecasted by the trained arnn model; 𝐿𝑡 = 𝛼 + ∑𝛼𝑖𝐿𝑡−𝑖 𝑛 𝑖= + ∑ Ψ(𝛾 𝑗 + ∑𝛾𝑖𝑗𝐿𝑡−𝑖 𝑛 𝑖= )𝛽𝑗 ℎ 𝑗= + 𝑡 using 𝑆𝑡 to estimate the ar( ) model, 𝑛 step ahead 𝑆𝑡+𝑛 is forecasted by the estimated ar( ) model; 𝑆𝑡 = 𝛼𝑆𝑡− + εt 𝑛 step ahead realized volatility is obtained by σt+n = 𝐿𝑡+n + 𝑆𝑡+n . data and forecasting evaluation . data we use the neural network enhanced two-component model defined in section to forecasting the volatility of the eur/usd, gbp/eur, gbp/jpy and gbp/usd exchange rates. high frequency exchange rate data (tick data) were obtained from oricode inc for the period september to august and included around . billion observations across days. the unobserved true volatility, in principle, can be estimated arbitrarily accurately using a measure of realized volatility calculated through the intraday returns (harris, stoja, & yilmaz, ). it is proved by (torben, tim, & francis, ) that the sum of squared intraday returns converges to the unobserved true volatility with the intraday interval approaching to zero. in this study, we use realized volatility as the proxy of the unobserved true volatility. this is obtained by aggregating the intraday squared returns using the approach given by (andersen & bollerslev, ): �̂�𝑟𝑣,𝑡 = ∑ 𝑟𝑡,𝑛 𝑁 𝑛= where �̂�𝑟𝑣,𝑡 is the realized volatility for time 𝑡, and 𝑟𝑡,𝑛 is the squared log return on time 𝑡 for interval 𝑛 (𝑛 = , ,… ,𝑁). in this paper, we use one-hour and one-day realized volatilities, both of which are constructed from millisecond log return, which is the highest frequency in our data. the one-hour realized volatility is obtained by aggregating , -millisecond log returns 𝑟𝑡,𝑛 in each hour. the data from september to december ( observations) is used for the initial estimation of the model, while the data from december to august ( ) is used for out- of-sample evaluation. the one-day realized volatilities is constructed by the sum of , , log returns in each trading day. for the one-day realized volatilities, the data from september to december ( observations) is used for the initial estimation of the model, while the remained data ( observations) is used for out-of-sample evaluation. table reports summary statistics for the one-hour and one-day realized volatilities of full observations of four exchange rates. panel a reports the mean, standard deviation, skewness and excess kurtosis and bera– jarque statistic and panel b reports the first six autocorrelation coefficients and the ljung–box q statistic for autocorrelation of six lags for the realized volatilities. p-values are also reported in parentheses. in the ljung– box q tests, the null hypothesis that the residuals of the returns are not autocorrelated is rejected in both one- hour and one-day realized volatilities. therefore the two realized volatilities are all highly autocorrelated. table summary statistics and autocorrelations mean standard deviation skewness excess kurtosis bera-jarque panel a: summary statistics gbp/usd hr . e- . e- . e+ . e+ . e+ gbp/jpy hr . e- . e- . e+ . e+ . e+ gbp/eur hr . e- . e- . e+ . e+ . e+ eur/usd hr . e- . e- . e+ . e+ . e+ gbp/usd daily . e- . e- . e+ . e+ . e+ gbp/jpy daily . e- . e- . e+ . e+ . e+ gbp/eur daily . e- . e- . e+ . e+ . e+ eur/usd daily . e- . e- . e+ . e+ . e+ panel b: autocorrelations ljung–box q gbp/usd hr . . . . . . . e+ ( . ) gbp/jpy hr . . . . . . . e+ ( . ) gbp/eur hr . . . . . . . e+ ( . ) eur/usd hr . . . . . . . e+ ( . ) gbp/usd daily . . . . . . . e+ ( . e- ) gbp/jpy daily . . . . . . . e+ ( . ) gbp/eur daily . . . . . . . e+ ( . ) eur/usd daily . . . . . . . e+ ( . ) in the example in the following figure (a-b), hodrick-prescott filter is applied to one hour realized volatility with smoothing parameter 𝜆 of (blue curve in figure (a)) of eurusd exchange rate from sep to dec to extract the long term component 𝐿𝑡 (red curve in figure (b)). figure (c-d) shows an example of one-day realized volatility with long term component extracted by hodrick-prescott filter with smoothing parameter 𝜆 of (assuming trading days per year, the 𝜆 is calculated by multiplied by the squared frequency of the data). (a) (b) (c) (d) figure (a) blue curve in the figure: one-hour realized volatility of eurusd exchange rate from sep to dec ; red curve: the long term component 𝑳𝒕of the realized volatility extracted by low-pass hodrick-prescott filter; (b) short term component of one-hour realized volatility of eurusd exchange rate; (c) blue curve in the figure: one-day realized volatility of eurusd exchange rate from sep to dec ; red curve: the long term component 𝑳𝒕of the realized volatility extracted by low- pass hodrick-prescott filter; (d) short term component of one-day realized volatility of eurusd exchange rate. . forecasting evaluation the proposed neural network enhanced two-component model is used to calculate out-of-sample forecastings of the realized volatilities of up to hours or days ahead across the evaluation period for the one-hour or one-day realized volatility respectively. as the benchmark, we select one-factor egarch and two-factor component garch of (lee & engle, ). in addition, we also employ four lags autoregressive neural network applied directly on realized volatility as one of the benchmark models. we therefore estimate four models: a) four lags autoregressive neural network enhanced two-component model (nne c); b) one- component egarch model; c) two-component garch model; and d) four lags autoregressive neural network model (nnonly). the four models are evaluated using one-hour and one-day realized volatilities respectively. for one-hour and one-day data, four models are initially estimated using the first , and , observations from sep to dec respectively and then the volatilities at different estimation periods are calculated. following (michael & christopher, ), we consider forecasting horizons of , , , , , and hours or days ahead for one-hour or one-day volatility respectively. we choose the root mean square error (rmse) with respect to the true realized volatility as the measure of the forecasting performance: 𝑅𝑀𝑆𝐸 = [ 𝑇 ∑ (𝜎𝑡(𝜏𝑡,𝜏𝑡+𝑇) − �̂�𝑡(𝜏𝑖,𝑡,𝜏𝑡+𝑇)) 𝑇 𝑡= ] ( ) for the shorter evaluation horizons of , , , we calculate the rmse over the forecasting horizon, i.e. (𝜏𝑡,𝜏𝑡+𝑇) = ( , ),( , ) and ( , ). for the three longer horizons, we calculate the rmse over time points, i.e. (𝜏𝑡,𝜏𝑡+𝑇) = ( , ),( , ) and ( , ). for the one-hour realized volatility, the rmse is to evaluate the average performance over days ( trading hours in each day). for the one-day realized volatility, the rmse is to evaluate the average performance over month ahead. . results . model parameter in theory, three-layer autoregressive neural network can approximate most of the functions as long as a sufficient number of hidden neurons is provided. in this paper, we construct the nne c with , , and neurons in the hidden layer and test the forecasting performance of the proposed model. the results are illustrated in table and table for one-hour and one-day realized volatility data respectively. for the comparison purpose, we also included the forecasting results with hidden neurons in table and table . from table , it is very clear that for the nne c, the forecasting accuracies are roughly the same by using different number of hidden neurons. the forecasting accuracy does not increase with the number of hidden neuron increases. this result is consistent in all forecasting horizons across four currencies. the only case that the forecasting accuracy rises with the increase of the number of hidden neurons is highlighted in table as gbp/usd rate at ( , ) horizon. however, the accuracy increase is as subtle as around e- . for all the cases in table , the forecasting accuracy differences by different number of hidden neurons are also tiny. the only three cases that the forecasting accuracy increases (although tiny) with the rise of the number of hidden neurons are highlighted in table : gbp/jpy rate at horizon ( , ), ( , ) and gbp/eur at horizon ( , ). in addition, the forecasting accuracies of nnonly model with different number of hidden neurons do not show significant differences in table and table as well. our results also conform to the previous researches in (kristjanpoller & minutolo, ) and (kristjanpoller, fadic, & minutolo, ). therefore, based on our experimental evaluations in table and table , we believe our proposed nne c can achieve the best forecasting performance under the structure of three-layer autoregressive neural network with hidden neurons. table forecasting performance of one-hour realized volatilities. this table reports the root mean square error (rmse) for the autoregressive neural network enhanced two-component model constructed by one hidden layer with and neurons (nne c) and autoregressive neural network model constructed by one hidden layer with , and neurons (nnonly). 𝜏 𝜏 nne c ( neurons) nne c ( neurons) nne c ( neurons) nnonly ( neurons) nnonly ( neurons) nnonly ( neurons) eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- table forecasting performance of one-day realized volatilities. this table reports the root mean square error (rmse) for the autoregressive neural network enhanced two-component model constructed by one hidden layer with and neurons (nne c) and autoregressive neural network model constructed by one hidden layer with and neurons (nnonly) 𝜏 𝜏 nne c ( neurons) nne c ( neurons) nne c ( neurons) nnonly ( neurons) nnonly ( neurons) nnonly ( neurons) eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e+ . e+ . e+ eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- eur / usd . e- . e- . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . e- . e- . experimental results following the configurations, we employ our experiments using nne c with lags and neurons. table and table report the rmse of the one-hour and one-day realized volatilities given by equation ( ) for four models over six forecasting intervals for four currencies respectively. overall, for all models in the experiments in table and table , the rmse measures fall at the first with the forecasting horizon and then rise for four currencies. this is due to the reason that initially the forecasting interval increases from five to and then to time points. after the horizon (𝝉𝒕,𝝉𝒕+𝑻) = (𝟏,𝟏𝟎𝟎), the forecasting interval is fixed at time points, and therefore the forecasting error rises as the horizon increases. in one-hour realized volatility evaluations in table , the nne c model shows the highest forecasting accuracy in out of cases (four currencies with each having horizons) followed by cgarch model. the only exception is the highlighted case in table : cgarch achieved rmse of . e- at horizon (𝟏𝟎𝟎,𝟐𝟎𝟎) on gbp/eur and was more accurate than the nne c, which had . e- rmse at the same case. at horizon of (𝟏,𝟏𝟎𝟎), the performance of cgarch model is lower but close to the performance of the nne c in all four currencies. in other horizons, the nne c is significantly more accurate than cgarch model. particularly, the decline of the forecasting accuracy of the cgarch model as well as egarch and nnonly model is obvious from horizon (𝝉𝒕,𝝉𝒕+𝑻) = (𝟏𝟎𝟎,𝟐𝟎𝟎) to (𝟐𝟔𝟎,𝟑𝟔𝟎) and then to (𝟒𝟎𝟎,𝟓𝟎𝟎). it shows that at longer forecasting horizon, the nne c has sufficiently stronger forecasting capability than the traditional models. table forecasting performance of one-hour realized volatilities. this table reports the root mean square error (rmse) for the autoregressive neural network enhanced two-component model constructed by one hidden layer with neurons (nne c neurons), component garch model, egarch model, and autoregressive neural network model (nnonly), for the forecasting interval 𝝉𝟏 to 𝝉𝟐, where 𝝉𝟏 = 𝒎𝒂𝒙(𝟏,𝝉𝟐 − 𝟏𝟎𝟎), for the one-hour realized volatilities of four currencies. 𝜏 𝜏 nne c ( neurons) cgarch egarch nnonly ( neurons) eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- in table , the nne c was significantly more accurate than other three models in all cases. in the highlighted cases in table , which include gbp/eur rate at ( , ) and ( , ) horizons, and gbp/eur rate at ( , ) horizon, cgarch model achieved the forecasting accuracies lower but close to the performance of the nne c. in other cases, the nne c remarkably outperforms all other models. it is worth noting that the nnonly model performs the worst in all cases while the nne c achieves the significantly accurate performance. our results conform to that of (zhang & qi, ) and (nelson, hill, remus, & o'connor, ), which concluded that neural network is not able to model volatility directly but neural networks built with deseasonalized data could produce significantly more accurate forecasting than with non-deseasonalized data. table forecasting performance of one-day realized volatilities. this table reports the root mean square error (rmse) for the autoregressive neural network enhanced two-component model constructed by one hidden layer with neurons (nne c neurons), component garch model, egarch model, and autoregressive neural network model constructed by one hidden layer with neurons (nnonly neurons), for the forecasting interval 𝝉𝟏 to 𝝉𝟐, where 𝝉𝟏 = 𝒎𝒂𝒙(𝟏, 𝝉𝟐 − 𝟏𝟎𝟎), for the one-day realized volatilities of four currencies. 𝜏 𝜏 nne c ( neurons) cgarch egarch nnonly ( neurons) eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e+ eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- eur / usd . e- . e- . e- . e- gbp / eur . e- . e- . e- . e- gbp / jpy . e- . e- . e- . e- gbp / usd . e- . e- . e- . e- . conclusions the fact that volatility comprises both a long-term trend component and a strongly oscillation short-term component has crucial implications for modelling and forecasting volatility over both short and long horizons. in this paper, we develop a simple but effective volatility-forecasting model. the model is based on a decomposition of intraday realized volatility into the long and short-term components using the low- pass hodrick-prescott filter. the three-layer autoregressive neural network with hidden neurons models the long-term component and the short-term component is modelled as a simple autoregressive process. therefore, we name the proposed model as “neural network enhanced two-component volatility model”. the model was thoroughly evaluated using high frequency tick data of four currencies across six forecasting horizons. the out-of-sample forecasting results consistently and significantly outperform the component garch and egarch models as well as the autoregressive neural network model, which is applied on modelling the volatility directly. the results reported in this paper are based on two simple structures: the low-pass hodrick-prescott filter uses a fixed smoothing parameter and the short-term component follows a first order autoregressive progress. future work would be to consider using an optimized smoothing parameter that decomposes a stationary short-term component. a higher order arma process would provide a better fit for the decomposed short- term component and may bring improved out-of-sample forecasting performance. moreover, a higher smoothing parameter brings more smoothed long-term component, which can be easier to forecasting by neural network with higher accuracy, and decomposes a more volatile short-term component, which may not follow a stationary process. indeed, it would be a sufficiently further improvement if considering the forecasting model as an optimization framework to find an optimal parameter, which decomposes a stationary short-term process while maintaining a smooth long-term component, which can be forecasted by autoregressive neural network with high accuracy. acknowledgement this research is partially supported by the scientific research funds for the returned overseas chinese scholars from state education ministry, nature and science funds from fujian province, china (grant no. j ), and the funds for young key program of education department from fujian province, china ( grant no. jz ) . references alizadeh, s., brandt, m. w., & diebold, f. x. ( ). range‐based estimation of stochastic volatility models. journal of finance, - . aljarah, i., faris, h., mirjalili, s., & al-madi, n. ( ). training radial basis function networks using biogeography-based optimizer. neural computing and applications, - . andersen, t., & bollerslev, t. ( ). answering the skeptics: yes, standard volatility models do provide accurate forecasts. international economic review, - . andreou, p. c., charalambous, c., & martzoukos, s. h. ( ). pricing and trading european options by combining artificial neural networks and parametric models with implied parameters. european journal of operational research, – . baxter, m., & king, r. g. ( ). measuring business cycles approximate band-pass filters for economic time series. review of economics and statistics, - . bollerslev, t. ( ). modelling the coherence in short-run nominal exchange rates: a multivariate generalized arch model. the review of economics and statistics, - . boyacioglu, m. a., & avci, d. ( ). an adaptive network-based fuzzy inference system (anfis) for the prediction of stock market return: the case of the istanbul stock exchange. expert systems with applications, – . brandt, m., & jones, c. ( ). volatility forecasting with range-based egarch models. journal of business and economic statistics, – . charalambous, c. (june ). conjugate gradient algorithm for efficient training of artificial neural networks. iee proceedings g - circuits, devices and systems, - . gorr, w. l. (june ). research prospective on neural network forecasting. international journal of forecasting, ( ), - . granger, c. w. ( ). long memory relationships and the aggregation of dynamic models. journal of econometrics . , - . granger, c. w., & joyeux, r. ( ). an introduction to long‐memory time series models and fractional differencing. journal of time series analysis, - . hagan, m. t., & menhaj, m. b. ( ). training feedforward networks with the marquardt algorithm. ieee transactions on neural networks, - . hajizadeh, e., seifi, a., zarandi, m. f., & turksen, i. ( ). a hybrid modeling approach for forecasting the volatility of s&p index return. expert systems with applications, – . harris, r. d., stoja, e., & yilmaz, f. ( ). a cyclical model of exchange rate volatility. journal of banking & finance, - . hodrick, r. j., & prescott, e. c. ( ). postwar us business cycles: an empirical investigation. journal of money, credit, and banking, - . jesús, r. j. ( ). a method with neural networks for the classification of fruits and vegetables. soft computing, – . jia, z., yi, c., yuan, y., xuemei, d., & yuhua, l. ( ). coarse and fine identification of collusive clique in financial market. expert systems with applications, – . jiang, y., ahmed, s., & liu, x. ( ). volatility forecasting in the chinese commodity futures market with intraday data. review of quantitative finance and accounting, - . kristjanpoller, & minutolo. ( ). forecasting volatility of oil price using an artificial neural network- garch model. expert systems with applications, , – . kristjanpoller, w., fadic, a., & minutolo, m. ( ). volatility forecast using hybrid neural network models. expert systems with applications, – . lee, g. g., & engle, r. f. ( ). a permanent and transitory component model of stock return volatility. in r. engle, & h. white, cointegration, causality, and forecasting: a festschrift in honor of clive w. j. granger (pp. - ). oxford: oxford: oxford university press. liu, q., yin, j., leung, v. c., zhai, j.-h., cai, z., & lin, j. ( ). applying a new localized generalization error model to design neural networks trained with extreme learning machine. neural computing and applications, – . mcdonald, s., coleman, s., mcginnity, li, y., & belatreche, a. ( ). a comparison of forecasting approaches for capital markets. ieee conference on computational intelligence for financial engineering & economics (cifer). london: ieee. mcdonald, s., coleman, s., mcginnity, t. m., & li, y. ( ). a hybrid forecasting approach using arima models and self-organising fuzzy neural networks for capital markets. the international joint conference on neural networks (ijcnn). dallas: ieee. mcelroy, t. ( ). exact formulas for the hodrick-prescott filter. econometrics journal, – . mustafaraj, lowry, & chen. ( ). prediction of room temperature and relative humidity by autoregressive linear and nonlinear neural network models for an open office. energy and buildings, – . nelson, m., hill, t., remus, w., & o'connor, m. ( ). time series forecasting using nns: should the data be deseasonalized first? journal of forecasting , – . poon, s.-h., & granger, c. w. ( ). forecasting volatility in financial markets: a review. journal of economic literature, ( ), - . ravn, m. o., & uhlig, h. ( ). on adjusting the hodrick–prescott filter for the frequency of observations. review of economics and statistics, - . restrepo, a. r., manotas, d. f., & lozano, c. a. ( ). self-generation of electricity, assessment and optimization under the new support schemes in colombia. ieee latin america transactions, - . rubio. ( ). least square neural network model of the crude oil blending process. neural networks, – . rubio, j. d., elias, i., cruz, d. r., & pacheco, j. ( ). uniform stable radial basis function neural network for the prediction in two mechatronic processes. neurocomputing, – . sharda, r., & patil, r. b. (oct ). connectionist approach to time series prediction: an empirical test. journal of intelligent manufacturing, – . siegelmann, h., horne, b., & giles, c. ( ). computational capabilities of recurrent narx neural networks. ieee transactions on systems, man, and cybernetics, part b (cybernetics), - . stock, j. h., & watson, m. w. ( ). forecasting inflation. journal of monetary economics, – . stock, j., & watson, m. ( ). dynamic factor models, factor-augmented vector autoregressions, and structural vector autoregressions in macroeconomics. in j. b. taylor, & h. uhlig, handbook of macroeconomics (pp. volume , – ). elsevier. svalina, i., galzina, v., lujić, r., & Šimunović, g. ( ). an adaptive network-based fuzzy inference system (anfis) for the forecasting: the case of close price indices. expert systems with applications, – . tim, b. ( ). generalized autoregressive conditional heteroskedasticity. journal of econometrics . , - . torben, a., tim, b., & francis, d. ( ). parametric and nonparametric measurements of volatility. in y. aït-sahalia, & l. p. hansen, handbook of financial econometrics. amsterdam: north-holland. w, b. m., & s, j. c. ( ). volatility forecasting with range-based egarch models. journal of business & economic statistics, - . walker, g. ( ). on periodicity in series of related terms. proceedings of the royal society of london, – . yi, c., yuhua, l., sonya, c., ammar, b., & martin, m. ( ). adaptive hidden markov model with anomaly states for price manipulation detection. ieee transactions on neural networks and learning systems, - . yule, g. u. ( ). on a method of investigating periodicities in disturbed series, with special reference to wolfer's sunspot numbers. philosophical transactions of the royal society, – . zhai, j., cao, y., yao, y., ding, x., & li, y. ( ). computational intelligent hybrid model for detecting disruptive trading activity. decision support systems, – . zhang, g., & qi, m. ( january ). neural network forecasting for seasonal and trend time series. european journal of operational research, – . zhang, p., & qi, m. ( ). neural network forecasting for seasonal and trend time series. european journal of operational research, – . pii: - ( ) -x computers math. applic. vol. , no. , pp. - , - / $ . + . printed in great britain. all rights reserved copyright © pergamon press pie electronic circuit diagnostic expert systems--a survey y. lmov at&t bell laboratories, holmdel, nj , u.s.a. (recewedll may ) abstract--the electronic circuit diagnostic problem is roughly formulated and subdivided into six subproblems. current literature and patents are surveyed with respect to the above six subproblems. some of the existing expert diagnostic systems as well as expert diagnostic shells are described and their limitations are outlined. a review of the relevant terms from ai is included. the bibliography list contains some references. . i n t r o d u c t i o n this paper discusses artificial intelligence methods in circuit pack testing. testing requires investment o f labor, capital and information about testing techniques. the added cost o f testing in the circuit pack production line is justified by several reasons: first, the defective products are not shipped; second, testing helps to develop product repair methodology; and, finally, testing helps to discover manufacturing problems, e.g. low yield due to a single fault may occur because o f malfunctioning insertion robot, a bad batch o f components, or a badly designed in-circuit test. the interested reader should refer to towill ( ) and williams et al. ( ) for overviews o f automatic testing. .i. f a u l t c a t e g o r i e s defects in circuit packs, in general, come from three sources: bad components, damage to components or the circuit board, and mistakes in the manufacturing process. since "different things go wrong at different times because o f different reasons" (pynn, ), determining an optimal manufacturing and test strategy becomes a challenging task. roughly speaking, one can define three different fault categories: device faults (e.g. unoperative capacitor) assembly faults (e.g. solder shorts, missing components, misoriented components, etc.) and operational faults, such as timing problems arising from apparently good components that fail to work together. according to this somewhat arbitrary fault categorization, a two-stage test strategy has been adopted by most circuit pack manufacturers: the inspection test stage and the functional test stage. inspection test is designed to examine the circuit pack for proper construction and it consists of shorts test, in-circuit test, and visual test. the in-circuit test examines the function o f each device separately from others. functional (system) test is designed to examine the operation o f the pack and thus tests the function o f assemblies o f components. as the complexity o f the circuit pack grows, the role o f testing becomes more important and the testing process becomes more difficult. the area o f design for testability is becoming a popular topic by necessity (williams and parker, ). with the advent of computer technology, automatic test equipment--the equipment for programmed application o f sequences o f stimuli, measurement, and comparison o f the results to the expected values---became the accepted norm. the issue o f software development for such automatic test equipment becomes both increasingly important and difficult as the complexity o f the circuit packs grows. . . a u t o m a t e d d i a g n o s t i c s the ability o f software to generate stimuli and to draw conclusions about the defective components for a given circuit pack is a subtle issue dependent on the experience o f the test engineer responsible for programming o f the automatic test equipment. if the test is to perform two basic ca~ava t s/*--e y. lirov tasks, i.e. verification (knowing whether everything is functioning as expected), and diagnosis (identifying what causes the unexpected behavior, so it can be fixed) then it is the diagnosis which becomes an increasingly difficult task as a result o f the growing complexity of the circuit packs. . . importance o f diagnostics diagnosis o f an electronic circuit usually refers to the process of determining the faulty component(s) that cause an undesired behavior (output) o f the given circuit for some (correctly) given input. importance o f diagnosis stems from both practical and theoretical considerations. . . . practical considerations. the practical considerations are primarily of an economic nature: there is a monetary value associated with every component in the pack, as well as a monetary value associated with the processes of assembly and soldering of the pack. thus, when a pack is declared faulty, it is desirable to replace only the faulty components. moreover, the process o f identification o f the faulty components (the diagnostics) consists o f a series o f tests each of which has an associated cost, expressed in such parameters as test setup time, component destruction, etc. the process o f identification o f the faulty components and their subsequent replacement should not be overly expensive. thus, observation : diagnostics o f circuit packs is primarily an economic problem (barber and yau, ). currently this problem is solved in the majority o f cases by humans who are called test analysts or diagnosticians. training o f such human analysts is expensive. their performance depends on the accumulated experience. indeed, observation : there are human experts that are extremely successful in the area o f diagnostics. therefore, it is very desirable to have computer programs which are capable o f performing diagnostics. however, observation : diagnostics problem (i.e. the construction of a computer algorithm which is capable to perform diagnostics) is np-hard (hyafil and rivest, ; ibarra and sahni, ; karp, ). . . . theoretical considerations. therefore, in the absence o f a better alternative, we resort to the construction o f programs that perform approximate diagnostics. in other words, we are willing to give up some o f the precision o f the final description of the faulty component(s) (solution) in order to gain in the length of time spent to obtain that solution. moreover, in many cases, in the absence o f rigorous mathematical methods to simulate the complex decision making processes which are performed by the human diagnosticians, we are forced to invent different heuristics in order to obtain in some sense similar diagnostic performance. therefore, the development of such programs is related to the field o f computer science which is commonly known by the name o f artificial intelligence. this relation is further extended to the fields o f control systems, since, as we shall show later, a diagnostics problem is also a planning or control problem. moreover, approximate solution approaches often employ probabilistic and fuzzy reasonings, thus relating the research in diagnostics to operations research and information theory. the time complexity considerations in the diagnostic and related algorithms requires applications and further developments o f the results belonging to the field o f computational complexity. and, o f course, the area o f application is related to electrical engineering. diagnostics belongs to the intersection of all o f the above fields. hence the theoretical importance o f diagnostics. . . current approaches two basic approaches and their combinations are currently utilized in industry for conventional diagnostic software development: the fault dictionary approach and the guided probe approach. the fault dictionary approach requires simulation of the faulty behavior of the system and subsequent storage o f the simulation results (as well as the fault assumptions) in the fault dictionary. a "misbehaved" output o f the pack under test can therefore be looked up in the fault dictionary. basically, two problems prevent the fault dictionary approach from being the widely accepted testing method. these are the overly extensive memory requirements and the diversity o f the faults. as a result, the fault dictionaries are usually incomplete and ambiguous. the guided probe approach requires only simulation o f the correctly functioning circuit pack. the basic tool in utilizing this approach is the blame-shifting mechanism applied after each electronic circuit diagnostic expert systems---a survey measurement. the upstream (from the test point view) components in the circuit are blamed only when the measurements do not agree with the expected results. the efficient sequencing o f measurements becomes the maii: problem when implementing the guided probe approach. . . outline o f the overview in order to appreciate and put into perspective the contributions o f the earlier published results on diagnostics, it is convenient to view them with relation to different aspects o f the diagnostics problem. such a perspective not only allows us to compare different approaches but also provides us with the opportunity to identify yet unsolved problems. the rest o f the paper is organized as follows: first we state the diagnostics problem in a more formal way so as to allow its subdivision into six subproblems. then we give a brief overview o f each of the subproblems and the tools that are used in their solution. . p r o b l e m s t a t e m e n t a circuit pack (plant) p = p(s, b) is modeled by its structure s and its behavior b. structure s = s(c, t) is defined by the set c of components and topology t on the set c. behavior b =b(io, r) is defined by two mathematical models: ( ) i model describing the input--output relationships o f every c e c and ( ) r model describing the reliability o f every c e c . a component c ~ c is called faulty if its observed behavior bo is different from be, the expected io behavior (be e io). problem (diagnostic problem). given a multicomponent plant p and a symptom indicating an anomaly on the expected io behavior, identify the subset cf c c o f faulty components. however, the identification part o f the above problem is itself a complex process. in particular it involves repetitive generation o f hypotheses and their support or disposal. proving or rejecting a hypothesis requires in turn to perform sequences o f measurements and subsequent sequences of symbolic computations. therefore, the identification requirement of problem can be satisfied by recursively solving the following: problem (classification problem). given a symptom and the results of previously performed measurements, obtain an estimate o f the faulty subset cf (generate hypothesis), and problem (planning problem). given an estimate o f cf, obtain the minimum expected cost plan (sequence) of measurements, so that when performed, the initial hypothesis will be either proved or rejected. note . the actual reasoning technique employed to prove or reject a hypothesis, although an important issue by itself, serves only a secondary role. diagnostic expert system exploits diverse sources of information to detect failures in a given unit under test (uut). the relevant information is concerned with the function of the different components, their interconnection, signal characteristics and their paths, and component reliability measures. moreover, since usually the initial information is insufficient to diagnose correctly, a number o f additional observations (tests) is needed to complete the diagnosis. the description o f available tests, their diagnostic value and costs is yet another kind o f information necessary for diagnosis. knowledge representation is the term used for the schemes o f storing the different kinds o f information in order to allow for its efficient processing, i.e., failure detection and cost-efficient test sequence generation. knowledge has to be adequately represented for quick manipulation, for ease in modification, for specifying distinctions between different concepts, and for viewing of multiple levels o f abstractions. knowledge representation is inherently related to the ways the knowledge is used. different problem-solving strategies which use the knowledge base define the inference engine o f the diagnostic expert system. problem (reasoning problem). find an efficient way o f reasoning about electrical circuits. (such reasoning involves manipulation o f data structures which represent both structural and behavioral data.) y. l m o v note . both problem and problem involve requirements to reason along both structural and behavioral lines. such computational requirements require novel approaches for solution of problem (representation problem). find an efficient way of representing plant p = p(s, b). efficiency of representation is determined by its level of completeness, consistency, transparency, and computability (winston, ). when solving problem by using heuristic approaches and in particular by utilizing knowledge of human experts, one can not overestimate the importance of problem (knowledge engineering problem). find efficient ways to extract the heuristics about circuit pack troubleshooting from human experts. and, when interfacing to analysts at the shop floor level, one must also face problem (user interface problem). find efficient ways to conduct a fruitful dialog with the human analysts performing diagnostics. we have identified six problems which belong to the field of artificially intelligent testing. now we turn to the approaches and tools which are being used to solve those problems. it is more convenient first to address the problems of representation and reasoning, then the planning and classification problems, and, finally the issues about knowledge engineering and user interfaces. . r e p r e s e n t a t i o n and r e a s o n i n g p r o b l e m s basically two approaches and their combinations have been used to solve diagnostics problems: the rule-based (shallow modeling) approach and the conceptual (deep) modeling approach. many diagnostic systems have their knowledge in the form of rules, representing separate and modular chunks of knowledge, e.g. "observations =~ hypothesis". often such knowledge may be contra- dictory, incomplete, or too large to be manipulated efficiently. in such cases one models the domain or system under consideration and reasons with this model to generate diagnostics. such models have been called causal, conceptual, or deep models. . . performance notions the choice of the correct approach or their combination depends on such factors as size and complexity of the diagnosed system, its reliability on the one hand, and its criticality and stability on the other hand. system size is usually defined in terms of the number of line replaceable units (lru). system complexity reflects the size of the effort required to track down a fault and it is a function of the number of feedback loops, number of outputs and inputs to lrus, etc. reliability of the diagnosed system is defined as the cumulative reliability of its lrus. note that the more reliable the system is the more difficult is the diagnostic task. the criticality of the diagnosed system is defined in terms of the length of time interval length within which the diagnosis must be completed. stability of the diagnosed system is defined by the number of modifications to the system per unit of time. . . rule-based appoach the rule-based approach can be further subdivided into deterministic decision trees, single relevant rule, and multiple relevant rule methods. . . . deterministic decision tree. the deterministic decision tree method is based on the construction of a hierarchy of questions in such a way that each combination of answers can be interpreted as a leaf on the decision tree. the main advantage of this approach is in its simplicity and in its suitability to the systems of high criticality. however, this approach is limited to the systems of low complexity and high stability. . . . single relevant rule. the single relevant rule approach is based on searching the rule base for the single rule which can be applied in a given situation or can confirm a given hypothesis. since such search procedures can be applied recursively, they have been called forward and backward chaining respectively. the advantages and disadvantages of the single relevant rules are similar to the ones of the deterministic decision trees. however, when applying this approach one must be aware of the requirements to maintain a valid rule base, i.e. no two rules may apply to the same situation (consistency), and each situation has to be covered by a rule (completeness). electronic circuit diagnostic expert systems---a survey optimality o f the rule base is obtained by ordering the rules in such a way that the expected time to search for the relevant rule is minimized. an entropy-based approach is usually utilized to obtain the optimal ordering o f the rule-base. sometimes a cost is associated with the firing action o f the rule or/and with the satisfaction o f the antecedent o f the rule (i.e. query o f the user, test setup time, etc.). in such cases the optimality o f the rule-base becomes a difficult problem. the requirements o f consistency, completeness, and optimality are very difficult to satisfy in advance when dealing with systems o f low stability and high complexity. . . . multiple relevant rules. to resolve the above problem, the multiple relevant rules approach has been developed by several authors. this approach requires application o f a heuristic to evaluate the chaining strength of the rule on a particular situation as well as a heuristic to combine the strengths o f several rules when forward or backward chaining is involved. a basic expert system building tool (a shell) has been recently patented (hardy et al., ) by teknoledge. the tool includes interactive knowledge base debugging, question generation, legal response checking, explanation, and certainty factors. (its source code in prolog takes pages and is available in the patent documentation.) however the expert systems based on this approach and having several hundreds o f rules again exhibit great difficulties in maintaining their knowledge base. also the multiple relevant rules approach does not contribute to the problem o f knowledge acquisition for the systems o f low stability. . . conceptual modeling approach conceptual modeling approach takes into consideration the original system description rather than using the knowledge compiled by humans into a decision tree or a set o f rules. the conceptual modeling approach can also be subdivided into normal functioning modeling and fault functioning modeling. the overall diagnostic strategies and their relationships were described in chandrasek- aran and milne ( ), who used a hierarchy o f four levels: structural, behavioral, functional and pat- tern matching. given structural representation (list o f components and their connectivity; forbus, ; bylander and chandrasekaran, ) using qualitative simulation and consolidation pro- posed methods to generate the behavioral description. to generate a functional description o f the de- vice using the behavioral description, one may use the approach o f de kleer which is based on teleo- logical (device intentions based) reasoning. the pattern matching diagnostic strategy can be applied after compiling (by a human expert or by a machine) the functional description o f the device. deep (structural, behavioral and functional) modeling can also be used to declare the innocence o f certain components o f the unit under test (uut; scarlet al., ), or to explain the observed malfunction (kuipers, ). using de kleer's ( ) approach, given a low level electronic description o f the components, one is able to deduce the outputs o f the circuit, and consequently diagnose faults in it. davis ( ) proposed computing the function and the inverse o f each component and then utilizing them to propagate o f combination o f inputs to compute the outputs or vice versa. digital troubleshooting becomes possible by comparing the expected values with the computed ones. a formal language for u u t description has been developed by (chandrasekaran, ) who compiles the above description into a set o f production rules which are used for diagnosis. m ilne ( ) uses structural reasoning at both the fault isolation to a functional area and possible faults proposal stages. using his "responsibility theory" appoach only four basic diagnostic rules are needed: two rules to propose faults using description o f structure and another two using description o f function. forbus ( ) develops a qualitative reasoning methodology to derive the diagnostic rules given a structural description o f objects and their topology. bylander and chandrasekaran ( ) propose a methodology to develop the description o f behavior o f composites o f components based only on the descriptions o f the components. . p l a n n i n g p r o b l e m ooel ( ) has shown that the number o f tests needed to pinpoint a fault is at worst linear in the number o f components. however, as we observed earlier, generating the best test sequence is np-complete (ibarra and sahni, ). y. lmov a test generation algorithm which uses the design information has been proposed some years ago by roth et al. ( ) and it is called the d-algorithm. the d-algorithm is based on activating simultaneously all the paths to the observable point in the system. such an approach is impractical for nowadays circuits because o f its high computational requirements. an algorithm o f cantone ( ) for deciding which test to perform is based on maximizing the ratio o f information gained and expected cost, computed over the available subcircuits. the algorithm is based on the gamma-miniaverage method (slagle and lee, ). a combi- nation o f structural information (topology o f components) and compiled knowledge (cost o f test and components' failure rates) is used to further isolate the fault to a single functional area. the d a r t algorithm (bennett et al., ) exploits the heirarchy of the circuit in order to deal with the issue o f exponential growth in the number o f computations. within each level d a r t deduces the suspects by reasoning backward to find a justification for a symptom and generates additional tests by working forward from a behavioral rule for one o f the suspects to observable outputs. for more on planning see also allen et al. ( ), de kleer ( ), hyafil et al. ( ), ibarra et al. ( ), konolige et al. ( ), loveland ( ), mcdermott ( , ), poage et al. ( ), rosenchein ( ), sacerdoti ( ), shirley ( ), slagle et al. ( ), vere ( ) and warren ( ). . c l a s s i f i c a t i o n p r o b l e m the diagnostic mode o f reasoning is a term used by kim and pearl ( ) to describe the inference process o f updating the causal model o f beliefs due to modifications in the model o f symptoms. kim and pearl ( ) describe a computational model for causal and diagnostic reasoning which is a generalization o f the bayesian methods previously applied to decision trees (ddi, ). . . b a y e s i a n m e t h o d s diagnostic reasoning requires construction o f means to assign and update a blame measure to parts o f the model, based on the observed behavioral deviations o f the physical system. the classical bayesian probabilistic approach suffers from the following shortcomings: ( ) the need to establish a p r i o r i probability o f success o f a test; ( ) the need to establish a priori failure rates of the components; ( ) the approach is only applicable when the events are mutally exclusive; and ( ) conditional probabilities for every combination o f evidence must be computed since the independence o f evidence cannot be assumed. the dempster-shafer (shafer, ) theory o f evidence has been proposed to remedy some o f the above problems. the theory still suffers from shortcomings , and . recently thompson and wojcik o f westinghouse patented a " m e t h o d and apparatus for system fault diagnosis and control" (thompson and wojcik, ). a domain-independent set of metarules is constructed which is able to build a rule network through which belief is propagated to detect defects. . . m u l t i p l e -fault a s s u m p t i o n the multiple-fault assumption poses an additional difficulty in diagnostics problems. peng and reggia ( ) developed a probability based approach to guide the backward chaining in such a way that the more likely situations are considered first. by only using the input/output description o f the components and the connectivity information, de kleer ( ) developed a procedure to incrementally accept or rule-out possible faults. see also edgar et al. ( ). fuzzy logic applications for failure analysis were tried by kitowski and ksazek ( ). the applied method involves solution o f fuzzy equations derived from a failure-symptom fuzzy relations matrix. sensor-based systems for realtime monitoring present particularly stringent requirements for expert diagnostic systems. a rule-based system utilizing multiple sensors for obtaining data about chemical parameters was recently patented (kemper et al., ; westinghouse electric). the system is utilized to monitor a steam turbine-generator power plant. electronic circuit diagnostic expert systems--a survey . . l e a r n i n g learning o f diagnostics systems has been explored only at the compiled knowledge base level reasoning: pazzani ( ) developed a failure-driven learning mechanism to update the rule-base when a diagnostic failure occurs. . knowli~dge e n g i n e e r i n g problem problem (knowledge engineering) has been addressed in cheeseman ( ), who considered the problem o f expert systems learning from data. in particular he proposed a method for extracting information from data to form the knowledge base for a probabilistic expert system. a rule acquisition method for a diagnostic expert system has been recently filed for european patent (alexander et al., ) by tektronix. the method includes a grammar for acceptable patterns which are automatically converted to rules. the rule acquisition process includes parsing o f an input sentence as it is received and guiding the expert through the sentence creation. the system is implemented in small talk and prolog. . user i n t e r f a c e problem problem (user interface) is considered in simmons ( ) and taie et al. ( ) who addressed the issues o f graphical and animated representation o f functioning o f devices. see also richer and clancey ( ) and brown et ai. ( ). . o p e r a t i n g i m p l e m e n t a t i o n s several diagnostic expert systems have been developed in industry during the last few years. a review o f the capabilities and limitations o f some o f the completed projects is given to exemplify the theoretical discussion o f the applicable methods given above. advisr (cooper et al., ) is a rule-based expert system developed by aai corporation, running on ibm pc/at, working in conjunction was atlas test program software, and designed to indicate faults on a navy a n / u s m - u u g (v) automatic tester. a single-fault assumption is made as well as the assumption about fully-reliable connectivities. ads (magliero et al., ) was developed by harris corporation as part o f the navy integrated diagnostic support system. ads integrates several knowledge representation paradigms such as pre-computed fault-trees, design topology, logical parameters historical records, and production rules. test selection is done by maximizing the "expected" entropy to cost ratio. the system is implemented in ada. fcmds (bursch et al., ) is developed by honeywell inc., in lisp. fcmds uses frame representation for its knowledge base which is subdivided into causal model kb and test kb. the causal model includes the design toplogy. the test base contains the basic procedural knowledge for performing tests. test selection is done similarly as in ads. test interpretation procedures involve proprietary algorithms to manipulate failure likelihood measures attached to lrus. this allows for multiple-fault assumption. a p u m a i d (mccown et al., ) is developed by allied-signal aerospace company (bendix). it has a temporal, event-based description o f the diagnostic target system, the system is viewed as a sequence o f functional events composed o f location and context, e.g. time and causality. a failure model is obtained as an event-based description. the rest o f the system is implemented in a rule-based fashion. the rules describe the relationships among events, components and parameters. fis (pipitone, ) is developed at the nrl's center for applied research in ai. it is based on the fault model which is a qualitative causal model composed o f a block diagram o f the u u t combined with a set o f causal rules for each block. processing hypothesized test failures through the fault model produces an "ambiguity" set o f possibly faulty components. a rule-based test cost generation is used to evaluate the variable costs o f the tests. the best test is chosen similarly to the ads. it is written in lisp and runs on sun. y. lmov ai-ferret (maguire, ) is developed at hughes aircraft company. it is written in interlisp-d and loops and runs on xerox ai workstation. it is basically similar to fis and it has been successfully applied to diagnose the tow missile system. the resulting rule base contains rules while the connectivity model contains over functions (modes). haidex (firdman, ) is another diagnostic expert system developed at hughes. it is based on the conceptual model of the normally functioning system and the set o f deviations from the normal function. the inference engine generates a troubleshooting strategy by decomposing the u u t into its subsystems. it is developed in zetalisp and runs on symbolics . m y c i n (shortlife, ) is one o f the first expert diagnostic systems designed at stanford to recognize bacterial infections in blood samples. it is based on purely rule-based approach and it contains about rules. each rule has a confidence factor associated with it and the inference engine is provided with an a d h o c computation scheme to combine the factors when using backward chaining. other m y c i n and medical applications related papers include adams ( ), benbasset et al. ( ), ben-bassat et al. ( ), bjerregard ( ), blois ( ), bosyj ( ), cantazarite et al. ( ), chandrasekaran ( ), chandrasekaran et al. ( , ), charniak ( ), cooper ( ), elstein et al. ( ), flehinger ( ), gomez et al. ( ), gorry ( ), kolodner et al. ( ), kulikowski ( , ), lipkin et al ( ), miller et al. ( ), mittal ( ), mittal et al. ( ), patil et al. ( ), patrick et al. ( ), regia ( ), regia et al. ( ), rubin ( ), szolovitz ( ) and weiss et al. ( ). some other expert systems ( a f h r l , ) include a r b y / n d s ( d u c k ) for communications networks troubleshooting, ace for preventive maintenance o f telephone cables, les (frame-based with about rules) for electronic maintenance, stamp, i d t (shubin et al., ), crib (hartley, ) and c r i t t e r (kelly et al., ). . system b u i l d i n g t o o l s the core o f an expert system has a knowledge base and an inference engine that operates on the knowledge base. user interface is considered at two different levels: when developing a system, and when deploying it. some o f the important concepts are summarized below. . . k n o w l e d g e representation there are three important aspects o f knowledge representation which must be available in any diagnostic system building tool. these are: object descriptions, actions, and certainties. the most convenient way to represent objects is by frames which are tabular data structures (records). a frame consists o f slots which are filled with either data about objects or relations to other frames. a built-in inheritance mechanism allows to inherit or override the information from other frames. actions are used to modify the relevant database. actions are usually implemented by rules or procedures. the degree or correctness o f data and/or knowledge is usually expressed by means o f belief functions, certainty factors, or probabilities. . . i n f e r e n c e approach the inference approach in a typical diagnostic expert system may be a combination o f forward and backward chaining, hypothetical reasoning, and blackboard mechanism. . . i. chaining. backward chaining is a method o f evaluating hypothesized conclusions to see whether they are supported by the evidence. it is usually implemented via i f - t h e n rules, starting with rules that have the hypothesized conclusions as their consequents. the set o f rules is then searched for those in which the antecedents include the previous conclusions. this process (backward chaining) is continued until the hypothesis is proved. if the p r o o f is unachievable, then some other hypothesis may be tried. forward chaining starts with the present data with is matched with the antecedents until a relevant rule is found. the consequent o f that rule determines the new present data and the process o f matching is repeated until a desired conclusion is reached. electronic circuit diagnostic expert systems--a survey . . . hypothetical reasoning. hypothetical reasoning (truth maintenance) refers to the solution in which hypotheses have to be assumed when initializing the search (e.g. forward or backward chaining.) nonmonotonic reasoning refers to the ability o f retracting the consequences o f the assumptions which have been found untrue during the search process. different tools handle nonmonotonic reasoning by using viewpoints, contexts or worlds, or nonchronological backtrack- ing. hypothetical reasoning literature includes doyle ( ), and mcdermott ( , ). the blackboard mechanism (hayes-roth, ) refers to a common data structure shared by a group of cooperating expert systems. an agenda often is used to control the development o f the solution on the blackboard. this mechanism is especially useful when several competing solutions are developed in parallel. . . . pattern matching. pattern matching is usually required in two mechanized reasoning instances: when matching rule atecedents to the current knowledge situation to fire the best applicable rule, and when matching partially filled frames to the existing knowledge base to determine the current knowledge. . . expert system shells a diagnostic expert system development tool is required to have the following capabilities: ( ) knowledge representation: structured rules, certainty factors, frames; ( ) inference engine: forward chaining, backward chaining, backward chaining, blackboard-like, truth maintenance, pattern matching; ( ) developer interface: kb editor, menu, check for consistency, graphics representation o f kb, graphics utilities to build end user interface, debug; and ( ) interface to general-purpose programming language (e.g. c). unfortunately, none o f the available products in the current market have all o f the above features, although all o f the products have many features in addition to the above list. the only products that come close to our requirements are art, kee and knowledge craft (see expert system shells list). however, the cost o f these tools is in the range o f $ , -$ , , besides additional expensive training. . . diagnostic tools two commercially available special purpose diagnostic expert system tools have been developed recently, namely, in-ate, and ai-test. we give a short outline o f their characteristics. . . . i n - a t e . in-ate (cantone et al., ) provides an excellent menu-driven syntax for specifying fault diagnosis expert rules. it also allows specification o f the signal flow and uses it when searching for the defective component. in additon it has frames-based knowledge representation and a mechanism to perform probabilistic reasoning. the system does have an interface to c, and it does have capabilities for automated testing the knowledge base consistency. it runs however only on a macintosh computer. the system does not have a separate electronics theory and measurements related data-base. therefore, each test point must have all the relevant information. thus, some duplication of information and difficulties in knowledge acquisition are inevitable. . . . ai-test. ai-test (ben-bassat et al., ) is developed by intelligent electronics and runs on ibm-pc/at. ai-test is comprised o f two knowledge bases---the universal knowledge base which contains general information relating to electronics theory, kinds o f measurements (dc, frequency, etc), signal types, etc. and the uut-specific knowledge base which includes the block diagram o f the u u t and tests descriptions. the inference engine interprets the test results (by updating the model o f beliefs) and chooses the next best test. no mechanism is provided to update the initial failure rates o f the components. also, the choice o f the appropriate kind o f test at a specific test point must be performed by the user. an ai system should be able to suggest the most informative test description automatically. the user interface has been improved greatly since the last release. . c o n c l u s i o n s new circuit pack manufacturing technologies (e.g. surface mount) contribute further to the complexity o f the circuit pack and thus make the testing even more difficult. for example a new camwa / ---f y. lmov approach in electronic design is the built-in-self-test (bist) approach (agarwal, ). this approach has the positive features o f testing the equipment at its normal operation speed, as well as running the test in parallel on all the devices. thus a malfunctioning device can in principle declare its faulty condition faster and seemingly reduce the difficulty o f the diagnostic task. .i. new challenges as happened before, the new technology has posed new testing challenges. in particular, bist technology requires separate independent testing o f the built-in test equipment. failing to test the built-in hardware may result in two kinds o f errors: a correctly functioning device declaring itself as faulty and thus being replaced for the wrong reason (such repeated mistakes will later be propagated to mistakes in the process control) and a misbehaving device declaring itself as being healthy, thus postponing the troubleshooting to later stages and therefore increasing the manu- facturing and maintenance costs. . . drawbacks o f the rule-based approach since the success o f mycin, the rule-based approach has been implemented several times in circuit pack diagnostics. the mechanisms that proved useful in medical applications were automatically applied in the electronics industry. however, in this industry only part o f the knowledge (general electronics laws) is static, while the rest o f the knowledge changes with every new design and with every new batch o f components. moreover, the usually available statistical data is not present any more. one the other hand, another kind o f data (connectivity and signal paths) is available. these differences must be recognized and utilized (e.g. in-ate and ai-test) in order to develop a successful expert system. . . classification and planning problems the diagnostic problem is still far from being resolved and it is currently an area o f active research and development. in particular in the area o f the classification problem there is a lack o f computationally efficient schemes to propagate the results o f tests. a prevailing assumption is the "perfect info" assumption underlying the construction of most o f the diagnostic expert systems. the mechanisms for dealing with partial, unreliable, and contradictory results are mainly based on the dempster-shaffer theory. the conditional probabilities (correlation data), which are usually available in medical applications are not generally available in electronic circuit pack diagnostics applications. thus, the applicability o f the "classical" belief propagating schemes is questionable. this issue requires a separate investigation in light of dealing with structural circuit description in addition to rules. the issue of dealing with noisy data also requires a specialized effort. in the area o f the planning problem there is lack o f mathematically sound basis for the best test selection. most of the existing systems either do not address the issue o f test planning or use an ad hoc method. computational efficiency is another issue which must be addressed separately. . . deep knowledge representation and reasoning although it is now quite obvious that a purely rule-based approach is not practical, there are not many pragmatic ways to implement a deeper kind o f knowledge besides the connectivity o f the circuit pack. the area o f knowledge representation and reasoning currently receives most of the ai community's attention. yet it is still unclear to what extent qualitative reasoning and functional knowledge can be used for practical applications. consequently all o f the existing diagnostic expert systems implementations are using a combination o f rule-based and connectivity knowlege with a mechanism for belief functions manipulations and an ad hoc (entropy vs cost based) best test selection heuristic. thus, efficiency o f such systems is difficult to assess. . . learning and user interface the least amount o f work is done in the areas o f learning and user interface. it seems that the most desirable approach to "engineer" knowledge is via a mechanized interview o f the design engineers. experiments have not been performed, nor have tools been constructed to achieve such goals. electronic circuit diagnostic expert systems--a survey . . more open problems only one system (ai-test) has implemented a separate global electronics knowledge base containing generally true facts about electronic circuits behavior, possible tests, etc. however, there is no known technique to match the best test to its proposed location and to its time occurrence. this test matching problem solution could provide an important aid in diagnostic expert system construction. and finally, applications o f new technologies (e.g. neural networks) have yet to be examined. acknowledgements--the author is grateful to on-ching yue and patricia wirth for many constructive suggestions that improved the readability of this paper. special thanks are due to nancy perine for her help in collecting the bibliography. a p a r t i a l l i s t o f e x p e r t s y s t e m s h e l l s . art, inference corporation, w. century bird, la, ca , u.s.a. . kee, intelli corp, e camino real west, mountain views, ca , u.s.a. . knowledge craft, carnegie group, commerce court, station square, pittsburgh, pa , u.s.a. . picon, lisp machine, tech drive, andover, ma , u.s.a. . s. . m. , teknoledge, embarcadero rd, p.o. box , palo alto, ca , u.s.a. . kes, software architecture and engineering, wilson blvd, suite arlington, va , u.s.a. . nexpert object newron data corp., high st., palo alto, ca , u.s.a. , b i b l i o g r a p h y adams, j., a probabilistic model of medical reasoning and mycin model. math. biosci. , - ( ). addis, t. r., towards an "expert" diagnostic system. cl tech. j. , - ( ). addis, t. r. and r. t. hartley, a fault finding aid using a content addressable file source. icl research and advanced development centre, technical report / , stevenage, u.k. ( ). ahuja, s., artificial intelligence environment of the analysis and classification of errors in discrete sequential processes, ph.d.dissertation, univ. maryland, baltimore ( ). aikins, j., prototypes and production rules: a knowledge representation for computer consultations. ph.d. dissertation, rep. no. stan_csd_ - , stanford university ( ). alexander, j., m. freiling, b. phillips and s. messick, rule acquisition for expert systems. european patent application . ( nov. ). ali, m. and d. a. scharnhorst, knowledge-based avionics for interactive fault assessment. wattec : th a. energy conf. knoxville, ten. (february ). ali, m. and d. a. scharnhorst, sensor-based fault diagnosis in a flight expert system. proc. ieee nd conf. artificial intelligence applications, miami, fla., (december ). andersen, s. k. and m. woldbye, knowledge representations for diagnostics and test planning in the domain of electromyography. th european conf. artificial intelligence proc., vol. i, pp. - ( ). andow, p. k., real time analysis of process plant alarms using a minicomputer. comput. chem. engng. , - ( ). andow, p. k., computer based alarm systems. proc. th ieee conf. trends in on-line computer control systems, iee publication no. , pp. - ( ). andow, p. k., propagation of faults in process plants: a state-of-the-art review. chemical process hazards with special reference to plant design, th int. syrup., icheme symposium series no. , pp. - ( ). apfelbaum, l., an expert system for in-circuit fault diagnosis. ieee proc. int. test conf. ( ). arabian, j. h., user requirements for a dynamic high speed functional tester (dhsft). proc. th automatic test electron. seminar~exhibit. june ( ). armstrong, d. b., a deductive method for simulating faults in logic circuits. ieee trans. comput. c , - ( ). afhrl, artificial intelligence in maintenance. afhl, usaf systems command, t. r. afhrl-tr- - ( ). barber, j. s., private communication. barber, j. and c. w. yau, private communication. barr, a. and e. feigenbaum, the handbook o f artificial intelligence. kaufmann press, calif. ( ). bartlett, s. l., c. cole and r. jain, expert system for visual solder joint inspection. proc. rd conf. a i applications, pp. - ( ). basil, w. and l. felkel, disturbance analysis systems. in human detection and diagnosis o f system failures (ed. j. rasmussen and w. b. rouse). plenum press, new york ( ). benbasset, j. and a. schiffmann, an approach to teaching the introduction to clinical medicine. ann. internal med. ( ), - ( ). ben-bassat, m. et al, pattern-based interactive diagnosis of multiple disorders---the medas system. ieee trans. pattern analysis machine lntell. pami , - ( ). ben-bassat, m., d. ben-arie, i. beniaminy, j. cheifetz and m. klinger, a case study with ai-test: an expert system for electronic troubleshooting. proc. autotestcon ' , pp. - ( ). bennett, j. s. and c. r. hollander, dart: an expert system for computer fault diagnosis. proc. th lnt. joint conf. artificial intelligence ( ) bjerregaard, b., j. holst-christensen, e. kalaja, j. lund.kristensen, f. t. debombal and h. d. horroeks, computer-aided diagnosis of the acute abdomen: a system from leeds used on copenhagen patients. proc. information processing work conf. tc ( ). y. lxgov blois marsden, s., clinical judgement and computers. n. engl. j. med. ( ), - ( ). bosyj, m. a., a diagnostic program for the design of procurement systems. mit project mit technical report ( ). breuer, m. a., functional partitioning and simulation o f digital circuits. ieee trans. comput. c , - ( ). breuer, m. a. and m. abramovici, fault diagnosis based on effect-cause analysis. proc th design automation conf., san francisco, ca (june ). breuer, m. and a. friedman, diagnosis and reliable design o f digital systems. computer science press. new york ( ). brown, a. l., qualitative knowledge, causal reasoning and the localization of failures. mit ai lab technical report ( ). brown, d. c. and b. chandrasekaran, design consideration for picture production in a natural graphics systems. comput. graphics ( ), - ( ). brown, j. s., remarks on building expert systems. proc. th int. joint conf. artificial intelligence, san francisco, ca, pp. - ( ). brown, j. s. and r. r. burton, multiple representations o f knowledge for tutorial reasoning. in representation and understanding (ed. d. bobrow and a. collins). academic press, new york ( ). brown, j. s. and r. r. burton, diagnostic models for procedural bugs in basic mathematical skills. cognitive sci. , - ( ). brown, j. s., r. r. burton and a. g. bell, sophie; a sophisticated instructional environment for teaching electronic troubleshooting (an example of ai in cai). bolt beranek & newman, inc., report ( ). brown, j. s., r. r. burton and a. g. bell, sophie: a setup towards a reactive learning environment. int. j. man-machine stud. , - ( ). brown, j. s., a. collins and g. harris, artificial intelligence and learning strategies. in learning strategies (ed. a. o'neil). academic press, new york ( ). brown, j. s., r. rubinstein and r. r. burton, reactive learning environment for computer assisted electronics instruction. bolt beranek & newman, inc., report ( ). brown, j. s., r. r. burton and j. de kleer, pedagogical, natural language and knowledge engineering techniques in sophie i, ii and ill. in intelligent tutoring systems (ed. d. sleeman and j. s. brown). academic press, new york ( ). burger, l. and e. zobor, on-line alarm analysis in the hierarchical control system of the wwr-sm research factor. nuclear power plant control and instrumentation, vol. , pp. - . vienna. academic press, new york ( ). burns, m. and j. pearl causal and diagnostic inferences: a comparison of validity. organizl behav. hum. perform. , - ( ). bursch, p., j. meisner and k. winegar, a pc based expert diagnostic tool. proc. a u t o t e s t c o n ' , pp. - ( ). cain, d. and e. zebroski, the conceptual design of a power plant safety panel. nucl. engng , - ( ). cantone, r., model-based probabilistic reasoning for electronic troubleshooting. proc. th int. joint conf. artifical intelligence, karlsruhe, germany, august ( ). cantone, r., w. lander, m. marrone and w. gaynor, inate/ : interpreting high level fault modes. automated reasoning corp., sunnyvale, va. ( ). cantone, r., w. b. lander, m. p. marrone, and w. gaynor, automated knowledge acquisition in in-ate using component information and connectivity. sigart newsletter no. ( ). catanzarite, v. and a. greenburg, neurologist--a computer program for diagnosis in neurology. proc. rd symp. computer applications in medical care, ieee, pp. - ( ). chandrasekaran, b., on evaluating ai systems for medical diagnosis. proc. m e d c o m p , catalog no. , ch - , ieee, new york, pp. - ( ). chandrasekaran, b. and t. bylander, functional representations and behaviour composition by consolidation: two aspects of reasoning about devices. sigart newsletter no. ( ). chandrasekaran, b. and s. mittal, deep versus compiled knowledge approaches to diagnostic problem-solving. int. j. man-machine stud. - ( ). chandrasekaran, b., f. gomex, s. mittal and j. w. smith. an approach to medical diagnosis based on conceptual structures. proc. th int. joint conf. artificial intelligence, pp. - ( ). chandrasekaran, b., s. mittal and j. w. smith, radex--towards a computer-based radiology consultant. in pattern recognition in practice (ed. e. s. gelsema and l. n. kanal), pp. - . north-holland, amsterdam ( ). chandrasekaran, b., s. mittal and j. w. smith, reasoning with uncertain knowledge: the m d x approach. in proc st ann. joint conf. american medical information association (amia ), (ed. d. a. b. lindberg et al.) pp. - . masson, new york ( ). chandrasekaran, b., d. d. sharma and d. w. miller, the application of knowledge-based systems to reactor operations. trans. a m spc. ( ). charniak e., the bayesian basis of common sense medical diagnosis. proc. third natn. conf. artificial intelligence, aaai, pp. - ( ). charniak, e. and d. mcdermott, introduction to artificial intelligence. addison-wesley, reading, mass. ( ). cheeseman, p., learning of expert systems from data. proc. a a a i ( ). chen, c. c. j. j. schwermann and g. w. sturm, private communication. chiang, h. y. p. et al., comparison of parallel and deductive fault simulation. eee trans. comput. c , ! - ( ). christenson, j. et al., implementation of an automated analysis system in an operating nuclear power plant. engng des. , - ( ). clancey, w. j., acquiring, representing, and evaluating a competence model of diagnostic strategy. technical report hpp- - , stanford university, stanford, calif. ( ). clancey, w. j., heuristic classification. technical report ksl- - , stanford university, stanford, calif. ( ). clancey, w. j., the epistemology of a rule-based expert system--a framework for explanation. artificial lntell. , - ( ). clancey, w. j., tutoring rules for guiding a case method dialogue. int. j. man-machine stud. , - ( ). clancey, w. j. and r. letsinger, neomycin: reconfiguring a rule-based expert system for application to teaching. proc. th int. joint conf. artificial intelligence, pp. - ( ). clocksin, w. and a. morgan, qualitative control. european conf. artificial intelligence , brighton, england, - july ( ). electronic circuit diagnostic expert systems---a survey cooper, g., nestor: a computer-based medical diagnostic aid that integrates causal and probabilistic knowledge. ph.d dissertation, department of computer science, stanford univ. calif. ( ). cooper, l. and t. o. cooper, a performance evaluation of a prototype expert system. proc. a u t o t e s t c o n ' , pp. - ( ). cross, s. e., an application of qualitative simulation to expert system plan justification. sigart newsletter no. ( ). cross, s. e., towards an expert system architecture for flight domain applications. proc. ieee national aerospace and electronics conf., dayton, ohio. vol. , pp. - ( ). cunningham, p., j. gleeson, s. hakiel and m. wheatley, diagnostic heuristics and perspectives. in research and development o f expert systems, vol (ed. m. a. bramer), pp. - ( ). davis, r., amplifying expertise with expert systems. in the a i business (eds p. h. winston and k. prendergast). mit press, cambridge ( ). davis, r., applications of meta-level knowledge to the construction, maintenance and use of large knowledge bases. ph.d thesis, stanford university [stan-cs- - , hpp- - ] ( ). davis, r., diagnosing via causal reasoning: paths of interaction and the locality principal. proc. aaai- , washington, d.c. ( ). davis, r., diagnosis based on description of structure and function. proc. am. associal. artificial intell. , pp. - ( ). davis, r., expert systems: where are we? and where do we go from here? a i mag. ( ), - ( a). davis, r., diagnosis reasoning based on structure and behavior. mit ai lab memo, no. , june ( ). davis, r., diagnostic reasoning based on structure and behavior of digital hardware. computer october, - ( ). davis, r., diagnostic reasoning based on structure and behavior. artificial lntell. , - ( ). davis, r., metarules: reasoning about control. artificial intell. , - ( ). davis, r., qualitative reasoning about physical systems. mit press, cambridge ( ). davis, r., reasoning from first principles in hardware troubleshooting. int. j. man-machine stud. , - ( ). davis, r. and h. shrobe, the hardware troubleshooting group. sigart newsletter no. ( ). davis, r. and h. shrobe, representing structure and behavior. artificial intell. , - ( ). davis, r. and h. shrobe, representing structure and behavior of digital hardware. ieee comput. october, - ( ). davis, r., h. shrobe, w. hanscher, k. wiechert, m. shirley and s. polit, diagnosis based on description of structure and function. proc. am. assoc. artificial intell. % ( ). dempster, a., a generalization of bayesian inference. j. r. statist. sci. serv. b. , - ( ). dhir, v. k. et al., pressurized water reactor (pwr) system and simulation and disturbance analysis for anomalous transients and degraded system conditions. proc. winter simulation conf., ieee, new york ( ). doyle, j., a truth maintenance system. artificial intell., , - ( ). doyle, j., analysis by propagation of constraints in elementary geometry problem solving. mit ai lab working paper ( ). doyle, j., truth maintenance systems for problem solving. mit ai lab. technical report ( ). duda, r., j. g. gaschnig and p. f. hart, model design in the prospector consultant system for mineral exploration. in expert systems in the microelectronic age (ed. d. michie), pp. - . edinburgh university press, edinburgh, scotland ( ). duda, r. o., p. e. hart, p. barrett, j. g. gaschnig, k. konolige, r. reboh and j. slocum, development of the prospector consultation system for mineral exploration: final report. technical report sri international, menlo park, calif. ( ). duda, r., et al., subjective bayesian methods for rule-based inference systems. proc. afips, washington d.c., pp. - ( ). dworzak, f. et al., design and implementation of a computerised system for evaluation of plant status with respect to safety technical regulations. int. symp. nuclear power plant control and instrumentation, munich, paper iaea-sm- / iaea ( ). ddi handbook for decision analysis. decision & design inc., mclean, va. ( ). earman, l. d., et al., hearsay-ii. comput. surveys , ( ). edgar, g. and m. petty, location of multiple faults by diagnostic expert systems. spie proc. , applications o f artifical intelligence ( ). edwards, j., covers and packings in a family of sets. bull. am. math. soc. , - ( ). eichelberger, e. b.,, hazard detection in combinational and sequential switching circuits. i b m j res. dev. march ( ). elstein, a. s., l. shulman and s. sprafka, medical problem solving: an analysis o f clinical reasoning. harvard university press, cambridge, mass. ( ). erman, l. d., f. hayes-roth, v. r. lesser and d. r. reddy, the hearsay-ii speech-understanding system: integrating knowledge to resolve uncertainty. a c m comput. surveys , - ( ). erman, l. d. and v. r. lesser, a multi-level organization for problem-solving using many diverse cooperating sources of knowledge. proc. th int. joint conf. artificial intelligence. pp. - ( ). fikes, r. e., deductive retrieval mechanisms for state description models. stanford research institute ai technical note ( ). fink, p. k. and j. c. lusth, expert systems for diagnostic expertise in the mechanical and electrical domains. ieee trans. systems man cybernet. smc , - ( ). firdman, e., haidex: tomorrow available today. through the looking glass ( ), - ( ). fitzgerald, m. w. et al., ess t m switching system software: diagnostics. a t & t tech. j. january ( ). flehinger, b. and r. engle, heme: a self-improving computer program for diagnosis-oriented analysis of hematologic diseases. i b m j. res. dev. november, - ). florcik d. and h. wong, a tester independent automated test preparation process for loaded boards. proc. int. test. conf., san francisco, ca ( ). forbus, k. d., interpreting observation of physical system. i e e e trans. systems man cybernet. smc , - ( ). forbus, k., qualitative reasoning about physical processes proc. th hcai, madrid, spain ( ). forbus, k., qualitative process theory. mit technical report ( ). y. lirov fox, m. s. and d. j. moslow, maximal consistent interpretations of errorful data in hierarchically modeled domains. proc. th int. joint conf. artificial intelligence, cambridge, mass. ( ). fox, m. s., s. lowenfeld and p. kleinovsky, techniques for sensor-based diagnosis. proc. th ijcai ( ). freund, t. g., applying knowledge engineering to tps development. proc. ieee autotestcon, pp. - ( ). frey, t. and r. a. kisner, survey of methods for improving operator acceptance of computerized aids. n u r e g cr- ( ). friedman, l., extended plausible inference. proc. th int. joint conf. artificial intelligence ( ). frohwerk, r. a., signature analysis: a new digital field service method. hewlett-packard j. may, - ( ). fusaoka, a., h. seki and k. takahashi, a description and reasoning of plant controllers in temporal logic. proc. th int. joint conf. artificial intelligence, karlsruhe, germany, pp. - ( ). garvey, t. d. et al., an inference technique for integrating knowledge from disparate sources. proc. th int. joint conf. artificial intelligence ( ). genesereth, m. r., diagnosis using hierarchical design models. proc. aaai- , pp. - ( ). genesereth, m. r., the use of design descriptions in automated diagnosis. artificial intell. , , - ( ). genesereth, m. r., the use of hierarchical design models in the automated diagnosis of computer hardware faults. hpp- - , stanford university heuristic programming project ( ). gimmy, k. l. and e. nomme, automatic diagnosis of multiple alarms for reactor control rooms. savannah river lab report no. dp-ms- l- ( ). goel, p., test generation cost analysis and projections. th design automation conf. proc. washington, d.c. ( ). gomez, f., knowledge organization and distribution for diagnosis. ieee trans. systems man cybernet. , - ( ). gomez, f. and b. chandrasekaran, knowledge organization and distribution for medical diagnosis. ieee trans. system man cybernet s m c l l , ( ). gordon, j. and e. shortliffe, a method for managing evidential reasoning in a hierarchical hypothesis space. artificial lntell. , - ( ). gorry, g. a., computer-assisted clinical decision-making. meth. inf. med. , ( ). grillmeyer, o. and a. j. wilkinson, the design and construction of a rule base and inference engine for test system diagnosis. int. test conf. proc., ieee computer society, pp. - ( ). guarro, s. and d. okrent, the logic flowgraph: a new approach to process failure modeling and diagnosis for disturbance analysis applications. u.s.a. nucl. technol. , - ( ). guarro, s. and d. okrent, on the implementation of disturbance analysis systems for detection and diagnosis of safety-related transients. nuclear power plant nucl. engng des. , - ( ). guida, g., inference based on causal knowledge: deep versus shallow reasoning. proc. unicorn seminar expert systems and optimisation in process control, london, december ( ). gunhold, r. and j. zettel, an expert system for automated testing of electrical components and circuits. a i and advanced computer technology conf. proc., p. ( ). hamscher, w. and r. davis, using structural and functional information in diagnostic design. mit a lab memo ( ). hamscher, w. and r. davis, diagnosing circuits with state--an inherently underconstrained problem. proc. conf. american association for artificial intelligence ( ). hardy, s. and r. joyce, a basic expert system tool. us patent ( march ). hartley, r. t., crib: computer fault-finding through knowledge engineering. computer march, - ( ). harvey, a. m. and j. bordley iii, differential diagnosis. saunders, philadelphia, penn. ( ). hasling, d. w., w. j. clancey and g. r. rennels, strategic explanations for a diagnostic consultation system. int. j. man machine stud. , - ( ). hayball, c., fault diagnosis in the wave soldering o f printed circuit assemblies. alvey ikbs research theme: expert systems, deep knowledge, workshop no. , - ( ). hayes-roth, b., the blackboard architecture: a general framework for problem solving? heuristic programming project report no. hpp- - , stanford university ( ). hayes-roth, b., a blackboard model of control. heuristic programming project report no. hpp- - , stanford university ( ). hayes-roth, b. and f. hayes-roth, a cognitive model of planning. cognitive sci. , - ( ). hayes-roth, f., d. waterman and d. lenat (eds) building expert systems. addison-wesley, new york ( ). hayes-roth, f., d. a. waterman and d. b. lenat, principles o f pattern-directed inference systems. in pattern-directed inference systems (ed. d. a. waterman and f. hayes-roth) pp. - . academic press, new york ( ). herbert, m. r . , a review of on-line diagnostic aids for nuclear power plant operators. nucl. energy , - ( ). hudlicka, e. and v. lesser, modeling and diagnosing problem-solving system behavior. ieee trans. systems man and cybernet. smc , - ( ). hudlicka, e. and v. r. lesser, meta-level control through fault detection and diagnosis. proc. conf. american association for artificial intelligence ( ). hughes aircraft corp., knowledge based systems (kbs) study for advanced mate systems. hac report no. fr - - ( ). hyafil, l. and r. l. rivest, constructing optimal binary decision trees in np-complete. lnfo. process. lett. , - ( ). ibarra, o. h. and s. sahni, polynomially complete fault detection problems. i e e e trans. comput. c , - ( ). johnson, p., a. duran, f. hassebroek, j. moiler and m. prietula, expertise and error in diagnostic reasoning. cognitive sci. , - ( ). josephson, j., explanation and induction. ph.d. dissertation, department of philosophy, ohio state univ. ( ). josephson, j. r., b. chandreseharan, j. r. smith and m. c. taumer, a mechanism for forming composite explanatory hypotheses. ieee trans. systems man cybernet. smc , - ( ). karp, r., reducibility among combinatorial problems. in complexity o f computer computations (eds r. miller and j. thatcher), pp. - . plenum press, new york ( ). kelly, v. and l. steinberg, the c r i t t e r system--analyzing digital circuits by propagating behaviors and specifications. proc. conf. american association for artificial intelligence, pp. - ( ). electronic circuit diagnostic expert systems--a survey kemeny, j. c., the report of the president's commission on the accident at three mile island. kluwer, washington d.c. ( ). kemper, c. t., j. c. bellows and p. j. klinosky, diagnostic apparatus. us patent ( february ). kidd, a. l., what do users ask?--some thoughts on diagnostic advice. in expert systems. warwick university ( ). kignchi, t., h. yoshida, h. motoda and s. kobayashi, the development of a plant operation guidance system utilizing knowledge engineering technique. j. atom. energy soc. jpn , - ( ). kim, j., convince: a conversational inference consolidation engine. ph.d. dissertation, university of california, los angeles ( ). kim, j. h. and j. pearl, convince: a conversational inference consolidation engine. eee trans. systems man cybernet. smcit, - ( ). king, j. j., artificial intelligence techniques for device troubleshooting. technical report csl- - , hewlett-pnckard ( ). kitowski, j. and e. ksiazek, fuzzy logic applications for failure analysis and diagnosis of a primary circuit of the h t r nuclear power plant. comput. phys. commun. , - ( ). de kleer, j., local methods for localizing faults in electronic circuits. mit ai lab. mem. ( ). de kleer, j., the origin and resolution of ambiguities in casual arguments. proc. th int. joint conf. artificial intelligence ( ). de kleer, j., an assumption-based truth maintenance system. artificial intell. ( ) - ( ). de kleer, j., a theory of plans for electronic circuits. mit ai lab. working paper ( ) de kleer, j., causal and teleological reasoning in circuit recognition. mit ai lab. technical report ( ). de kleer, j., extending the atms. artificial intell. ( ), - ( ). de kleer, j., local methods for localization of faults in electronic circuits. mit ai lab. memo ( ). de kleer, j. and j. s. brown, mental models of physical mechanisms and their acquisition. in cognitive skills and their acquisition. edbaum, new york ( ) de kleer, j. and j. s. brown, a qualitative physics based on confluences. artificial intell. , - ( ). de kleer, j. and b. williams, reasoning about multiple faults. proc. th nam. conf. artificial intelligence, philadelphia, pp. - ( ). de kleer, j. and g. j. sussman, propagation of constraints applied to circuit synthesis. memo no. , mit ( ). kochan, s., n. landis and d. monson, computer guided probing techniques. proc. int. test conf. san francisco, ca ( ). kolodner, j. l. and r. m. kolodner, using experience in clinical problem solving: introduction and framework. ieee trans. systems man and cybernet. smci , - ( ). konolige, k., a deductive model of belief. proc. th int. joint conf. artificial intelligence, karlsrube, west germany, pp. - ( ). konolige, k. and n. j. nilsson, multi-agent planning systems. proc. aaai-i, stanford, calif. pp. - ( ). kuipers, b., qualitative simulation as causal explanation. ieee trans. systems man cybernet. smc , - ( ). kuipers, b., commonsense reasoning about causality: deriving behavior from structure. artificial intell. , - ( ). kulikowski, c. a., pattern recognition approach to medical diagnosis. ieee trans. systems sci. cybernet. ssc , - ( ). kulikowski, c. a., artificial intelligence methods and systems for medical consultation. ieee trans. pattern analysis med. lntell. , - ( ). lai, k.-w., functional testing of digital systems. ph.d. thesis, carnegie-mellon university department of computer science, technical report cmu-cs- - ( ). lee, f. p., process computer alarm and disturbance analysis: review of the state-of-the-art. comput. chem. engng , - ( ). leitch, r., towards intelligence control systems. proc. unicom seminar on expert systems and optimisation in process control, london, december ( ). lenat, d. b. beings: knowledge as interacting experts. proc. th int. joint conf. artificial intelligence, tbilisi, u.s.s.r. pp. - . ( ). lin, j. c. et al., a deterministic methodology for application to a diagnosis oriented disturbance analysis system. nucl. engng des. , - ( ). lind, m., the use of flow models for plant diagnosis. in human detection and diagnosis o f system failures (ed. j. rasmussen and w. b. rouse), pp. - . plenum press, new york ( ). lindsay, r. k., b. g. buchanan, e. a. feigenbaum and j. lederberg applications o f artificial intelligence for organic chemistry: the dendral project. mcgraw-hill, new york ( ). lipkin, m. and j. d. hardy, mechanical correlation of data in differential diagnosis of hematological diseases. j. am. med. assoc. , - ( ). long, w., a program for the management of heart failure. sigart newsletter no. ). loveland, d. w., selecting optimal test procedures from incomplete test set. proc. st int. syrup. policy analysis and information science, duke university, durham, n.c., pp. - ( ). macaloney, b. g., fault spectrum: an analysis of system level test with proposed solutions. proc. int. test conf. ( ). maclean, c. and p. wilde, knowledge-based electronic circuit diagnosis. proc. expert systems , technical conf. bcs sges, cambridge, u.k. ( ). magliero, a., r. leong and r. bethel, ads-the idss adaptive diagnostic system. proc. autotestcon ' , pp. - ( ). magulre, j., ai-ferret troubleshooting assistant. advanced products laboratory, support systems, hughes aircraft co., long beach, calif. ( ). martin-soils, g. a. et al., fault tree synthesis for design and real time applications. trans. icheme , - ( ). masui, s., j. mcdermott and a. sobel, decision making in time-critical situations. proc. th int. joint conf. artificial intelligence, karlsruhe, west germany ( ). matsumoto, k., t. sakagnchi and t. wake, fault diagnosis of a power system based on a description of the structure and function of the relay system. expert systems july, - ( ). y. lirov mccown, p. and d. lewy, apu m a i d - - a diagnostic expert system using heuristic and causal reasoning. proc. a u t o t e s t c o n ' , pp. - ( ). mcdermott, d., circuit design as problem solving. in artificial intelligence and pattern recognition in computer aided design (ed. latombe). north-holland, amsterdam ( ). mcdermott, d., non mototonic logic ii: non-monotonic modal theories. research report no. department of computer science, yale univ. ( ). mcdermott, d. v., planning and acting. cognitive sci., , - ( ). mcdermott, j., ri: a rule-based configurer of computer systems. technical report cmu-cs- - , carnegie-mellon university, pittsburgh ( ). merry, m., apex : an expert system shell for fault diagnosis. gec j. res. , ( ). miller, r., h. pople and j. myers, intrnist- , an experimental computer-based diagnostic consultant for general internal medicine. n. engl. j. med. , - ( ). milne, r., strategies for diagnosis. ieee trans. systems man cybernet. smci , ( ). milne, r., artificial intelligence applied to condition health monitoring. chart. mech. engng may ( ). milne, r., diagnosing faults through responsibility. proc. st int. conf. applied artificial intelligence, southampton, england ( ). milne, r., diagnosing faults through responsibility. a c m annual conf. association o f computing machinery, denver, col., pp. - ( ). milne, r., fault diagnosis and expert systems. th int. workshop expert systems and their applications, avignon, france ( ). milne, r., fault diagnosis through responsibility. proc. th int. joint conf. artificial intelligence, los angelos, pp. - ( ). milne, r., reasoning about structure, behavior and function. sigart newsletter no. , pp. - ( ). milne, r., using ai in the testing of printed circuit boards. national aerospace and electronics conf. dayton, ohio ( ). milne, r. w., the theory of responsibilities. sigart newsletter no. ( ). mingers, j., expert systems--experiments with rule induction. j. op. res. soc. , - ( ). minsky, m., a framework for representing knowledge. in the psychology o f computer vision (ed. p. winston), pp. - . mcgraw-hill, new york ( ). mitchell, t., l. steinberg, r. g. smith, p. schooley, v. kelley and h. jacobs, representations for reasoning about digital circuits. proc. th int. joint conf. artificial intelligence, pp. - ( ). mittal, s., design of a distributed medical diagnosis and database system. ph.d. dissertation, department of computer and information science, ohio state univ. ( ). mittal, s., b. chandrasekaran and j. smith., overview of m d s - - a system for medical diagnosis. proc. rd symp. computer applications in medical care, ieee, pp. - ( ). moore, r., expert systems in process control: applications experience. ist int. conf. applied artificial intelligence., southampton, england ( ). morris, r. and r. nado, representing actions with an assumption-based truth maintenance system. proc. aaai- , philadelphia, penn. ( ). mostow, d. j., a problem solver for making advice operational. proc. aaai- , pp. - ( ). mullis, r., expert system for vlsi tester diagnostics. proc. ieee test conf. ( ). nadig, h. j., signature analysis---concepts, examples, and guidelines. hewlett-packard j. may, - ( ). nau, d. and j. reggia, relationship between abductive and deductive inference. proc. st int. workshop expert database systems, pp. - ( ). nawal, h., v. lesser and e. milies, diagnosis using a formal theory of a signal processing system. ieee trans. systems man cybernet. smci , - ( ). nelson, w. r., reactor: an expert system for diagnosis and treatment of nuclear reactor accidents. proc. nd conf. american association for artificial intelligence, pp. - ( ). nii, h. p. and n. aiello, age (attempt to generalize): a knowledge-based program for building knowledge-based programs. proc. th int. joint conf. artificial intelligence, tokyo, japan ( ). nilsson, n. j., principles o f artificial intelligence. tioga, palo alto, calif. ( ). patil, r., p. szolvits and w. schwartz, causal understanding of patient illness in medical diagnosis. proc. th int. joint conf. artificial intelligence, vancouver, canada ( ). patrick, e. a. et al., review of pattern recognition in medical diagnosis and consulting relative to a new system model. ieee trans. systems man. cybernet. smc , - ( ). pau, l. f., an adaptive signal classification procedure: application to aircraft engine monitoring. pattern recognition , - ( ). pau, l. f., applications of pattern recognition to failure analysis and diagnosis. in human detection and diagnosis o f system failures (ed. j. rasmussen and w. b. rouse) pp. - . plenum press, new york ( ). pau, l. f., failure diagnosis and performance modelling. dekker, new york ( ). pau, l. f., failure diagnosis systems. north-holland, amsterdam ( ). pau, l. f., failure diagnosis by an expert system and pattern classification. pattern recognition lett. , - ( ). pau, l. f., integrated testing and algorithms for visual inspection of integrated circuits. ieee trans. pattern analysis medicine intell. pami , - ( ). pazzani, m. j., failure driven learning of fault diagnosis heuristics. ieee trans. systems man cybernet. smc , - ( ). pearl, j., fusion, propagation, and structuring in belief networks. artificial intell. , - ( ). pearl, j., heuristics: intelligent search strategies for computer problem solving. addison wesley, new york ( ). peng, y., a formalization of parsimonious covering and probabilistic reason in abductive diagnostic inference. ph.d. dissertation, department of computer science univ. maryland, baltimore ( ). peng, y. and j. reggia, a probabilistic causal model for diagnostic problem-solving--part : integrating symbolic causal inference with numeric probabilistic inference. ieee trans. systems man cybernetics. smc , - ( ). peng, y. and j. reggia, a probabilistic causal model for diagnostic problem-solving, part two: diagnostic strategy. ieee trans. systems man cybernet. smci , - . ( ). electronic circuit diagnostic expert systems---a survey perkins, w. a. and t. j. laffey, les: a general expert system and its applications. spie proc. , applications o f artificial intelligence, pp. l - ( ). pipitone, f., the fis electronic troubleshooting system. computer. july, - ( ). poage, j. f. and e. j. mccluskey, derivation of optimum tests for sequential machines. proc. th a. symp. switching circuit theory logic designs, pp. - ( ). pople, h., the formation of composite hypotheses in diagnostic problem solving: an exercise in synthetic reasoning. proc. th int. joint conf. artificial intelligence, cambridge, mass. ( ). pople, h., on the mechanization of abductive logic. proc. rd int. joint conf. artificial intelligence, pp. - ( ). putzolu, g. r. and j. p. roth, a heuristic algorithm for testing of asynchronous circuits. ieee trans. comput. c , - ( ). pynn, c., strategies for electronics test. mcgraw-hill, new york ( ). quinlan, j. r., learning efficient classification procedures and their application to chess end game. in machine learning: an artificial intelligence approach (ed. r. michalski, j. carbonnell and t. mitchell). springer, new york ( ). ramsey, j., diagnosis using automatic test equipment and an artificial intelligence expert system. masters thesis, air force institute of technology ( ). rasmussen, j. and a. jense, mental procedures in real-life tasks: a case study of electronic trouble shooting. ergonomics , - ( ). reggia, j., abductive inference. proc. expert systems in government symp., pp. - ( ). reggia, j., computer-assisted medical decision making. in applications o f computers in medicine (ed. m. schwartz), pp. - , ieee press ( ). reggia, j. and d. nau, an abductive non-monotonic logic. proc. workshop non-monotonic reasoning, aaai, pp. - ( ). reggia, j. and s. tuhrim, an overview of methods for computer-assisted medical decision making. in computer-assisted medical decision making, vol. i (ed. j. reggia and s. tuhrim), pp. - . springer, new york ( ). reggia, j., d. nau, p. wang and y. peng, a formal model of diagnostic inference. infor. sci. , - ( ). reggia, j., d. nau and p. wang, diagnostic expert systems based on a set covering model. int. j. man-machine stud. november, - ( ). reggia, j., b. perricone, d. nau and y. peng, answer justification in abductive expert systems for diagnostic problem solving. ieee trans. biomed. engng be , - ( ). reiter, r., a theory of diagnosis from first principles. department of computer science, univ. toronto, toronto, report on tr- / ( ). richer, m. h. and w. j. clancey, guidon watch: a graphic interface for browsing and viewing a knowledge-based system. ieee comput. graphics applic. ( ), - ( ). rieger, c. and m. grinbert, the declarative representation and procedural simulation of causality in physical mechanisms. proc. th int. joint conf. artificial intelligence, pp. - ( ). robinson, j. a., a machine-oriented logic based on the resolution principle. j. assoc. comput. machinery , - ( ). rosenschein, s. j., plan synthesis: a logical perspective. proc. th int. joint conf. artificial intelligence, vancouver, canada ( ). roth, j. p., computer logic, testing, and verification. computer science press, new york ( ). roth, j. p., diagnosis of automata failures: a calculus and a method. i b m j. res. dev. october, - ( ). roth, j. p., w. g. bouricius and p. r. schneider, programmed algorithm to compute tests to detect and distinguish faults in logic circuits. ieee trans. electronic comput. eci ( ). roylance, g. l., qualitative analysis of operational amplifier circuits. mit dept. of ef.&cs, ms thesis proposal, march ( ). rubin, a., the role of hypotheses in medical diagnosis. in proc. th int. joint conf. artificial intelligence, pp. -- ( ). ryan, p. and a. j. wilkinson, knowledge acquisition for ate diagnosis. ieee proc. int. test conf. pp. - ( ). sacerdoti, e. d., a structure for plans and behavior. elsevier, north-holland, amsterdam ( ). samuel, a., some studies in machine learning using the game of checkers ii. recent progress. i b m j. res. dev. , - ( ). scarl, e. a., j. r. jamieson and c. i. delaune, diagnosis and validation through knowledge of structure and function. ieee trans. systems man cybernet. smci , - ( ). scarl, e. a., j. r. jamieson and c. i. delaune, process monitoring and fault locations at the kennedy space center. newsletter no. ( ). schneider, p. r., on the necessity to examine d-chains in diagnostic test generation--an example. i b m j. res. dev. november, ( ). sembugamoorthy, v. and b. chandrasekaran, functional representation of devices and compilation of diagnostic problem solving systems. cognitive sci. , - ( ). seshu, s., on an improved diagnosis program. ieee trans. electronic comput. eci , - ( ). seshu, s. and d. n. freeman, the diagnosis of asynchronous sequential switching systems, lee trans. electronic comput. ecll, - ( ). sharer, g., a mathematical theory o f evidence. princeton univ. press ( ). shirley, m. h., digital test generation from hierarchical models and failure symptoms. masters thesis, mit ai lab ( ). shirley, m. and r. davis, digital test generation from symptom information. ieee vlsi workshop ( ). shortliffe, e., computer-based medical consultations: mycin. elsevier, new york ( ). shortliffe, e., m y c i n - - a rule-based computer program for advising physicians regarding antimicrobial therapy selection. stanford univ. ai lab., memo aim- ( ). shortliffe, e. and b. buchanan, a model of inexact reasoning in medicine. math. biosci. , - ( ). simmons, f., c. chee, s. laughton, l. wood, m. purvis and d. throop, self explanatory device descriptions. sigart newsletter no. ( ). singhal, k. and j. vlach, symbolic analysis of analog and digital circuits. ieee trans. circuits systems. cas , - ( ). slagle, j. r. and r. c. t. lee, application of game tree searching techniques to sequential pattern recognition. commun. acm, , - ( ). y. lirov stallman, r. m. and g. j. sussman, forward reasoning and dependency-directed backtracking in a system for computer-aided circuit analysis. artificial intell. , - ( ). stallman, r. m. and g. j. sussman, problem solving about electrical circuits. in artificial intelligence: an m i t perspective (ed. p. h. winston and r. h. brown), pp. - . mit press ( ). stanfill, c., mack, a program which deduces the behaviour of machines from their forms. sigart newsletter no. ( ). stefik, m. j., planning with constraints. artificial intell. , - ( ). stefik, m. j., planning and meta-planning. artificial lntell. , - ( ). steinberg, l. and v. kelly, the c r i t t e r system: analyzing digital circuits by propagating behaviors and specifications. proc. nd national conf. on artif. intelligence, washington d.c., pp. - ( ). sussman, g. j., a computational model of skill acquisition. mit ai lab technical report ( ). sussman, g. j., electrical design--a problem for artificial intelligence research. mit ai memo ( ). sussman, g. j., slices at the boundary between analysis and synthesis. in artificial intelligence and pattern recognition in computer aided design (ed. b. latombe). north-holland, amsterdam ( ). sussman, g. j. and r. m. stallman, heuristic technique in computer-aided circuit analysis. ieee trans. circuit systems cas ( ), ( ). swabey, m., use of deep and shallow knowledge in electronics diagnosis. proc. unicom seminar on expert systems and optimization in process control, london ( ). swanson, d. b. computer simulation of expert problem solving in medical diagnosis. ph.d. dissertation, univ. minnesota ( ). swartout, w. r., explaining and justifying expert consulting programs. proc. th int. joint conf. artificial intelligence, pp. ~ ( ). szolovits, p. and s. pauker, categorical and probabilistic reasoning in medical diagnosis. artificial lntell. , - ( ). szygenda, s. a. and e. w. thompson, modeling and digital simulation for design verification diagnosis. ieee trans. comput. c , - ( ). taie, m. r. and s. srihari, modeling connections for circuit diagnosis. proc. rd conf. a applications, pp. - ( ). taie, m. r., s. n. srihari, j. geller and s. c. shapiro, device representation using instantiation rules and structureal templates. proc. canadian a i conf- , presses de l'universit du qu bec, montr al, pp. - ( ). taie m. r., j. geller, s. n. srihari and s. c. shapiro, knowledge based modelling of circuit boards. reliability and maintenance syrup. proc., pp. - ( ). tanaka, t., representation and analysis of electrical circuits in a deductive system. proc. th int. joint. conf. artificial intelligence, west germany, pp. - ( ). tanaka, t., simulation of thinking process on electronic circuits in natural language i, i , iii. technology report of kyushu univ., vol. , p. ( ). tanaka, t., structural analysis of electrical circuits in a deductive system. in computer expert systems (ed. l. bolc). springer, new york ( ). tendolkar, n. n. and r. l. swann, automated diagnostic methodology for the ibm processor complex. i b m j. res. and dev. , - (t ). thompson, t. f. and r. m. wojcik, methods and apparatus for system fault diagnosis and control. u.s. patent ( march ). towill, d. r., automatic testing of control systems---past present and future. j. inst. electronic radio engrs ( ), - ( ). tsiang, s. h. and w. ulrich, automatic trouble diagnosis o f complex logic circuits. bell syst. tech. j. ( ), - ( ). tulloss, r. e., fault dictionary compression: recognizing when a fault may be unambiguously represented by a single failure detection. proc. test conf. ( ). tulloss, r. e. et al., the diagnostic organization and retrieval-algorithm system. the engineer. summer ( ). tversky, a. and d. kahneman, causal schemata in judgements under uncertainty. in progress in social psychology (ed. m. fishbein). erlbaum, hillsdale, n.j. ( ). ulrich, e. g., t. baker and l. r. williams, fault test analysis techniques based on simulation. proc. th design automation workshop, pp. - ( ). van melle, w., a domain independent production-rule system for consultation programs. proc. th int. joint conf. artificial intelligence, pp. - ( ). vere, s., planning in time: windows and durations for activities and goals. ieee trans. pattern analysis machine intell. pami , - ( ). wang, d. t., an algorithm for the detection of test sets for combinational logic networks. ieee trans. comput. c , - ( ). warren, d. h. d., generating conditional plans and programs. proc. a i s b summer conf., university of edinburgh, u.k., pp. - ( ). washio, t., m. kitamura, k. kotajima and k. sugiyama, semantic network approach to automated failure diagnosis in nuclear power plant. proc. ans int. topical meet. computer applications for nuclear power plant operation and control, pasco, wash. ( ). weiss, s., c. a. kulikowski, s. amarel and a. safir, a model-based method for computer-aided medical decision-making. artificial lntell. , - ( ). weiss, s. m. and c. a. kulikowski, expert: a system for developing consultation models. proc. th lnt. joint conf. artificial intelligence, tokyo, japan, pp. - ( ). white, b. y. and j. r. fredericksen, quest: qualitative understanding o f electrical systems troubleshooting. sigart newsletter no. ( ). wilkinson, a. j., benchmarking an expert system for electronic diagnosis. proc. a u t o t e s t c o n ' , pp. - ( ). wilkinson, a. j., a method for test system diagnostics based on the principles of artificial intelligence. int. test conf., , proc. ieee, pp. - ( ). wilkinson, a. j., mind: an inside look at an expert system for electronic diagnosis. design test mag. august, - ( ). williams, t. l., isolating multiple faults in nds. technical report , smart systems technology, falls church, va. ( ). williams, t. w. and k. p. parker, design for testability--a survey. i e e e trans. comput. c , - ( ). winston, p., artificial intelligence. addison-wesley, new york ( ). yau, c. w., private communication ( ). yau, s. s. and s. c. yang, multiple fault detection for combinational logic circuits. ieee trans. comput. c , - ( ). intelligent tourism recommender systems: a survey expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a review intelligent tourism recommender systems: a survey http://dx.doi.org/ . /j.eswa. . . - /� elsevier ltd. all rights reserved. ⇑ corresponding author at: science & technology park for tourism and leisure, c/ joanot martorell, , vila-seca, catalonia, spain. tel.: + . e-mail addresses: joan.borras@pct-turisme.cat (j. borràs), antonio.moreno@urv. cat (a. moreno), aida.valls@urv.cat (a. valls). joan borràs a,b,⇑, antonio moreno b, aida valls b a science & technology park for tourism and leisure, c/joanot martorell, , vila-seca, catalonia, spain b intelligent technologies for advanced knowledge acquisition (itaka), departament d’enginyeria informàtica i matemàtiques, universitat rovira i virgili, av. països catalans, , tarragona, catalonia, spain a r t i c l e i n f o a b s t r a c t article history: available online june keywords: recommender systems tourism ontologies planning clustering recommender systems are currently being applied in many different domains. this paper focuses on their application in tourism. a comprehensive and thorough search of the smart e-tourism recommend- ers reported in the artificial intelligence journals and conferences since has been made. the paper provides a detailed and up-to-date survey of the field, considering the different kinds of interfaces, the diversity of recommendation algorithms, the functionalities offered by these systems and their use of artificial intelligence techniques. the survey also provides some guidelines for the construction of tour- ism recommenders and outlines the most promising areas of work in the field for the next years. � elsevier ltd. all rights reserved. . introduction the amount of information available in the world wide web and its number of users have experienced an enormous increase in the last decade. all this information may be particularly useful for those users who plan to visit an unknown destination. informa- tion about travel destinations and their associated resources, such as accommodations, restaurants, museums or events, among oth- ers, is commonly searched for tourists in order to plan a trip. how- ever, the list of possibilities offered by web search engines (or even specialised tourism sites) may be overwhelming. the evaluation of this long list of options is very complex and time consuming for tourists in order to select the one that fits better with their needs. personalization techniques (gao, liu, & wu, ) aim to provide customised information to users based on their preferences, restrictions or tastes. they are particularly relevant in recom- mender systems (adomavicius & tuzhilin, ), whose objective is to filter irrelevant options and to provide personalised and rele- vant information to each particular user. in the tourism field, travel recommender systems (ricci, ) aim to match the characteris- tics of tourism and leisure resources or attractions with the user needs. these systems are especially useful if they can automati- cally learn the user’s preferences through the analysis of her expli- cit or implicit feedback (sieg, mobasher, & burke, ). explicit data may be given by the user in different ways, for instance whenever she specifies her cultural interests by filling in a form. implicit interests can be inferred by the system through the analysis of the behaviour of the user. in many cases recommender systems not only take into account the preferences of the tourist but they also analyze the dynamic context (dey & abowd, ) of the trip. this is especially useful when tourists are already at the destination and they are willing to use their mobile devices to customise their trips in real time. the context can include aspects like the tourist’s location, the time of the visit or the current weather (lamsfus, alzua-sorzabal, martin, salvador, & usandizaga, ). approaches that take con- text into account can send suggestions proactively, depending on the current state of the tourist. for example, a museum that was planned to be visited today may be too far from the visitor and she may not have enough time to reach it, so the plan scheduled for tomorrow may be changed to include the visit to the museum. the last ten years have witnessed an explosive increase in the use of mobile technology among tourists. therefore, e-tourism systems (buhalis, ) provide a good opportunity for mobile ser- vices that help visitors by offering recommendations based on their preferences and their current context. this paper focuses on tourism recommender systems that employ artificial intelligence (ai) techniques at some point. some examples can be the following: � intelligent autonomous agents may analyse the behaviour of a user, learn automatically the user profile and provide proactive recommendations depending on the current context (batet, moreno, sánchez, isern, & valls, ). http://crossmark.crossref.org/dialog/?doi= . /j.eswa. . . &domain=pdf http://dx.doi.org/ . /j.eswa. . . mailto:joan.borras@pct-turisme.cat mailto:antonio.moreno@urv.cat mailto:antonio.moreno@urv.cat mailto:aida.valls@urv.cat http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa fig. . interfaces used in the reviewed works (in%). j. borràs et al. / expert systems with applications ( ) – � some systems go beyond offering a list of recommended tourist attractions and use automated planners to schedule these recommendations within a route that can span several days (vansteenwegen, souffriau, vanden berghe, & van oudheusden, ). � other approaches take into account the opening and closing times of the attractions, or the time needed to go from one point of interest to another, hence offering a detailed timetable of the visit. however, this is a very complex planning and scheduling problem for which it is difficult to guarantee an optimal solu- tion. some systems solve this complexity with the use of ai opti- mization techniques, such as ant colony or meta-heuristic iterative methods (lee, chang, & wang, ). � automatic clustering algorithms may be applied to classify tour- ists with similar tastes or similar features (gavalas & kenteris, ). � approximate reasoning methodologies, like fuzzy logic or bayesian networks, may be used to manage the imprecision related to the inference of the preferences of the user (huang & bian, ). � reasoning procedures, like rule-based systems, are also employed to deduce the user’s preferences (lamsfus, alzua-sorzabal, martin, & smithers, ). � formalisms developed in the knowledge representation field of ai, like ontologies, are commonly used to represent (and reason about) the tourism domain knowledge (moreno, valls, isern, marin, & borràs, ). the contributions of this article are twofold. on the one hand, it provides a comprehensive review of the tourism recommender systems published in scientific journals and conferences since , with an especial focus on the ones that employ ai tech- niques. commercial products are not considered in this review, as it is usually not possible to know how they have been designed and implemented. several aspects of these systems are analysed, such as their interface, their functionalities, the recommendation mechanisms and the ai methods and techniques employed. this review provides an up-to-date state of the art of the field of intel- ligent tourism recommenders, which may be useful not only to the scientists working in this field but to designers and developers of intelligent recommender systems in other domains. on the other hand, the paper also provides guidelines to be followed in the design of this kind of intelligent systems and an outline of some of the most promising research lines that may be pursued in the near future. the reminder of this review is structured as follows. in the next section we analyze which interfaces are commonly used by tour- ism recommender systems to interact with users, discussing espe- cially the differences between mobile and web-based approaches. after that we survey the main functionalities offered by these sys- tems, ranging from the recommendation of a tourist destination to table review of user interfaces. interface references web + mobile venkataiaha et al. ( ),(lamsfus et al. ( ), niaraki and kim ( ( ), borràs et al. ( a), ruotsalo et al. ( ) and umanets et a only web coelho, martins, and almeida ( ), huang and bian ( ), lucas, l montes ( ), jannach et al. ( ), mínguez et al. ( ), sebastià colomo-palacios, gonzález-carrasco, and ruiz-mezcua ( ), linaza et al. ( ), montejo-ráez et al. ( ), sebastià et al. ( ), garci ( ), kurata and hara, ( ), lucas et al. ( ) and savir et al. ( only mobile castillo et al. ( ), ceccaroni et al. ( ), garcía-crespo et al. ( martínez-santiago et al. ( ), noguera et al. ( ), braunhofer et yang and hwang ( ) the automatic construction of a detailed complex schedule of a visit of several days to a certain area. section comments the rec- ommendation methods employed by e-tourism recommenders, focusing on content-based and collaborative approaches. the next section exposes the use of ai techniques from different fields like multi-agent systems, approximate reasoning, knowledge represen- tation, etc. a comparison with previous surveys on tourism recom- menders is given in section . the paper concludes with a global analysis of the surveyed systems and some suggestions of lines of future work in the field. . interface this section analyses the user interfaces of recent tourism rec- ommender systems. most of them offer a web-based interface and/or an interface specifically designed to be used in mobile devices. table classifies the most relevant e-tourism recom- menders in these two broad categories, and fig. shows the per- centage of surveyed systems in each of them. a web-based interface is the option chosen by most of the systems, since it per- mits an easy access from any computer connected to the web without any kind of downloading, installation and configuration. however, due to the enormous increase in the use of smart phones connected to the web in the last years, more than half of the reviewed systems have specific interfaces for mobile devices. there are some recommender systems that have been designed as desktop applications and do not offer any of the two usual kinds of interfaces (e.g., kurata, ). this kind of applications can usu- ally be implemented more quickly than the mobile or web-based ones; however, they require downloading and installing the pro- gram, which is not comfortable to most of the tourists that want to get recommendations as simply as possible without being both- ered by technical details. ), vansteenwegen et al. ( ), gavalas and kenteris ( ), rey-lópez et al. l. ( ) aurent, moreno, and teisseire ( ), lee et al. ( ), ruiz-montiel and aldana- et al. ( ), yang ( ), borràs et al. ( ), garcía-crespo, lópez-cuadrado, , aguirregoikoa, garcía, torres, and aranburu ( ), lorenzi et al. ( ), luberg a et al. ( ), wang et al. ( ), koceski and petrevska ( ), gyorodi et al. ) ), yu and chang ( ), ricci et al. ( ), martin et al. ( ), batet et al. ( ), al. ( ), garcia et al. ( ), meehan et al. ( ), rojas and uribe ( ) and fig. . personalized route through tainan city (from lee et al., ). j. borràs et al. / expert systems with applications ( ) – the following subsections review some approaches based on web or mobile interfaces. . . web-based recommenders the use of a web-based interface is the most common option adopted by e-tourism recommenders. this kind of interfaces allows tourists to look for information in a user-friendly manner. users normally have a rich interaction with the system using a wide screen which allows displaying a large amount of data extended with maps, images or even high quality videos. more- over, the mouse permits to interact easily with the computer and move through maps, perform zoom actions, select items or even drag and drop them. this is very useful for tourists when they are still in the planning stage of their trips. nevertheless, web- based applications are usually not designed to be used during the stay since most of the tourists will not have easy access to comput- ers with internet connection. although an increasing number of tourists have mobile handsets or tablets with internet connection, the information-ridden web pages usually shown by recommend- ers cannot be easily read or manipulated on such small screens. in the remainder of this section we comment some interesting fea- tures exploited in web-based interfaces to improve the interaction with the users. venkataiaha, shardaa, and ponnadaa ( ) report the design of two visualisation systems (called discrete and continuous) for a tourism recommender and compare the interaction of the users in both cases. the former provides a high quantity of information in the screen at the same time, and it was determined that users needed too much time and effort to understand it. the latter aggre- gates all the information into a single video clip that combines the most relevant media content, including text, photographs and videos. the approach shown in lee et al. ( ) is one of the firsts that embeds google maps services in their web pages (fig. ) in order to plot the travel route on a map, so that tourists can follow the per- https://developers.google.com/maps/ (last access march ). sonalised itinerary to enjoy cultural heritage and local gourmet food during their stay at tainan city. other web-based recommender systems that display in a map the places scheduled to be visited in a single day are e-tourism (sebastià, garcia, onaindia, & guzman, ), city trip planner (vansteenwegen et al., ), otium (montejo-ráez, perea- ortega, garcía-cumbreras, & martínez-santiago, ) and enosig- tur (borràs et al., a). in this last system the user introduces her socio-demographic information and general preferences (left image of fig. ), and then she will receive recommendations of attractions spread out over the province of tarragona, a wide region of spain (right image of fig. ). the last web page shows a high quantity of information, such as geo-localised attractions classified by categories, a route indicating driving times between the suggested activities, their approximate visiting times, etc. the vibe virtual spa advisor (jannach, zanker, & jessenitschnig, ) keeps an avatar-based conversation with the tourist in order to acquire the user’s visit requirements through personalised forms. the main point of this approach is its dynamicity. if a new attribute has to be added to the product catalogue, it is automati- cally taken into account not only in the recommendation and pref- erence elicitation processes, but also in the web interface which is changed accordingly. the web site has a section for domain experts, in which they can add or modify logical conditions that govern the conversational and recommendation procedures. wang, zeng, and tang ( ) show how semantic web technol- ogies may be integrated with web . services to leverage each other’s strengths. to do so they proposed an ontology-based tour- ism recommender that allows the automatic and dynamic integra- tion of heterogeneous on-line travel information. the platform is built in ruby on rails with view extensions to create rich ajax web-based applications. they also use third party services to pro- vide additional features, such as google map, yahoo weather, and wikitravel. . . mobile recommendations systems that offer mobile interfaces have increased consider- ably in the last few years, due to the large number of users acquir- https://developers.google.com/maps/ fig. . web version of enosigtur: introduction of initial user profile and route planner (from borràs et al., a). j. borràs et al. / expert systems with applications ( ) – ing mobile devices with internet connection or, more recently, the well-known smartphones. mobile devices are small and their internet connection is usually slow; thus, the quantity of informa- tion that can be shown in these devices cannot be compared with a standard web page. therefore, mobile tourism recommender sys- tems have to make an effort to provide only the information that is essential for the user, and it must be well structured in order to be displayed correctly in small screens. moreover, the user’s interaction with the system is limited, since even the basic actions made in web-based interfaces (scrolling, introducing text) are not that easy. however, it is fair to say that the latest smartphones with bigger touchscreens provide a better user interaction. fur- thermore, the main advantage of mobile devices is that they allow the use of the system in any place with an internet connection, so that tourists may access information, discover places or modify their trips during the stay. besides, most mobile systems are equipped with gps and the recommender may know the present location of the user and it may offer geo-referenced information, advice or recommendations based on this knowledge. one of the first approaches in the field that used mobile systems was reported in yu and chang ( ). this system, designed for pdas, offers location-based recommendation services to support personalised tour planning. recommendations are based on fig. . prototype system for windows mo tourists’ preferences, location and time. fig. shows the mobile user interface in four separated screenshots. the first one shows the different mobile tourism services (restaurant, hotel, sightsee- ing spot, user profile, and tour plan recommendation). the second image illustrates the interface for setting user preferences. the third screenshot shows the recommended tour plan with informa- tion about the places to visit, such as names, descriptions, photos or visiting time frames. finally, the last image displays the tour plan on google maps. another approach compatible with pdas is mtrs (gavalas & kenteris, ). the authors argue that tourists may have prob- lems to connect with the internet, either because they are in a rural area or because they are foreigners and cannot afford the roaming costs abroad. they propose to solve this problem by installing an infrastructure to support proximity detection and a cost-effective means for remote content update. in fact, they propose to use small- to medium-scale wireless sensor networks. through this infrastructure, they introduce the concept of ‘con- text-aware rating’, in which user ratings uploaded through fixed internet connection infrastructures (located at the rated places) are weighted higher to differentiate them from users that pro- vide an evaluation using the internet away from the visited place. bile devices (from yu & chang, ). j. borràs et al. / expert systems with applications ( ) – another product using mobile devices is mapmobyrek (ricci, nguyen, & averjanova, ) that exploits quite well its interface by showing recommendations in lists and on maps. this system permits to compare two items with their characteristics displayed side-by-side in order to decide the one that is preferred. geoasis (martínez-santiago, ariza-lópez, montejo-ráez, & ureña-lópez, ) acts as a tourist guide that describes the places to visit while the tourist approaches the recommended locations. the system uses the mobile gps device to know the tourist location and speed in order to estimate the available time to give the expla- nations. users can interact with the system in two ways: using a tactile interface or using a voice-based interface (voice recognition and text-to-speech software). despite the existence of several mobile tourism recommenders, not many of them use the newest technologies in mobile devices, such as the android or iphone platforms. some examples that use these popular and rising platforms are reviewed below. fig. . preference elicitation from the mobile interf fig. . mobile version of enosigtu the moretourism (rey-lópez, barragáns-martínez, peleteiro, mikic-fonte, & burguillo, ) android-based platform provides information about tourist resources through the use of mashups, integrating images, videos, augmented reality services, geo-loca- tion, guide services, access to urban networks, etc. another approach that uses android platforms is enosigtur (borràs et al., a). fig. shows some screenshots of the mobile version of enosigtur: a list of recommended places, the route of the trip and detailed information of an attraction. livecities (martin, alzua, & lamsfus, ) uses the notification service of android systems to provide push information according to the user context. this information can be plain text, audio, video or html. the sts system (braunhofer, elahi, ricci, & schievenin ) is a powerful android application with a good design inter- face that permits the user to enter accurate information about her interests and opinions on the trip and the visited attractions (see fig. ). ace of the sts system (braunhofer et al. ). r (from borràs et al., a). fig. . ipad version of the guideme system (umanets et al. ). fig. . functionalities applied in the reviewed approaches (in%). j. borràs et al. / expert systems with applications ( ) – the recent guideme system (umanets, ferreira, & leite ) features a good implementation for mobile devices since its designers have not only developed an app for phones but also for tablet devices. in particular, the app is built for the ios platform and it is adaptive to the screen sizes with specific adjustments for both iphone and ipad devices. fig. shows screenshots of its ipad version. reja (noguera, barranco, segura, & martínez, ) also works for ios platforms. . functionalities in this section we describe the general functionalities provided by the reviewed tourism recommender systems. table catalogues the approaches in four broad groups, depending on the services they offer: suggestion of a destination and construction of a whole tourist pack, recommendation of suitable attractions in one specific destination, design of a detailed multi-day trip schedule, and social capabilities. fig. gives a visual estimation of the percentage of systems that offer each of them. these aspects are commented in more detail in the following subsections, with examples of the most prominent proposals. table review of user functionalities. functionalities references destination/tourist packs seidel et al. ( ), yu and chang ( ), lorenzi et al. ( suggest attractions castillo et al. ( ), coelho et al. ( ), ceccaroni et al. ( et al. ( ), lee et al. ( ), niaraki and kim ( ), ruiz-m mínguez et al. ( ), ricci et al. ( ), sebastià et al. ( ), gavalas and kenteris ( ), kurata ( ), linaza et al. ( et al. ( ), rey-lópez et al. ( ), sebastià et al. ( ), g koceski and petrevska ( ), martínez-santiago et al. ( ) , lucas et al. ( ), meehan et al. ( ), rojas and uribe and hwang ( ) trip planner castillo et al. ( ), coelho et al. ( ), ceccaroni et al. ( ( ), niaraki and kim ( ), yu and chang ( ), míngue linaza et al. ( ), luberg et al. ( ), montejo-ráez et al. et al. ( ), batet et al. ( ), borràs et al. ( a), kurata social aspects coelho et al. ( ), ceccaroni et al. ( ), garcía-crespo et ( ), garcia et al. ( ), garcia et al. ( ), meehan et al . . travel destination and tourist packs some of the reviewed systems focus on the recommendation of a destination that suits the user’s preferences. this is the case of systems like personaltour (lorenzi, loh, & abel, ), itchy feet ) and koceski and petrevska ( ) ), garcía-crespo et al. ( ), huang and bian ( ), lamsfus et al. ( ), lucas ontiel and aldana-montes ( ), yu and chang ( ), jannach et al. ( ), vansteenwegen et al. ( ), yang ( ), borràs et al. ( ), fenza et al. ( ), ), lorenzi et al. ( ), luberg et al. ( ), martin et al. ( ), montejo-ráez arcia et al. ( ), wang et al. ( ), batet et al. ( ), borràs et al. ( a), , braunhofer et al. , garcia et al. ( ), gyorodi et al. , kurata & hara ( ), ruotsalo et al. ( ), savir et al. ( ), umanets et al. ( ) and yang ), garcía-crespo et al. ( ), huang and bian ( ), lucas et al. ( ), lee et al. z et al. ( ), sebastià et al. ( ), vansteenwegen et al. ( ), kurata ( ), ( ), rey-lópez et al. ( ), sebastià et al. ( ), garcia et al. ( ), wang & hara , lucas et al. ( ) and savir et al. ( ) al. ( ), vansteenwegen et al. ( ), rey-lópez et al. ( ), sebastià et al. . ( ), umanets et al. ( ) and yang and hwang ( ) j. borràs et al. / expert systems with applications ( ) – (seidel, gärtner, pöttler, berger, & dittenbach, ) and mytravelpal (koceski & petrevska, ). personaltour is used for travel agencies to help their costumers to find the best travel pack- ages according to their preferences. once the recommendation process is finished, a rated list of options is presented to the cos- tumer. table shows an example of the hotel recommendation service. after that, the customer can rate each item of each travel service. itchy feet not only recommends tourism destinations but also provides purchasing services for booking a trip and assistance from professional travel agents. users make search requests, which are handled by autonomous agents that search for information in the internal database as well as in external data sources. the results are shown to the user through the interface, where recommended items (flights and hotels) can be selected and purchased. mytravelpal (koceski & petrevska, ) first recommends areas of interest over a region graphically (see fig. ), where the size of the circle indicates the level of affinity with the user. once the user focuses on a particular area, their tourist resources are also shown and sized depending on the affinity to the user profile. . . ranked list of suggested attractions most tourist recommender systems tend to suggest places once the user has decided the destination of the trip or she is already there. these systems are more complex, since they can suggest a large number of attractions, accommodations, restaurants or even temporal events. in this context the capability of recommenders to fig. . mytravelpal – recommend table example of hotel recommendation in personaltour (adapted from (lorenzi et al. ( )).) id hotel name city hotel category libertel paris economic palladium punta cana resort amadeus milan economic riu palace cancun first westin aruba economic classify and rank only those elements considered important for a particular user among the huge quantity of available information is very useful. with the support of these systems the users can find interesting places in an efficient way and even discover unex- pected ones that may be of their interest. the activities to be rec- ommended are normally stored in a static database, although some systems (e.g. otium, montejo-ráez et al., ) extract auto- matically information about events from the web to ensure that they always provide updated information. this kind of recommender systems (e.g. borràs et al., ; fenza, fischetti, furno, & loia, ; ruiz-montiel & aldana- montes, ; sebastià et al., ) usually provide a list of activ- ities that match the user profile, have been visited and/or posi- tively evaluated by similar users in the past, or are similar to activities previously enjoyed by the user. thus, they include mech- anisms to compare the user preferences with the features of an object, or to compare the similarities between two users or two objects (more details on the recommendation techniques are given in section ). the selection of the recommended items may also take into account contextual factors, like the present location of the user (noguera et al., ). some systems are also capable of justifying the provided recommendations (e.g. jannach et al., ). smartmuseum (ruotsalo et al., ) is an example of a more complex recommendation system, which detects automatically if the user is outdoors or indoors, based on her location. for the first case, it can display the recommendations on a map. for indoor sce- narios, it gives a list of the most relevant objects according to the ation of regions of interest. . room category room type swimming pool wifi standard double no yes luxe double yes yes standard single no yes luxe double yes yes luxe double yes yes fig. . list of recommendations of specific objects for indoor scenarios. j. borràs et al. / expert systems with applications ( ) – user’s preferences. this is useful, for instance, in museums, where the number of objects to see may be relatively high (fig. ). . . planning a route there are several projects that not only provide a list of the places that fit better with the user’s preferences but also help tour- ists to create a route through several attractions. ct-planner (kurata, ; kurata & hara, ) offers tour plans, as shown in fig. , that are refined gradually as the user’s expresses her preferences and requests (duration, walking speed, reluctance to walk, etc.). it displays a radar chart that represents the user’s preferences and a cartoon character as a navigator, in order to enrich the sense of user-friendliness and interactivity. there are several systems that provide an initial set of recom- mended activities (or an initial plan), with which the user can directly interact to add more activities, remove activities, select an activity to be visited, change the order of visit, etc. the planning component of the recommender system takes into account impor- tant factors like the expected duration of the visit, the opening and closing times of the attractions and the distance between them. some relevant examples include enosigtur (borràs et al., a), city trip planner (vansteenwegen et al., ), cruzar (mínguez, berrueta, & polo, ), smart city (luberg, tammet, & järv, ), otium (montejo-ráez et al., ) and e-tourism (sebastià et al., ). a more detailed review of trip planning functionalities is available in (vansteenwegen & souffriau, ). some advanced recommenders, like samap (castillo et al., ) and patac (ceccaroni, codina, palau, & pous, ), are capable of analysing the connection possibilities between the activities using different means of transport (walking, by bike, by car, or by public transport). some of these systems incorporate more complex geographical information systems (gis) to manage the geographical data associ- ated to the touristic points and events. (huang & bian, ) argued that it is computationally unfeasible to maintain large amounts of spatial data and use them in planning procedures. hence, they used existing geospatial web service technologies, in concrete the esri arcweb service, to obtain the location of the attractions, the distance between them given their street address, and driving directions between two attractions. geoasis (martínez- santiago et al., ) continually calculates the position and the speed of the user. the estimated time to reach a place is considered in order to create the plan in real time. the key aspect is the predic- tion of where the user will be in the immediate future: in a city, near a city or on the road. if the user is already in a city, the planning algo- rithm checks the nearest places to the user without taking into http://www.esri.com/news/arcuser/ /arcweb.html (last access march ). account the route or the speed, since it is considered that the user is close to them. if the user is near a city, the planning algorithm checks the most relevant attractions in it. if the user is on the road, but not near a city, then the planning algorithm is more complex because it considers temporal constraints. the plan is not computed by the server but by the client application, since it is constantly checking the location by gps. routes are computed using google maps as an external resource. once the visit plan has been completely defined, the user may wish to retrieve the full schedule to follow the route. this retrieval can take different forms. systems like samap (castillo et al., ) or enosigtur (borràs et al., a) allow downloading a pdf file that contains a geo-referenced map with a detailed explanation of the plan. in others, like city trip planner (vansteenwegen et al., ) and otium (montejo-ráez et al., ), the user can download the route to a mobile phone. . . social aspects several projects (e.g. ceccaroni et al., ; garcia, torre, & linaza, ; umanets et al., ; vansteenwegen et al., ) have paid special attention to the inclusion of social functionalities that allow users to share material (pictures, comments, evalua- tions) and interact with other tourists. these aspects may be very interesting to help to promote the use of a recommender among the visitors of a particular city. recommenders like moretourism (rey-lópez et al., ) and itchy feet (seidel et al., ) allow users not only to interact over popular social networks but also to create location-based activity groups that can be employed to post comments, join groups for doing common activities or interact with other users. the system e-tourism (garcia, sebastià, & onaindia, ) allows to create plans that accommodate the pref- erences of a whole group of visitors. in itravel (yang & hwang, ) users communicate among them with mobile peer-to-peer communications to send ratings of attractions. their navigation map not only displays the location of attractions but also the position of near-by users with which it is possible to communicate. fig. shows a map with recommended attractions (green pins) and nearby users (blue pins). the visit system (meehan, lunney, curran, & mccaughey, ) applies sentiment analysis techniques (using the alchemy api ) to analyse the updates about a given attraction in twitter and facebook and identify if users are expressing positive or negative comments about it. this information is shown with green and red colours by the system in its interface, so that the user may easily identify those places visitors are loving most today and which not. . recommendation techniques in e-tourism recommender systems have been classically classified, accord- ing to the way in which they analyze the information of the user and filter the list of items, into content-based, collaborative and demographic systems (burke, ; manouselis & costopoulou, ; montaner, lópez, & de la rosa, ). in this section we introduce these three paradigms, analyzing its use in current tour- ism recommender systems. in addition, we present the different user models that have been employed in those systems. . . approaches to recommendation content-based (cb) systems calculate a degree of similarity between the users and the items to be recommended. the process alchemyapi ( ). alchemy api: transforming text into knowledge. [online] http://www.alchemyapi.com/. http://www.esri.com/news/arcuser/ /arcweb.html http://www.alchemyapi.com/ fig. . ct-planner user interface (from (kurata, )). fig. . navigation map of itravel (from yang & hwang, ). j. borràs et al. / expert systems with applications ( ) – is carried out by comparing the features of the item with respect to the user’s preferences. so, it is assumed that both users and alter- natives share a common representation (e.g., they are composed of the same set of attributes or keywords). the output of the compar- ison process is usually an overall performance score, which indi- cates the degree of matching between the user’s profile and each alternative. the higher the score is, the higher the performance of the alternative for a given user. sometimes these methods also take into account the rating history of the user. in this approach, the recommendation system relies on having an accurate knowl- edge of the user’s preferences to be able to select the appropriate items. this kind of approaches may suffer from the ‘‘cold start’’ problem when a new user enters in the system, because we can elicit poor knowledge about the user in an initial stage. some solutions to this problem are explained later in this section. in cb systems the recommendation process is mainly focused on defining an appropriate measure to compare a user and an item. the two most common approaches are the aggregation of ratings and distance calculation. � when the user profile is represented as a rating vector with the degree of interest of the user in each attribute, each rating can be interpreted as a performance score that can be used to eval- uate an alternative. the goal is then to calculate an overall inter- est score for a certain alternative. the simplest approach consists of using an aggregation operator to combine the user ratings on the concepts that define a certain alternative (batet et al., ). more sophisticated aggregation methods have also been applied, like ahp (analytic hierarchy process) (huang & bian, ; niaraki & kim, ). � when the items and the users are described by a list of key- words, some similarity measures can be applied. for example in lamsfus et al. ( ) items and users are described using concepts from an ontology, which defines archetypes of tourists (e.g. cultural, sportive or adventurous), and the cosine similarity between the two vectors (user and item) is calculated. a similar approach is proposed in gyorodi, gyorodi, and dersidan ( ) with ad-hoc hierarchies of tags for locations that are rated by users. the locations ratings are then compared to the user’s tags. in garcía-crespo et al. ( ) a feature-based similarity algorithm is applied, using several ontologies as reference. in fenza et al. ( ) classification rules are automatically gener- ated and later used to define the matching degree between the user and the item. j. borràs et al. / expert systems with applications ( ) – other methods of recommendation exploiting ai techniques are explained in section , such as the ones using rules or probabilistic information. collaborative (cl) systems make recommendations based on groups of users with similar preferences. the similarity between users is normally computed by comparing the ratings that they give to some of the items. when the system identifies who are the people that share similar interests with the current user, then the items that those people liked are recommended to this user. in this approach, some feedback about the provided recommenda- tions is necessary, in order to know which items the user has liked or disliked (e.g. which places she has enjoyed visiting). two types of cl methods are distinguished: user-based and item-based. the former finds neighbours of a target user by matching her opinions with the ones of the other users in the system. the latter builds groups by finding similarities on the items that the users liked (or disliked) in the past. two weak points are recognised in cl systems: ‘‘data sparsity’’ and ‘‘grey sheep’’. the former occurs when the number of ratings from users is small in comparison with the total number of items, so that the probability of finding users that rate the same items is too low to make good estimations. the latter, ‘‘grey sheep’’, refers to a user with a profile different from the rest of users of the sys- tem. in this case, it is difficult to find appropriate items to recom- mend because we do not have information about similar users. finally, this approach also suffers from the scalability problem if the community of users is large. demographic-based (dm) systems rely on the demographic data of the user (e.g. age, country of origin, level of studies, etc.). in this case, the recommendation is not based on the user’s interests and preferences but on her personal characteristics. in this approach, users are usually assigned to a certain stereotypical class depend- ing on their demographic data, so that the members of the same group share a common demographic profile. the system has inter- nal knowledge about the standard preferences of each stereotype, which is used to provide the recommendations to the users. the definition of stereotypes of tourists is not new in this field. many studies have defined segments of tourists according to their behav- iour in different cities or territories (brewer, ; marques, ; tsung-chiung, chyong-ru, & wan-chen, ). specific stereo- types provide precise descriptions of what tourists want and how they act in different situations. this information is normally used as a guide to conduct business with tourists, but it can also be exploited in recommender systems. since each of the approaches has some drawbacks, the combi- nation of different techniques is also a widespread practice. fig. shows the distribution of the types of recommendation techniques used in the field of tourism recommenders, in percent- ages. half of the works use a mixture of techniques ( %), combining fig. . recommendation techniques used in the reviewed works (in%). mainly cb methods with cl filtering or with dm techniques. the rest of the systems apply a single approach, having a clear predominance for the techniques based exclusively on the description of the con- tent of the alternatives ( % of the reviewed papers). hybrid systems can integrate these techniques in different ways. three approaches can be distinguished: . selection of the method: the system incorporates dm, cb and cl methods, but only one of them is applied depending on the par- ticular situation of each user. for example, the first time the user arrives, a method based on demographic data is used. later on, if similar users can be found, a cl recommendation is made. otherwise, a cb procedure is applied. this is the case of martínez, rodríguez, and espinilla ( ), huang and bian ( ) or noguera et al. ( ). . sequential use: each recommendation technique is used in dif- ferent stages of the process. for example, speta (garcía-crespo et al., ) has four steps: first, contextual information such as the location or time are used to make the first selection of appropriate options; second, a more fine grained set of results is obtained using knowledge-based filtering techniques, by cal- culating the semantic similarity between the user preferences and the touristic services; third, preferences and cl techniques are used to refine the set of options; finally, in the fourth step, a vector of preferences is used to make the final selection. in braunhofer et al. ( ) cl filtering with dm and personal information is applied in a training phase to build a prediction model in different contexts. once the model has been trained, cb techniques generate the list of recommendations by com- puting ratings for each item based on the current and predicted values. . integrated use: both cb and cl techniques are combined during the execution. for example, in sigtur (borràs et al., ) differ- ent ratings are calculated to estimate the interest of an activity for a target user. ratings are obtained using dm-clustering, cl-clustering and cb similarity; afterwards, these ratings are merged to find an overall quality rating for each item and make the filtering of the best tourist attractions. in lucas et al. ( ) the users are classified into groups using simultaneously per- sonal demographic data (dm), information about the content of the items previously selected by the user (cb) and the infor- mation of other users (cl). then a set of fuzzy rules is automat- ically generated so that new users can be automatically classified into several groups (with different membership degrees). the list or recommended items is finally derived from a prediction based on the groups the user belongs to. in the survey we have observed an increasing trend in the exploitation of cl filtering techniques since , mainly in hybrid systems. more precisely, from to only % of systems used such method, whereas since the percentage has increased to % (see table ). . . representation of the user profile a key component in recommender systems is the user profile, which stores the information related to the user preferences and permits to make personalised recommendations. different user profile models have been developed. the simplest model associates to each user a list of preferred keywords or categories in which items are pre-classified. however, this information is usually too general to provide accurate recommendations. a more widespread approach consists of storing a vector with numerical ratings corre- sponding to the attributes of the items. this rating indicates the degree of interest of the user with respect to each attribute. this vector approach facilitates the inclusion of other types of features table review of recommendation methods used. recommendation method reference all lucas et al. ( ), ruiz-montiel and aldana-montes ( ), borràs et al. ( , a), batet et al. ( ), koceski and petrevska ( ), braunhofer et al. , garcia et al. ( ), lucas et al. ( ) and meehan et al. ( ) cb + cl castillo et al. ( ), garcía-crespo et al. ( ), fenza et al. ( ), rey-lópez et al. ( ), noguera et al. ( ) and rojas & uribe cb + dm coelho et al. ( ), ceccaroni et al. ( ), lamsfus et al. ( ), niaraki and kim ( ), mínguez et al. ( ), yang ( ), martin et al. ( ), sebastià et al. ( ) and garcia et al. ( ) cl + dm gavalas and kenteris ( ) cb huang and bian ( ), lee et al. ( ), seidel et al. ( ), yu and chang ( ), jannach et al., ( ), ricci et al. ( ), sebastià et al. ( ), vansteenwegen et al. ( ), garcía-crespo et al. ( ), kurata ( ), linaza et al. ( ), lorenzi et al. ( ), luberg et al. ( ), montejo- ráez et al. ( ), martínez-santiago et al. ( ), gyorodi et al. ( ), kurata and hara ( ) and ruotsalo et al. ( ) cl savir et al. ( ), umanets et al. ( ) and yang and hwang ( ) dm wang et al. ( ) j. borràs et al. / expert systems with applications ( ) – in the profile, such as demographic information, as can be seen in codina and ceccaroni ( ), garcia et al. ( ), mínguez et al. ( ) and kurata and hara ( ). these basic models can be extended by means of some ai knowledge representation techniques. one possibility is to use semantic models, in which the domain knowledge usually takes the form of an ontology. for example, the keyword list may contain names of classes in the ontology (i.e. concepts) and the ratings vec- tor can also be defined in the space of the concepts in the ontology. uncertainty models have also been incorporated into some recom- mender systems to be able to handle the credibility associated to the information stored in the profile. the rating values are uncer- tain in many situations, especially when they are not given explic- itly but have to be inferred from the user interaction with the domain items. a confidence degree can be associated to each rating in the profile and can be used as a weighting factor in the exploi- tation stage. more details about these techniques are given in sections . and . , respectively. in order to produce up-to-date personalised recommendations to the same user along time, the user model has to be updated. feedback information is used to modify the profile when any change on the user’s preferences is detected. this information can be collected explicitly or implicitly. explicit feedback is obtained by means of the direct interaction with the user. the deci- sion maker is requested to fill in some form (giving her opinion on different values of the criteria or indicating her location) or to rate a set of alternatives. this approach provides precise knowledge because the data is given directly by the user. however, it is usually considered quite an intrusive way of elicitation, and many users are not keen on spending time in answering this kind of questions. techniques based on implicit feedback aim at collecting the user information by analysing her behaviour in the system, such as table review of the ai techniques used. ai techniques references multi-agent systems castillo et al. ( ), ceccaroni et al. ( ), lee et al. ( optimization techniques castillo et al. ( ), lee et al. ( ), garcia et al. ( ), va ( ) automatic clustering castillo et al. ( ), garcía-crespo et al. ( ), martínez e noguera et al. ( ), lucas et al. ( ), kurata & hara management of uncertainty garcía-crespo et al. ( ), huang and bian ( ), lamsfu et al. ( ), lucas et al. ( ), meehan et al. ( ) and r knowledge representation castillo et al. ( ), ceccaroni et al. ( ), lamsfus et al. ( ( ), garcía-crespo et al. ( ), wang et al. ( ), alon ( ), ruotsalo et al. ( ) and moreno et al. ( ) the alternatives that are selected, purchased or viewed. more sophisticated tools study the sequence of actions done by the user on a certain alternative, or even the amount of time spent with each one. the main advantage of these methods is that an addi- tional effort from the user is not required. for example, in savir, brafman, and shani ( ), the travellers modify a trip plan and the system observes the modifications to adjust the budget, daily time, maximum travel distance, transportation constraints, etc. other works (albanese, chianese, d’acierno, moscato, & picariello, ; albanese, d’acierno, moscato, persia, & picariello, ; moscato, picariello, & rinaldi, ) analyze the sequence of acces- ses to information, to deduce which is the most accessed object after viewing a specific one. they use both low level (colour, tex- ture) and high level (author, type) features of objects to compare the usage behavioural patterns. despite the advantages of implicit feedback, this kind of data is more uncertain than explicit informa- tion, so less confidence must be given to it when the profile is mod- ified. in this review we observed that around % of the works use only explicit feedback, whereas the rest combine explicit and implicit feedback. the most common approach in explicit feedback collection consists of requiring the user to vote the different options proposed by the system. the most sophisticated approach is presented in the vibe system (jannach et al., ), which builds personalised dialogues to gather new knowledge about the user preferences. . use of ai techniques in tourism recommender systems this section makes a brief review of the main ai techniques and tools employed in tourism recommender systems in the last years, which are summarised in table . ), seidel et al. ( ), sebastià et al. ( ) and lorenzi et al. ( ) nsteenwegen et al. ( ), garcia, vansteenwegen, et al. ( ) and meehan et al. t al. ( ), fenza et al. ( ), gavalas and kenteris ( ), batet et al. ( ), , ruotsalo et al. ( ) and moreno et al. ( ) s et al. ( ), lamsfus et al. ( ), pinho et al. ( ), wang et al. ( ), hsu uotsalo et al. ( ) ), lee et al. ( ), sebastià et al. ( ), sebastià et al. ( ), garcia et al. so et al. ( ), batet et al. ( ), martínez-santiago et al. ( ), lucas et al. j. borràs et al. / expert systems with applications ( ) – . . multi-agent systems agents are autonomous and proactive software entities capable of obtaining information from their environment and acting in an intelligent way upon it in order to try to accomplish a set of goals or objectives. multi-agent systems are groups of agents that com- municate between themselves to share information and resources, coordinate their activities and cooperate in the joint efficient solu- tion of a distributed problem (wooldridge, ). turist@ (batet et al., ) is an agent-based system that pro- vides personalised recommendations on cultural activities. the architecture of the system is shown in fig. . there is one agent for each kind of cultural activity, which maintains a small database with the events of that type available in the city (museums are the exception, as there is one specific agent for each museum in the city). the user interacts with the system through a graphical inter- face provided by a user agent. a broker agent mediates the commu- nication between the user agents and the cultural activities agents. the user can make specific queries, can evaluate an activity that she has attended, or can ask for a personalised recommendation. the core of turist@ is the recommender agent, which maintains a user profile for each tourist. this profile is initialised with some basic information on high-level cultural interests provided by the user when she uses the system for the first time. the recommender agent dynamically and automatically refines this initial knowledge about the user preferences by analysing the user’s queries and evaluations. the user agent can also provide proactive recommen- dations, because it knows the position of the user in the city and can suggest cultural activities that fit the user’s preferences and are located in the vicinity. the system uses both cb and cl recom- mendation techniques. the idea of having an initial profile and refining it by analysing the explicit (evaluations) and implicit (actions) activities of the tourist is also given in ceccaroni et al. ( ). that work proposes to have a profile management agent, which not only initializes the profile (by fitting the user into stereotyped classes) but also mod- ifies it depending on the feedback provided by the tourist. in this agent-based proposal there are information service agents that retrieve touristic information from databases and ontologies, and a personalization agent that, given the user profile and the available touristic data, applies cb recommendation techniques to select the items that should be suggested. in personaltour (lorenzi et al., ) there is a set of travel agents, and each of them is specialised in the recommendation of flights, hotels or attractions. when a new costumer arrives and fig. . multi-agent based architecture expresses her preferences, these agents collaborate among them- selves in order to propose a travel package to the tourist. the user can later evaluate each of the components of the package, provid- ing a feedback to the system so that the degree of expertise of each travel agent can be conveniently updated. some recommenders (e.g. castillo et al., ; lee et al., ; sebastià, giret, & garcia, ) ‘‘agentify’’ the different compo- nents of the system (the interface with the user, the capture of her requirements and preferences, the analysis of the suitability of each attraction, the creation of a route among the selected points of interest), although there is not any kind of complex communica- tion or coordination between them. in all these systems the agents seem to work in a sequential fashion, without any kind of coordi- nated effort. therefore, the full potential of distributed, concurrent and coordinated behaviour of agents is not employed. . . optimization techniques many tourism recommender systems have to solve complex planning and scheduling problems, which are well known to be np complete and, therefore, cannot be optimally solved in an effi- cient way. in some cases, researchers have opted for the use of dif- ferent kinds of optimization techniques which, although in many cases they do not guarantee the optimal solution, offer an afford- able computational cost. one example is the agent-based travel route recommender for tainan (lee et al., ), that uses ant colony optimization tech- niques. in these methods a set of autonomous entities (which rep- resent the ants) cooperate through pheromone-mediated indirect and global communication to find a good solution to the travelling salesman problem (in this case, to plan a route that goes through different points of interest around the city). ct-planner (kurata & hara, ) uses a genetic algorithm to construct the plan to visit a city. in each iteration of a cyclic process it considers a population of different possible plans, which are evaluated according to their utility for the user; the best ones are mutated and recombined via crossover to generate another population for the next iteration. after a certain number of iterations, the best plan is finally selected. the authors of the visit system (meehan et al., ) pro- pose to make recommendations adapted to the context of the user, that is composed of different factors (location, time, weather, social media sentiment and user preferences). in that work, they suggest the idea of using an artificial neural network to assess the relevance of each context component for each user. of turist@ (from batet et al., ). j. borràs et al. / expert systems with applications ( ) – some heuristic procedures to build travel itineraries were explored in the city trip planner system and related works (garcia, arbelaitz, linaza, vansteenwegen, & souffriau, ; garcia, vansteenwegen, arbelaitz, souffriau, & linaza, ; vansteenwegen et al., ). one possibility is the use of iterated local search, a meta-heuristic iterative method that builds sequences of solutions generated by a local search. the heuristic perturbs the solution found by the local search (a route to visit some city attractions) to create a new solution. then, it takes the best solution as the new starting solution for the local search. the process is repeated until a termination criterion is met. another option that was studied is the use of meta-heuristic itera- tive greedy randomised adaptive search methods (souffriau, vansteenwegen, vanden berghe, & van oudheusden, ). in each iteration a list of possible visits is generated from an initial solution which contains only the start and end of each tour. those visits that have a heuristic value below a certain threshold are eliminated. a random visit from the remaining list is selected and applied to the current solution. most of the tourism recommender systems that build personalised routes or itineraries implement an ad-hoc planning mechanism, but some of them apply more classical domain- independent ai planning techniques. for instance, in the samap system (castillo et al., ) the use of heuristic, a⁄ and hierarchi- cal temporal planners was explored. . . automatic clustering many tourism recommenders employ techniques based on cl filtering, in which the users of the system are partitioned into groups that share some common characteristics. the basic idea of these methods is that it can be appropriate to recommend to the fig. . portion of the samap domain ontol user those items that have been positively valued by similar tour- ists. the concept of similarity employed to group users may be based on demographic information, on the general preferences of the users over diverse types of touristic activities, or on the explicit ratings of individual activities. in any case, the automatic cluster- ing tools developed in ai may be successfully used to classify the tourists. this section comments different alternatives that have been used in touristic recommender systems. a very simple way of associating a new user with similar past users of the system is to employ the k-nearest neighbours approach (dasarathy, ), calculating which are the k past users of the sys- tem who were more similar to the current one (e.g. martínez et al., ; noguera et al., ). having done that, the information on those users may be employed to provide recommendations (e.g., the activities that were more highly valued for them). in samap (castillo et al., ) the similarity between users is based on the preferences expressed over the concepts of a domain ontology (a portion of it may be seen in fig. ). for instance, the system could easily infer that a user that likes cinema is more similar to a user that enjoys theatre than to another that prefers sport activities. scalability is one of the main problems to be addressed when using this method. a common option to group the users into different classes is to use the k-means algorithm (e.g. gavalas & kenteris, ; pinho, da silva, moreno, de almeida, & lopes, ). the initial seeds of the k desired clusters are established in some application-dependent way. then there is an iterative process in which, in every step, the objects are sorted into the nearest cluster and the cluster pro- totypes are recalculated. the method converges to a solution when the objects belong to the same clusters in two consecutive itera- tions. in moreno et al. ( ) the k-means algorithm is applied with three different purposes: to obtain a set of initial tourist segments, ogy (adapted from castillo et al., ). j. borràs et al. / expert systems with applications ( ) – to obtain classes of users with similar demographic characteristics, and to classify users according to the explicit ratings they have pro- vided. in the first case, the historical data of , questionnaires was used to compute a set of generic tourist types (segments). each of them had an associated prototype and a preference level for each kind of activity. new users fill a small form with some basic personal and preference data and it is possible to compute the segment to which they belong, guiding the initial recommen- dations. the demographic classification is mainly based on the country of origin, the travel group composition, the travel budget and the accommodation type. aggregation operators like owa (yager, ) and lsp (dujmovic & nagashima, ) are used to join the similarities on these different kinds of data. the classifica- tion of the users according to their ratings uses the well-known pearson correlation as similarity measure. the recommender system described in fenza et al. ( ) pro- poses the use of the uncertain version of k-means, fuzzy c-means. the result of this algorithm is a fuzzy partition of a set of objects into clusters, so that each object has a degree of membership between and to each cluster, and the addition of the degrees of membership to all the clusters is . this algorithm is both applied to users and to touristic points of interest (pois). after the definition of clusters of users and pois, the system is able to derive rules that characterise them, that are used to integrate new users and new pois to the clusters in which they fit better. this work also proposes to build association rules, which explain the relationship between clusters of users (plus contextual infor- mation) and clusters of pois. these rules permit to determine the kind of touristic activities that should be recommended to a certain type of users. very similar techniques are employed in the psis (personalised sightseeing information system) recommender (lucas et al., ). turist@ (batet et al., ) also employs cl filtering recommen- dation techniques that require the definition of classes of similar users. the clustering is applied every time that new users join the system, so classes are periodically recomputed. the employed clustering system is clusdm (valls, ), which builds a hierarchy of classes taking into account the interests of the users in general kinds of activities and their demographic data. the tree generated by the algorithm can be cut at different levels to generate parti- tions with the desired number of classes. the use of support vector machines (svms) as a classification technique in tourism recommender system is suggested in the speta system (garcía-crespo et al., ). tourist preferences on several kinds of tourist activities are stored in a vector, and the characteristics of each activity are also stored in the same way. thus, svms may be used to compute the distance between the fig. . use of a bayesian network to detect the preferred user’s preferences and the recommendable items, so that the most appropriate ones can be efficiently found. . . management of uncertainty the task of recommending activities to a tourist is not simple, as there is not any clear and precise relationship between the charac- teristics and preferences of a visitor and the pois available at a given destination. some of the techniques developed in the ai field of approximate reasoning have been proposed to represent and rea- son about this uncertain relationship. one possibility is to use bayesian networks (pearl, ). a bayesian network is an acyclic graph in which edges represent relationships of causality or influence between nodes. nodes that do not have any parent have an associated probability table, indi- cating how likely they are to occur. nodes that have n parents have a conditional probability table of n nodes, indicating how likely they are to occur depending on the presence (or absence) of their parents. a very simple use of bayesian networks is presented in hsu, lin, and ho ( ), where a number of attributes (age, nation- ality, occupation, income, travel motivation, etc.) influence directly on the probability that a certain touristic point is interesting for the user. the initial bayesian network was built after the analysis of more than questionnaires. a more complex application of this kind of networks is given in huang and bian ( ) and wang et al. ( ). they propose a network (see fig. ) in which the age, occupation and personality influence the type of user which, along with the travel motivation, influences the probability of the user liking a certain kind of touristic destinations. specific touristic events are not included in the network. another common option to manage uncertainty is the use of fuzzy logic. a fuzzy variable make take as values a series of linguistic labels. each linguistic label has an associated fuzzy set, in which every value in the domain of reference is assigned a membership value to the set between and . in that way, fuzzy logic provides a generalisation of standard logic. fuzzy sets and fuzzy reasoning may be used to represent the preferences of the user and to calcu- late how they fit with the characteristics of a tourist attraction (garcía-crespo et al., ; lee et al., ), to obtain the degree of membership of each user to different groups of users (pinho et al., ) or to represent contextual aspects of the journey (meehan et al., ). for instance, if the weather conditions are represented with a value between and , instead of using a simple boolean value for good/bad weather, it is possible to make a more fine grained analysis of the weather conditions and reason about its influence on the recommendation of each cultural activity. kind of tourist activities (from huang & bian, ). j. borràs et al. / expert systems with applications ( ) – some touristic recommender systems also employ a rule-based approach, but without the addition of a fuzzy component. for instance, in the concert system (lamsfus et al., ; lamsfus et al., ) there are rules that detect the events to be recom- mended depending on the user preferences and the context, such as this one: hasfoodpreferencesrule: (v? red:type dcl:visitor), (?v dcl:haspre- ferences ?p), (?p red:type dcl:foodpreferencesdemographics), (?v dcl:usesdevice ?d), (?d dcl:isconnectedtonetwork ?n), (?n dcl:haslo- cation ?l), (?l dcl:hasenvironment ?e), (?e dcl:offerskindoftourism- concepts ?s), (?s dcl:isrestaurantoftypevegetarian ?r) => print (?r dcl:istourismserviceofferedtovisitor ?v) . . knowledge representation recommender systems in e-tourism need, as any knowledge- based intelligent system, a way to represent in an efficient way the domain knowledge, so that it can be used in their reasoning processes. the knowledge representation and reasoning tech- niques developed in ai are adequate tools for this purpose. in par- ticular, nowadays the most common way of representing domain knowledge is the use of ontologies. an ontology describes a shared and explicit formal conceptualization of a given domain. its main components are classes (representing concepts, usually organised in some kind of hierarchical structure), taxonomical and non-taxo- nomical relationships (sánchez & moreno, a, b), axioms (sánchez, moreno, & del vasto-terrientes, ) and instances (representing specific objects). there are several tourism recommenders that employ ontolo- gies to formalize the domain knowledge. most systems have gen- eric ontologies that store information about different aspects that have to be taken into account in the recommendation of cultural activities. for example, the system described in wang et al. ( ) includes a generic travel ontology (with information about accommodation, restaurants, transport, shopping, culture, etc.) and a user ontology in which the demographic characteristics of the tourists and their preferences are modelled. other systems like geoasis (martínez-santiago et al., ), samap (castillo et al., ) (see fig. ), smartmuseum (ruotsalo et al., ) and the one proposed in alonso et al. ( ) also include ontologies to model the different kinds of touristic activities and to be able fig. . portion of the tourism ontology devel to reason on them in a semantic fashion. these systems use ontol- ogy-based similarity measures to deduce if two kinds of activities are similar, and this knowledge may also be used to compute the similarity between users and provide recommendations based on cl filtering techniques. two systems that make a heavy use of ontologies are sigtur (borràs et al., ; moreno et al., ) and e-tourism (garcia et al., ; sebastià et al., ; sebastià et al., ). they use domain ontologies that detail the different kinds of leisure activi- ties available in the city. for instance, fig. shows a portion of the sigtur ontology, which has more than concepts organised in a hierarchy of levels. sigtur stores in each node of the ontology the preference of the user (and the confidence of the system on that preference). these values are initialised with a small questionnaire filled by the user when entering the system (preferences on the most generic con- cepts are transmitted through their children to all the hierarchy). when the user performs actions on the recommended activities (e.g. adds an activity to the current travel plan) the system updates the preferences on the concepts associated to that activity and spreads this information through their parents in the ontology. thus, there is an ontology-based dynamic management of the user preferences (borràs, valls, moreno, & isern, b). the e-tourism system follows a similar approach, in which the user preferences are updated after analysing the explicit ratings she has provided. there are systems that propose the use of a set of ontologies, instead of having a single integrated ontology. patac (ceccaroni et al., ) includes separate ontologies about cultural activities, restaurants, entertainment, hotels, etc. the authors propose to link those ontologies with standard temporal and geo-location ontolo- gies provided by the w c consortium and with a user model ontology that contains different kinds of touristic stereotypes. concert (lamsfus et al., ) models all the context associated to a travel in a network of ontologies called contology, which includes separate ontologies that deal with aspects like tourism services, preferences, activities or travel motivations. normally the ontologies used by recommender systems are designed ad-hoc for a specific application and built manually. a way to reduce the cost of the construction of the ontology (ruíz- martínez, miñarro, castellanos, garcía, & valencia, ) is to pop- ulate it in automatic fashion, by analysing electronic resources (e.g. oped in sigtur (from moreno et al., ). j. borràs et al. / expert systems with applications ( ) – web pages), extracting the appropriate information about tourist activities and creating the associated instances. a similar proposal was made in vicient, sánchez, and moreno ( ). . related reviews there have been some previous papers explaining the applica- tion of recommender systems in the tourism area, which are chro- nologically mentioned in this section. ricci ( ) and staab and werthner ( ) explain in a very generic way the characteristics of travel recommender systems with some examples, without making an exhaustive review or comparing different approaches. werthner ( ) gives also a very generic description of technolog- ical approaches applied to tourism, where some examples related to artificial intelligence are mentioned. however, it does not pro- vide any review or comparative analysis of different systems. berka and plößnig ( ) provides a brief guide on how to design recommmender systems for tourism, but it does not attempt to make a survey of the area either. the survey that is more similar to this work is kabassi ( ). it is mainly a classification of tourism recommender systems (until early ) under different criteria: kind of objects they recommend (hotels, flights, restau- rants, etc.), hardware support (computer or handheld device), indi- vidual/group recommendations, explicit/implicit acquisition of information from the user, recommendation technique (content- based, collaborative or hybrid) and personalization techniques (mainly decision-making tools and bayesian networks). the authors of that paper basically group the systems in these catego- ries, without making a deep analysis or explanation of all these possibilities. it does not provide any guideline on how to build this kind of systems and it does not consider the latest advances in the last five years, which are the basis of our study (advanced geolocal- isation capabilities of mobile phones and tablets, context-aware recommendations, semantic management of preferences, use of social networks, etc.). gretzel ( ) makes an analysis of tourism recommenders from the point of view of social sciences, not from the technological perspective. the author of this paper argues that intelligent systems are necessary in the tourism domain because there are many complex aspects to be managed: the mobility of tourists, the increased risk and uncertainty experienced in unfa- miliar environments, the distributed nature of information sources, the idiosyncratic quality of tourism decision-making, the multi- faceted nature of tourism experiences, and the interdependency of suddecisions. a description of some systems that tackle those issues is done. the author also comments the main issues on the design and the evaluation of those systems, focusing on the user interaction, the context, the social perspective and the decision making process to maximise tourists utility; however, this work does not cover the use of intelligent techniques. the main recommendation methods applied in tourism are reviewed in felfernig et al. ( ). this paper presents some exam- ples of the use of these techniques, but they are not deeply described nor compared. this paper emphasises some interesting topics like group recommendation and context-aware recommen- dations in mobile devices. finally, it is worth mentioning the paper (vansteenwegen & souffriau, ), that makes a deep overview of systems built between and that compose trip plans, although they only comment this single functionality. the authors compare each of the reviewed references in terms of these plan- ning functionalities: personal interest estimation, selection and routing, mandatory points of interest, dynamic recalculation (update plan in real time when unexpected events occur), multiple day decision support (enable plans for multiple days), opening hours, budget limitations, max-n type (limitation of activity types per day), mandatory types, weather dependency, scenic routes (build paths with beautiful views rather than the shortest ones), hotel selection, public transportation and group profiles. this paper describes how the orienteering problem and its extensions can be used to model trip planning functionalities. in summary, as far as we know, there is not any recent survey of tourism recommenders with the technological focus, novelty and breadth of coverage of this paper. . conclusions this final section contains a brief summary of the work pre- sented in this paper, describing some points that should be taken into account by scientists aiming to design and develop tourism recommender systems, and an outline of several lines of future work. . . summary and guidelines to develop tourism recommenders tourism recommender systems give personalised and relevant suggestions to tourists whenever they visit unknown places. they provide support tools to make the process of deciding what to do more manageable. in this article we have reviewed tourism recom- mender systems published mainly in ai-related scientific journals and conferences since . we first analysed the interfaces used by these systems and we pointed out the predominance of web-based approaches, which are especially useful for tourists when they are planning a visit before the stay. however, lately the usage of mobile platforms has widely increased, since they allow a direct access to the infor- mation about attractions during the stay. moreover, they also per- mit to personalise and contextualise the gathered information, for instance taking the current location of the user into account. how- ever, we have noticed that new mobile platforms such as android or iphone have been weakly exploited. since these platforms are currently being widely used for tourists when visiting places, it is necessary to address the development of applications for those systems and to create responsive web designs that permit to adapt the content to any viewing device. tourism recommender systems, as we have seen, not only manage textual information, but most of them use images, pictures and interactive maps. we have also analysed the main functionalities of the reviewed tourism recommender systems. most of the approaches suggest points of interest in a destination according to the user prefer- ences. these suggestions may be shown in a ranked list ordered by importance or they may be integrated in a scheduled plan. usu- ally the system provides only a list and a planner that the user can employ to build manually the detailed plan. however, some approaches can make this process automatically, hence providing an almost complete route taking the context of the user into account, such as location, time and opening hours, among others. this is a novel issue and it may be extended in the future with more functionalities and contextual elements. another aspect added in the last years is the use of social features that allow tourists to share information among friends. the recommendation process is a crucial aspect in tourism advi- sory systems, hence we have analysed the main mechanisms used in the reviewed articles. the most popular approaches use content- based, collaborative and demographic-based techniques. these techniques suffer from several problems when applied individu- ally. hence, a good practice is the combination of several techniques together to overcome the mentioned drawbacks. we observed that the most recent approaches follow this trend and propose hybrid recommendation methods, including also contex- tual information. however, there are still a large number of approaches that apply only content-based methods. in tourism j. borràs et al. / expert systems with applications ( ) – recommender systems it is also a good practice to include a diver- sification mechanism to avoid content-based problems and to offer unexpected and off-the-beaten-track alternatives to tourists. recommenders need to learn and manage the user profile to make accurate personalised suggestions. this user profile can be repre- sented by a vector of words or be extended with semantic and uncertainty models which offer a richer representation of the user’s preferences. these preferences can be acquired explicitly or implicitly. the most common method is the acquisition of explicit information. however, we consider interesting to apply a combination of both methods. even though implicit information is inherently more uncertain, it is also less intrusive for users and it is easy to collect it directly by monitoring the interaction of the users with the system. we have also shown how a wide range of artificial intelligence techniques are already being employed in e-tourism recommend- ers. knowledge representation and reasoning techniques are com- monly used to represent and reason about the tourism domain. in this field it is especially promising the introduction of semantic measures of similarity between users and items (or between users), taking advantage of detailed ontological representations of the domain. intelligent autonomous agents may be used to provide proactive recommendations depending on the context; therefore, their use may be especially relevant in the case of mobile-based recommenders. approximate reasoning techniques are applied to manage the uncertainty on the user’s preferences, making them an ideal option when the user feedback is only impli- cit. reasoning procedures of different kinds may be employed to infer the user’s preferences. automatic clustering algorithms may be successfully used to classify tourists with similar tastes or sim- ilar features. finally, optimization techniques offer cost-effective solutions to complex planning and scheduling problems when the system wants to build automatically a multi-day route. as a result of this analytic review, some basic guidelines that can be followed in the design and development of tourism recom- mender systems may be given: � the use of both a web and a mobile version of a recommender system can be suggested as a good practice, since the web ver- sion can be comfortably used at the home of the tourist to select the destination, the activities to be carried out and a global planning of the routes, whereas the mobile version (that should run on the latest and most widely spread platforms) may permit to follow the route in situ, receive more detailed information about the visited places, receive proactive recommendations depending on the user preferences, or even share the travel experience through social networks. the mobile version should provide a rich graphical interaction with the user, to enhance the tourist experience (e.g. augmented reality approaches) and also exploit heavily contextual information (e.g. the current time, the location of the user, etc.) to improve the accuracy of the recommendations and adapt it to the dynamically changing circumstances of the trip. it is sensible to design dedicated web interfaces for mobile devices, or even better to create respon- sive web designs that provide an optimal viewing experience for any device (from personal computers to mobile phones or tablets). a very interesting topic is to extend the current formats with videos, -dimensional objects or augmented reality, as some approaches already start to do. � another way of getting more information from the user and to improve the information that the recommender has on her interests is the exploitation of all the data provided by social networks and other web . applications, including the rela- tionships between users and the different kinds of content they provide (comments, pictures, ratings). this aspect is certainly very relevant in the tourism field, due to its highly social nature. thus, it is important to include in tourism recommenders as many possibilities of sharing information (pictures, videos, comments, ratings, localisation, etc.) as possible. the analysis of the social relationships of the users is a recent area of work that can surely lead towards the discovery of more accurate rec- ommendations that fit better with the user’s tastes, by taking into account the opinions of her closest friends, weighting the opinions depending on the strength of the relationship with the acquaintance, etc. � in any kind of recommender systems it is essential to have the precise information about the user’s interests stored in her pro- file. the particular characteristics of the tourism field offer the possibility to define new mechanisms to analyse the interaction of the user not only with the recommender system but also with the related environment, in order to learn automatically (and dynamically adapt) the user’s preferences. in particular, contextual recommendations are key in the success of any recommender of tourist activities. � it is also becoming increasingly clear that, in order to provide precise recommendations, it is necessary to move away from purely textual information and represent in a semantic way (e.g. through the use of ontologies) both the preferences of the user and the features of the different kinds of cultural and leisure activities. having this structured information, it is possi- ble to define and use complex semantic similarity techniques to compare users, compare objects or compare the preferences of the user with the characteristics of the objects. . . lines of future work this section briefly comments three relevant issues that are cur- rently being studied in the development of e-tourism recommend- ers: the diversification of the suggestions provided to the user, the use of social data available in current web . applications, and the improvement of the recommendations by leveraging the extra capabilities of mobile devices. � content-based systems focus on recommending items similar to the user’s profile, which may cause overspecialized results, leaving aside other items that might be interesting for the user. this is an important issue in some applications in the field of tourism. some recommender systems aim at making publicity of ‘‘different’’ or new sorts of activities which may be ignored by most visitors (e.g. a new restaurant or a new guided tour). it has also been argued that a smart recommender should pro- vide a diversified list of recommendations (e.g., even if the sys- tem knows that the user is interested in going to the beach, it is not very exciting to show a list of ten different beaches and not to suggest other kinds of related activities). in savir et al. ( ) a measure of balance between the number of attractions of a certain type and the minimum rating threshold is proposed in order to keep a fixed diversity level in the activities proposed in a trip. the sigtur recommender system borràs et al. ( ) also includes a diversification mechanism that aims to widen the range of suggested activities. in ruotsalo et al. ( ) the objects of a museum are gathered in clusters sharing the same features so that the recommendation procedure picks a repre- sentative number of objects from each cluster to increase the diversity of the proposal made to the visitor. � recent recommender systems exploit the power of new web- based applications, like social networks. in addition to offering social functionalities, these tools facilitate the use of collabora- tive filtering techniques, since this kind of technologies permit new forms of rating items or collecting user information at an individual level or at a social level. they are known as social rec- ommender systems (noel et al., ). these tools can be used j. borràs et al. / expert systems with applications ( ) – both to identify groups of items and to build groups of users. for example, in moretourism (rey-lópez et al., ) the users have an associated tag cloud with terms relevant to their profile, and a new tag is created for each attraction based on the tags of the users who liked it. this information is used to compare the tag clouds of users and items and find coincidences. tasticwiki obtains information about the user interactions with the items by analyzing the searches, readings and editions in a wiki (ruiz- montiel, molina-castro, & aldana-montes, ). this informa- tion is used to calculate the satisfaction degree that an article in the wiki has for a certain user. another example is found in speta (garcía-crespo et al., ), which maintains a social net- work profile of the user, so that the user’s contact data is taken into account in order to analyze the interactions between the users. trust is another component that appears when dealing with social recommenders. it has been argued that ratings of credible users should be treated with higher weights than oth- ers (gavalas & kenteris, ). � a special characteristic in tourism, which distinguishes it from other domains in which recommenders have been applied, is the mobility of the users, which may need recommendations in different moments and in different places. for this reason, this particular type of recommender systems has started to incorporate context-aware techniques. the success of this approach is due to the widespread use of mobile devices, as introduced in section . . many tourism recommenders run on phones, so the user’s location can be used to guide the filter- ing of the items to be shown (kurata, ; lamsfus et al., ; yang, ). not only the current location of the user is impor- tant, but also the places that have already been visited (gavalas & kenteris, ; umanets et al. ). other features that are considered as contextual information in tourism recommender systems are, for instance, the current weather to decide if it is more appropriate to recommend indoor or outdoor activities (braunhofer et al. ; garcía-crespo et al., ; gavalas & kenteris, ) or the motion speed and time to generate plans (noguera et al., ). in the system described in niaraki and kim ( ) a complex model of the context is considered for constructing personalised route plans. the context information is organised on a hierarchy, including aspects related to the traf- fic, weather, safety (like telephone booth, side road parking, medical centre, etc.), facilities (gas station, etc.) and tourist attractions (fishing zone, recreation place, seaside, etc.). in amato, mazzeo, moscato, and picariello ( b) four main parameters for the context are set: (i) time (time needed by the user to reach the place, the opening/closing times, etc.); (ii) location of the user and the place; (iii) weather and environ- mental conditions (e.g. temperature, humidity, rainfall degree, wind, season, moment of the day, etc.); (iv) social factors (num- ber of users close to the place and number of positive/negative feedbacks). moreover, the same authors extended the work (amato, mazzeo, moscato, & picariello, a) to indoor scenar- ios to analyze room crowd, room fitness, network performances, location and time interval. they use a pre-filtering strategy to select those alternatives that satisfy the user’s needs and a post-filtering strategy to arrange the recommended items based on their contextual values. this dimension is devised as a cru- cial point in the success of recommender systems in tourism, due to the inherent mobile behaviour of the users in this specific application domain. acknowledgements this work has been supported by the spanish research projects damask (data mining algorithms for semantic knowledge, tin - ), shade (semantic and hierarchical attributes in decision making, tin - ), and the collaboration agreement between the itaka research group and the science & technology park for tourism and leisure in the area of technological applications for tourism mobility management. references adomavicius, g., & tuzhilin, a. ( ). toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions. ieee transactions on knowledge and data engineering, , – . albanese, m., chianese, a., d’acierno, a., moscato, v., & picariello, a. ( ). a multimedia recommender integrating object features and user behavior. multimedia tools and applications, ( ), – . albanese, m., d’acierno, a., moscato, v., persia, f., & picariello, a. ( ). a multimedia recommender system. acm transactions on internet technology (toit), ( ), . alonso, k., zorrilla, m., confalonieri, r., vázquez-salceda, j., iñana, h., palau, m., calle, j., & castro, e. ( ). ontology-based tourism for all: recommender and information retrieval system for interactive community displays. in proceedings of eighth international conference on information science and digital content technology (pp. – ). amato, f., mazzeo, a., moscato, v., & picariello, a. ( b). exploiting cloud technologies and context information for recommending touristic paths. in intelligent distributed computing vii – proceedings of the seventh international symposium on intelligent distributed computing, idc , prague, czech republic (pp. – ). amato, f., mazzeo, a., moscato, v., & picariello, a. ( a). a recommendation system for browsing of multimedia collections in the internet of things. in internet of things and inter-cooperative computational technologies for collective intelligence (pp. – ). berlin, heidelberg: springer. batet, m., moreno, a., sánchez, d., isern, d., & valls, a. ( ). turist@: agent-based personalised recommendation of touristic activities. expert systems with applications, , – . berka, t., & plößnig, m. ( ). designing recommender systems for tourism. in proceedings of the th international conference on information technology in travel & tourism (enter’ ). borràs, j., de la flor, j., pérez, y., moreno, a., valls, a., isern, d., et al. ( ). sigtur/e- destination: a system for the management of complex tourist regions. in r. law, m. fuchs, & f. ricci (eds.), information and communication technologies in tourism (pp. – ). springer. borràs, j., moreno, a., valls, a., ferré, m., ciurana, e., salvat, j., russo, a., & anton- clavé, s. ( a). uso de técnicas de inteligencia artificial para hacer recomendaciones enoturísticas personalizadas en la provincia de tarragona. in ix congreso turismo y tecnologías de la información y las comunicaciones, málaga, spain (pp. – ). isbn: - - - - . borràs, j., valls, a., moreno, a., & isern, d. ( b). ontology-based management of uncertain preferences in user profiles. in s. greco, b. bouchon-menier, b. bouchon-menier, g. colletti, m. fedrizzi, b. matarazzo, & r. yager (eds.), information processing and management of uncertainty in knowledge-based systems part ii. communications in computer and information science (ccis) (vol. , pp. – ). springer-verlag. braunhofer, m., elahi, m., ricci, f., & schievenin, t. ( ). context-aware points of interest suggestion with dynamic weather data management. in information and communication technologies in tourism (pp. – ). springer international publishing. brewer, j. d. ( ). tourism and ethnic stereotypes: variations in a mexican town. annals of tourism research, ( ), – . buhalis, d. ( ). etourism: information technology for strategic tourism management. englewood cliffs: prentice hall. burke, r. ( ). hybrid recommender systems: survey and experiments. user modeling and user-adapted interaction, ( ), – . castillo, l., armengol, e., onaindía, e., sebastiá, l., gonzález-boticario, j., rodríguez, a., et al. ( ). samap: an user-oriented adaptive system for planning tourist visits. expert systems with applications, , – . ceccaroni, l., codina, v., palau, m., & pous, m. ( ). patac: urban, ubiquitous, personalized services for citizens and tourists. in third international conference on the digital society (icds ), february – , , cancun, méxico (pp. – ). codina, v., & ceccaroni, l. ( ). taking advantage of semantics in recommendation systems. in r. alquézar, a. moreno, & j. aguilar (eds.) (pp. – ). amsterdam, the netherlands, the netherlands: ios press. coelho, b., martins, c., & almeida, a. ( ). adaptive tourism modeling and socialization system. in cse ‘ proceedings of the international conference on computational science and engineering (vol. , pp. – ). dasarathy, b. ( ). nearest neighbour (nn) norms: nn pattern classification techniques. ieee computer society press. dey, a., & abowd, g. ( ). towards a better understanding of context and context-awareness. in huc ’ : proceedings of the first international symposium on handheld and ubiquitous computing. dujmovic, j., & nagashima, h. ( ). lsp method and its use for evaluation of java ides. international journal of approximate reasoning, , – . felfernig, a., gordea, s., jannach, d., teppan, e., & zanker, m. ( ). a short survey of recommendation technologies in travel and tourism. oegai journal, ( ), – . http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h j. borràs et al. / expert systems with applications ( ) – fenza, g., fischetti, e., furno, d., & loia, v. ( ). a hybrid context aware system for tourist guidance based on collaborative filtering. in proceedings of ieee int. conference on fuzzy systems (pp. – ). gao, m., liu, k., & wu, z. ( ). personalisation in web computing and informatics: theories, techniques, applications, and future research. information systems frontiers, , – . garcia, a., arbelaitz, o., linaza, m. t., vansteenwegen, p., & souffriau, w. ( ). personalized tourist route generation. in f. daniel, f. m. facca (eds.), th international conference on web engineering , icwe : vol. . lncs (pp. – ). garcia, i., sebastià, l., & onaindia, e. ( ). on the design of individual and group recommender systems for tourism. expert systems with applications, , – . garcia, a., torre, i., & linaza, m. t. ( ). mobile social travel recommender system. in information and communication technologies in tourism (pp. – ). springer international publishing. garcia, a., vansteenwegen, p., arbelaitz, o., souffriau, w., & linaza, m. t. ( ). integrating public transportation in personalised electronic tourist guides. computers & operations research, special issue: transport scheduling, , – . garcía-crespo, a., chamizo, j., rivera, i., mencke, m., colomo-palacios, r., & gómez- berbís, j. m. ( ). speta: social pervasive e-tourism advisor. telematics and informatics, , – . garcía-crespo, a., lópez-cuadrado, j. l., colomo-palacios, r., gonzález-carrasco, i., & ruiz-mezcua, b. ( ). sem-fit: a semantic based expert system to provide recommendations in the tourism domain. expert systems with applications, ( ), – . gavalas, d., & kenteris, m. ( ). a web-based pervasive recommendation system for mobile tourist guides. personal and ubiquituous computing, , – . gretzel, u. ( ). intelligent systems in tourism: a social science perspective. annals of tourism research, ( ), – . gyorodi, r., gyorodi, c., & dersidan, m. ( ). an extended recommendation system using data mining implemented for smart phones. international journal of computers and technology, ( ), – . hsu, f. m., lin, y. t., & ho, t. k. ( ). design and implementation of an intelligent recommendation system for tourist attractions: the integration of ebm model, bayesian networks and google maps. expert systems with applications, , – . huang, y., & bian, l. ( ). a bayesian network and analytic hierarchy process based personalized recommendations for tourist attractions over the internet. expert systems with applications, ( ), – . jannach, d., zanker, m., & jessenitschnig, m. ( ). developing knowledge-based travel advisor systems: a case study. in n. sharda (ed.), tourism informatics: visual travel recommender systems, social communities, and user interface design (pp. – ). hershey, pa: information science reference. kabassi, k. ( ). personalizing recommendations for tourists. telematics and informatics, ( ), – . koceski, s., & petrevska, b. ( ). empirical evidence of contribution to e-tourism by application of personalized tourism recommendation system. annals of the alexandru ioan cuza university – economics, ( ), – . kurata, y., & hara, t. ( ). ct-planner : toward a more user-friendly interactive day-tour planner. in information and communication technologies in tourism (pp. – ). springer international publishing. kurata, y. ( ). ct-planner : more flexible and interactive assistance for day tour planning. enter . in information and communication technologies in tourism (pp. – ). innsbruck, austria. lamsfus, c., alzua-sorzabal, a., martin, d., & smithers, t. ( ). an evaluation of a contextual approach to visitor information system. in r. law, m. fuchs, & f. ricci (eds.), proceeding of the enter conference (pp. – ). innsbruck, austria: springer-verlag (pp. – ). lamsfus, c., alzua-sorzabal, a., martin, d., salvador, z., & usandizaga, a. ( ). human-centric ontology-based context modelling in tourism. in proceedings of the international conference on knowledge engineering and ontology development, funchal, madeira, portugal, october – , (pp. – ). lee, c. s., chang, y. c., & wang, m. h. ( ). ontological recommendation multi- agent for tainan city travel. expert systems with applications, , – . linaza, m. t., aguirregoikoa, a., garcía, a., torres, j. i., & aranburu, k. ( ). image- based travel recommender system for small tourist destinations. in information and communication technologies in tourism (pp. – ). innsbruck, austria. lorenzi, f., loh, s., & abel, m. ( ). personaltour: a recommender system for travel packages. in proceedings of the ieee/wic/acm international conference on intelligent agent technology, iat , lyon, france, august – . ieee/wic/acm international conferences on web intelligence and intelligent agent technology. luberg, a., tammet, t., & järv, p. ( ). smart city: a rule-based tourist recommendation system. in r. law, m. fuchs, & f. ricci (eds.), information and communication technologies in tourism . enter , january th– th , innsbruck, austria. springer-verlag. lucas, j., laurent, a., moreno, m., & teisseire, m. ( ). fuzz-cba: classification à base de règles d’association floues et systèmes de recommandation. rencontres francophones sur la logique floue et ses applications, – . lucas, j., luz, n., moreno, m., anacleto, r., figueiredo, a., & martins, c. ( ). a hybrid recommendation approach for a tourism system. expert systems with applications, ( ), – . manouselis, n., & costopoulou, c. ( ). analysis and classification of multi- criteria recommender systems. world wide web, ( ), – . marques, s. s. ( ). fabrication of image-destination: the impact of stereotypes at the level of tourism. revista turismo & desenvolvimento, . martin, d., alzua, a., & lamsfus, c. ( ). a contextual geofencing mobile tourism service. in r. law, m. fuchs, & f. ricci (eds.), proceeding of the enter conference (pp. – ). innsbruck, austria: springer-verlag. martínez, l., rodríguez, r. m., & espinilla, m. ( ). reja: a georeferenced hybrid recommender system for restaurants. in proceedings of ieee/wic/acm international conference on web intelligence and intelligent agent technology (pp. – ). martínez-santiago, f., ariza-lópez, f., montejo-ráez, a., & ureña-lópez, a. ( ). geoasis: a knowledge-based geo-referenced tourist assistant. expert systems with applications, ( ), – . meehan, k., lunney, t., curran, k., & mccaughey, a. ( ). context-aware intelligent recommendation system for tourism. in proceedings of the th ieee international conference on pervasive computing and communications (pp. – ). mínguez, i., berrueta, d., & polo, l. ( ). cruzar: an application of semantic matchmaking to e-tourism. in y. kalfoglou (ed.), cases on semantic interoperability for information systems integration: practices and applications (pp. – ). hershey, pa: information science reference. montaner, m., lópez, b., & de la rosa, j. l. ( ). a taxonomy of recommender agents on the internet. artificial intelligence review, ( ), – . montejo-ráez, a., perea-ortega, j. m., garcía-cumbreras, m. a., & martínez- santiago, f. ( ). otium: a web based planner for tourism and leisure. expert systems with applications, , – . moreno, a., valls, a., isern, d., marin, l., & borràs, j. ( ). sigtur/e-destination: ontology-based personalized recommendation of tourism and leisure activities. engineering applications of artificial intelligence, , – . moscato, v., picariello, a., & rinaldi, a. m. ( ). towards a user based recommendation strategy for digital ecosystems. knowledge-based systems, , – . niaraki, a. s., & kim, k. ( ). ontology based personalized route planning system using a multi-criteria decision making approach. expert systems with applications, , – . noel, j., sanner, s., tan, k. -n., christen, p., xie, l., bonilla, e. v., abbasnejad, e., & della penna, n. ( ). new objective functions for social collaborative filtering. in proceedings of the st int. conf. on world wide web (pp. – ). noguera, j. m., barranco, m. j., segura, r. j., & martínez, l. ( ). a mobile d-gis hybrid recommender system for tourism. information sciences, , – . pearl, j. ( ). probabilistic reasoning in intelligent systems. networks of plausible inference. morgan kaufmann. pinho, j., da silva, b., moreno, m., de almeida, a. m., & lopes, c. ( ). applying recommender methodologies in tourism sector. highlights in practical applications of agents and multiagent systems. in advances in intelligent and soft computing (pp. – ). springer. rey-lópez, m., barragáns-martínez, a. b., peleteiro, a., mikic-fonte, f. a., & burguillo, j. c. ( ). moretourism: mobile recommendations for tourism. in ieee int. conf. on consumer electronics (pp. – ). ricci, f. ( ). travel recommender systems. ieee intelligent systems, ( ), – . ricci, f., nguyen, q. n., & averjanova, o. ( ). exploiting a map-based interface in conversational recommender systems for mobile travelers. in n. sharda (ed.), tourism informatics: visual travel recommender systems, social communities, and user interface design (pp. – ). hershey, pa: information science reference. rojas, g., & uribe, c. ( ). a conceptual framework to develop mobile recommender systems of points of interest. in international workshop on advanced software engineering. proceedings jornadas chilenas de computación, chile, november . ruíz-martínez, j. m., miñarro, j. a., castellanos, d., garcía, f., & valencia, r. ( ). ontology population: an application for the e-tourism domain. international journal of innovative computing, information and control, ( ), – . ruiz-montiel, m., & aldana-montes, j. ( ). semantically enhanced recommender systems. on the move to meaningful internet systems: otm workshops (vol. , pp. – ). springer. ruiz-montiel, m., molina-castro, j. j., & aldana-montes, j. f. ( ). tasticwiki: a semantic wiki with content recommendation. in proceedings of the fifth workshop on semantic wikis-linking data and people. seventh extended semantic web conference (pp. – ). ruotsalo, t., haav, k., stoyanov, a., rochee, s., fanid, e., deliaic, r., et al. ( ). smartmuseum: a mobile recommender system for the web of data. web semantics: science, services and agents on the world wide web, , – . sánchez, d., & moreno, a. ( a). pattern-based automatic taxonomy learning from the web. ai communications, ( ), – . sánchez, d., & moreno, a. ( b). learning non-taxonomic relationships from web documents for domain ontology construction. data and knowledge engineering journal, , – . sánchez, d., moreno, a., & del vasto-terrientes, l. ( ). learning relation axioms from text: an automatic web-based approach. expert systems with applications, , – . savir, a., brafman, r., & shani, g. ( ). recommending improved configurations for complex objects with an application in travel planning. in proceedings of the seventh acm conference on recommender systems (pp. – ). acm. sebastià, l., garcia, i., onaindia, e., & guzman, c. ( ). e-tourism: a tourist recommendation and planning application. international journal on artificial intelligence tools, ( ), – . http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h j. borràs et al. / expert systems with applications ( ) – sebastià, l., giret, a., & garcia, i. ( ). a multi agent architecture for tourism recommendation. trends in practical applications of agents and multiagent systems. advances in intelligent and soft computing, , – . seidel, i., gärtner, m., pöttler, m., berger, h., & dittenbach, m. ( ). itchy feet: a d e-tourism environment. in n. sharda (ed.), tourism informatics: visual travel recommender systems, social communities, and user interface design. hershey, pa: igi global. sieg, a., mobasher, b., & burke, r. ( ). learning ontology-based user profiles: a semantic approach to personalized web search. ieee intelligent informatics bulletin, ( ), – . souffriau, w., vansteenwegen, p., vanden berghe, g., & van oudheusden, d. ( ). the planning of cycle trips in the province of east flanders. omega, , – . staab, s., & werthner, h. ( ). intelligent systems for tourism. ieee intelligent systems, ( ), – . tsung-chiung, w., chyong-ru, l., & wan-chen, y. ( ). segmenting indigenous tourists from a serious leisure perspective. journal of vacation marketing, ( ), – . umanets, a., ferreira, a., & leite, n. ( ). guideme – a tourist guide with a recommender system and social interaction. in conference on electronics, telecommunications and computer (cetc), lisboa, december . valls, a. ( ).clusdm: a multiple criteria decision method for heterogeneous data sets (ph.d. thesis). ai communications (pp. – ). vansteenwegen, p., & souffriau, w. ( ). trip planning functionalities: state-of- the-art and future. information technology and tourism, ( ), – . vansteenwegen, p., souffriau, w., vanden berghe, g., & van oudheusden, d. ( ). the city trip planner: an expert system for tourists. expert systems with applications, ( ), – . venkataiaha, s., shardaa, n., & ponnadaa, m. ( ). a comparative study of continuous and discrete visualisation of tourism information. enter conference, international federation for information technology and travel & tourism (ifitt), innsbruck, austria, – january . vicient, c., sánchez, d., & moreno, a. ( ). an automatic approach for ontology- based feature extraction from heterogeneous textual resources. engineering applications of artificial intelligence, ( ), – . wang, w., zeng, g., & tang, d. ( ). bayesian intelligent semantic mashup for tourism. concurrency and computation: practice and experience, , – . werthner, h. ( ). intelligent systems in travel and tourism. in ijcai (vol. , pp. – ). wooldridge, m. ( ). an introduction to multiagent systems. john wiley and sons. yager, r. ( ). on ordered weighted averaging operators. operators in multi- criteria decision making. ieee transactions on systems, man and cybernetics, , – . yang, w. s., & hwang, s. y. ( ). itravel: a recommender system in mobile peer- to-peer environment. the journal of systems and software, ( ), – . yang, y. ( ). semantic user model inferences for travel recommender systems. in n. sharda (ed.), tourism informatics: visual travel recommender systems, social communities, and user interface design (pp. – ). hershey, pa: information science reference. yu, c. & chang, h. ( ). personalized location-based recommendation services for tour planning in mobile tourism applications. in proceedings of the th international conference on e-commerce and web technologies, linz, austria, september – : vol. . lecture notes in computer science (pp. – ). http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h intelligent tourism recommender systems: a survey introduction interface . web-based recommenders . mobile recommendations functionalities . travel destination and tourist packs . ranked list of suggested attractions . planning a route . social aspects recommendation techniques in e-tourism . approaches to recommendation . representation of the user profile use of ai techniques in tourism recommender systems . multi-agent systems . optimization techniques . automatic clustering . management of uncertainty . knowledge representation related reviews conclusions . summary and guidelines to develop tourism recommenders . lines of future work acknowledgements references real-time credit card fraud detection using computational intelligence real-time credit card fraud detection using computational intelligence jon t.s. quah *, m. sriganesh school of electrical and electronic engineering, nanyang technological university, nanyang avenue, singapore , republic of singapore abstract online banking and e-commerce have been experiencing rapid growth over the past few years and show tremendous promise of growth even in the future. this has made it easier for fraudsters to indulge in new and abstruse ways of committing credit card fraud over the internet. this paper focuses on real-time fraud detection and presents a new and innovative approach in understanding spending patterns to decipher potential fraud cases. it makes use of self-organization map to decipher, filter and analyze customer behavior for detection of fraud. � elsevier ltd. all rights reserved. keywords: self-organizing map; unsupervised learning; risk scoring; transaction . introduction the fast and wide reach of the internet has made it one of the major selling channels for the retail sector. in the last few years, there has been a rapid increase in the number of card issuers, card users and online merchants, giving very little time for technology to catch-up and prevent online fraud completely. statistics shows that on-line banking has been the fastest growing internet activity with nearly % of the population in the us actively participating in it (fighting fraud on the internet, ). as overall e-com- merce volumes continued to grow over the past few years, the figure of losses to internet merchants was projected to be between $ and $ billion in the year . recent sta- tistics by garner group place online fraud rate between . % and . %, with auction fraud accounting to nearly half of the total incidents of fraud on the internet (online banking, ). considering the current trends of e-com- merce volumes, the projected loss is $ . billion in the year , with $ . billion in the us alone (statistics for gen- eral & online fraud, ). . . how and where does fraud begin in order to understand the severity of credit card fraud, let us briefly look into the mechanisms adopted by fraud- sters to commit fraud. credit card fraud involves illegal use of card or card information without the knowledge of the owner and hence is an act of criminal deception. fraudsters usually get hold of card information in a variety of ways: intercepting of mails containing newly issued cards, copying and replicating of card information through skimmers or gathering sensitive information through phish- ing (cloned websites) or from unethical employees of credit card companies. phishing involves acquiring of sensitive information like card numbers and passwords by masquerading as a trust- worthy person or business in an electronic communication such as e-mail (schneck, ). fraudsters may also resort to generation of credit card numbers using bin (bank iden- tification numbers) of banks. a recent scheme of triangula- tion takes fraud fighters many days to realize and investigate (bhatla, prabhu, & dua, ). in this method, the fraudster operates through an authentic-looking website, where he advertises and sells goods at highly discounted prices. the unaware buyer submits his card information and buys goods. the fraudster then places an order with a genuine merchant using the stolen card information. he then uses - /$ - see front matter � elsevier ltd. all rights reserved. doi: . /j.eswa. . . * corresponding author. tel.: + ; fax: + . e-mail address: itsquah@ntu.edu.sg (j.t.s. quah). www.elsevier.com/locate/eswa available online at www.sciencedirect.com expert systems with applications ( ) – expert systems with applications mailto:itsquah@ntu.edu.sg the stolen card to purchase other goods or route funds into intractable accounts. its only after several days that the mer- chant and card owners realize about the fraud. this type of fraud causes initial confusion that provides camouflage for the fraudster to carry out their operations. . . impact of fraud it is interesting to note that credit card fraud affects card owners the least because their liability is limited to the trans- actions made. the existing legislations and cardholder pro- tection policies as well as insurance schemes in most countries protect the interests of the cardholders. however, the most affected are the merchants, who, in most situations, do not have any evidence (eg. digital signature) to dispute the cardholders’ claim of misused card information. mer- chants end up bearing all the loses due to chargeback, ship- ping cost of goods, card issuer fees and charges as well as their own administrative costs. excessive fraudulent cases involving the same merchant can drive away customers, cause card issuer banks to withdraw service and also result in loss of reputation and goodwill (yu & singh, ). card issuer banks have to bear the administrative cost of investi- gations into fraud cases as well as infrastructure costs of set- ting up the required software and hardware facilities to combat fraud. they also incur indirect costs through trans- action delays. studies show that the average time lag between the fraudulent transaction date and chargeback notification can be as high as days, thereby giving fraudsters sufficient time to cause severe damage (bhatla et al., ). . . fraud detection and prevention the negative impacts of fraud make it very clear and necessary to put in place an effective and economical fraud detection system. recent technological advancements to combat fraud have contributed number of solutions in this area (bhatla et al., ; all points protection, ). fraud detection techniques involving sophisticated screen- ing of transactions to tracking customer behaviour and spending patterns are now being developed and employed by both merchants as well as card issuer banks (tan & thoen, ). some of the recently employed techniques include transaction screening through address verification systems (avs), card verification method (cvm), per- sonal identification number (pin) and biometrics. avs involves verification of address with zip code of the cus- tomer while cvm and pin involve checking of numeric code that is keyed in by the customer. biometrics might involve signature or fingerprint verification. rule-based methods and maintaining of positive and negative lists of customers and geographical regions are also used in prac- tice. data mining and credit scoring methods focus on sta- tistical analyses and deciphering of customer behaviour and spending patterns to detect frauds (huang, chen, & wang, ). neural networks are capable of deriving pat- terns out of databases containing historical transactions of customers. these neural networks can be ‘trained’ and are ‘adaptive’ to the emerging new forms of frauds. deployment of sophisticated techniques and screening of every transaction alone will not reduce losses. it is necessary to employ an effective and economical solution to combat fraud (turney, ). such a solution should not only detect fraud cases efficiently but also turn out to be cost-effective. the idea is to strike a balance between the cost involved in transaction screening and review and the losses due to fraud- ulent cases. analyses show that review of only . % of trans- actions can result in reducing fraud losses accounting to . % of total value of transactions. while a review of as high as % of transactions can reduce the fraud loses drastically to . %, but that increases review costs exorbitantly (bhatla et al., ). the estimated cost of not using anti-fraud soft- ware was about $ billion in (statistics for general & online fraud, ). the key to minimize total costs is to categorize transac- tions and review only the potentially fraudulent cases. this should involve deployment of a step-by-step screening, filter- ing and review mechanism. a typical deployment can involve initial authentication of transactions through pin, expiry date on card, avs and cvm. a second level of screening can involve comparing with positive and negative lists as well as rules based on customers, geographical regions, ip addresses and policies. risk and credit scoring with pattern and behaviour analyses can come next, followed by manual review. this classifies and filters out transactions as genuine or fraudulent in every step and as a result only a few transac- tions would require further manual review. such a solution reduces the overall processing delay as well as total costs involved in manpower and administration. the focus of this paper will now shift to risk scoring and behavioral pattern detection using neural networks. . neural networks in fraud detection – literature review neural networks have been extensively put to use in the areas of banking, finance and insurance. they have been successfully applied into credit scoring of customers, bank- ruptcy or business failure prediction, stock price forecast- ing, bond rating, currency prediction and many more areas (stock analysis, ; quah & srinivasan, ). in the area of fraud detection and prevention, neural net- works like feed-forward networks with back-propagation have found immense applications (fraud brief – avs & cvm, clearcommerce corporation, ; the evolving threat of card skimming, fairisaac, ; clearcom- merce fraud prevention guide, ). usually such appli- cations of neural networks systems involve knowing about the previous cases of fraud, to make systems learn the var- ious trends. fraud cases are statistically analyzed to derive out relationships among input data and values for certain key parameters in order to understand the various patterns of fraud. this knowledge of fraud trends is then iteratively taught to feed-forward neural networks, which can success- fully identify similar fraud cases occurring in the future. j.t.s. quah, m. sriganesh / expert systems with applications ( ) – https://isiarticles.com/article/ pii: - ( ) - socrates: a flexible toolkit for building logic-based expert systems r corlett*, n davies*, r khan*, h reichgeltt and f van harmelent this paper describes the architecture o f the socrates tool- kit ./'or building expert systems. the authors analyse the problems associated with existing expert system tools and propose a solution based on the use o f logic and meta-level inference. the abstract architecture f o r the toolkit is described which embodies this combination o f logic and meta-level inference. this architecture can be instantiated to create a system that is specialized for a particular application. this specialization process can be seen as a methodology f o r building expert systems. the three stages o f this methodology are discussed in detail, along with descriptions o f how the socrates toolkit supports it. the current implementation o f socrates, plus a number o f applications o f the toolkit are described, and the open problems are discussed. keywords: expert systems, toolkit, architecture, building methodology most o f the tools that are currently available for constructing expert systems fall into two categories, namely expert system shells and high level programming language environments. both o f these types o f tool suffer from a number o f shortcomings which limit their useful- ness. the first type o f available tools, 'shells', are usually constructed by abstraction from an existing expert system. thus, a shell normally consists o f an inference engine and an empty knowledge base, and some de- bugging and explanation facilities. buyers o f a shell often believe, and manufacturers often claim, that the shell is appropriate for a number o f different applications. however, a large number o f people have expressed dis- satisfaction with expert system shells. the two most frequently heard complaints (e.g. a l v e y ' ) are first, that the view that the inference engine which was successful *gec research, marconi research centre, chelmsford, uk tdepartment of artificial intelligence, university of edinburgh, edinburgh eh yl, uk - / / - $ . © in one application will also be successful in other applications is unwarranted, and second, that the knowledge representation scheme often makes the expression o f knowledge in another d o m a i n awkward, if not impossible. f o r example, a production rule language that was designed for solving a classification problem using backward chaining would probably not be very suitable for solving a design or planning problem. on the other hand, proponents o f high level program- ming language environments (sometimes known as 'hybrid systems'), such as loops , k e e or a r t , can be seen as taking a more pluralistic position: instead o f providing knowledge engineers with a single pre- fabricated inference engine, one provides them with a large set o f tools, each o f which has proven useful in other applications. loops, for example, provides object-oriented programming with inheritance, a pro- duction rule interpreter and active values, as well as lisp. while we accept that hybrid systems are useful as tools for program development, we would claim that they are less useful as tools for building expert systems. their main problem is that they provide the knowledge engineer with a bewildering array o f possibilities, and little, if any, guidance as to the circumstances in which any o f these possibilities should be used. unless used by experienced programmers, high level programming environments encourage an ad hoc programming style in which no attention is paid to a principled analysis o f the problem at hand to see which strategy is best suited for its solution. the conclusion that is drawn from the problems associated with shells and high level programming language environments is that a number o f different "models o f rationality' are needed, and that different applications require different models o f rationality. when constructing an expert system, the knowledge engineer then has to decide which model o f rationality is appropriate to the application at hand. a model o f rationality has both a 'static' and a 'dynamic' aspect. these two aspects correspond to the knowledge about butterworth & co (publishers) ltd knowledge-based systems the application area plus a strategy that describes how to use this knowledge when solving a problem. the 'interpretation models' from breuker and wielinga s correspond closely to our models o f rationality. the distinction between the static and the dynamic aspects of a model of rationality corresponds to a dis- tinction one can make between two different aspects of an expert system. first, there is the 'domain' in which the expert system is to solve problems. for example, the domain of an expert system may be electronics, or internal medicine. secondly, there is the 'task' which the knowledge engineer wants the expert system to perform. for example, the task of a system can be diagnosing a faulty electronic circuit or designing a new circuit. it is interesting to note that the problems with shells reflect these two aspects of an expert system. the first complaint about shells concerning the expressive- ness of the knowledge representation language is related to the structure of the domain. the second complaint concerning the rigidity of the inference engine is related to the task of the expert system. as pointed out by chandrasekaran ~, typical expert system tasks such as diagnosis, planning, monitoring, etc. seem to require particular control regimes. socrates allows the use of a variety of logical languages and a variety of control regimes to solve the problem o f the lack of flexibility associated with shells. by using logic as its unifying framework, and by provid- ing guidelines for the choice of both the representation language and the control regime, socrates avoids the unstructured richness of the hybrid systems. these points are discussed in detail below. u s i n g l o g i c f o r k n o w l e d g e r e p r e s e n t a t i o n socrates uses a 'logical language' as the main formalism to implement the static aspects of the required models of rationality for different domains. on the one hand, logical formalisms are rich enough to provide different models of rationality, while on the other hand the use of logic provides a unifying framework for the system which saves it from the unstructured richness of the hybrid systems. this choice o f logic as the main formalism implies that logical languages will serve as the representational scheme, while logical deduction will be the paradigm for the inference engine. many advantages come with the use of logic as the main knowledge representation formalism. logic comes with a formal semantics, it has well understood proper- ties regarding completeness, soundness, decidability, etc., and it has great expressive power. for a further elaboration on these arguments, see corlett, davies, khan, reichgelt and van harmelen . u s i n g m e t a - l e v e l i n f e r e n c e f o r c o n t r o l the correspondence between the two aspects of a model of rationality and the problems with shells suggests that a model of rationality should be computationally realized as a knowledge representation formalism plus a control regime for using this formalism. a number of arguments can be given for the explicit and separate representation of control knowledge. first, a system with i - . . . . . . . . . . . . u ~ r - f ' r o n ' t end . . . . . . . . . . . . . . "i l . . . . . . . . - i / a - , , [ - _ - - - - - - _- ; _ . . . . . . . . . . . . ] figure . socrates architecture an explicit representation of its control strategies is easier to develop, debug and modify, as argued by davis s, bundy and welham , clancey ~° and aiello and levi . second, the separation of control knowledge from domain knowledge (or 'object-level knowledge') increases the reusability of the system: the same object- level knowledge can be used for different purposes, and the same control knowledge can be used in different domains s. ~ o. finally clancey ~ stresses the importance of explicit control knowledge for the purpose of explanation. the separation of control knowledge from domain knowledge allows domain knowledge to be purely declarative in nature. while formulating domain knowledge we do not have to worry about efficiency, only about the 'representational adequacy' (in the words of mccarthy). in the control knowledge, on the other hand, the efficiency of the problem solving process is the most prominent aspect ('computational adequacy'). a b s t r a c t a r c h i t e c t u r e o f s o c r a t e s as argued in many other places in the literature (see ~ .~ .~ - among others), an architecture with a separate object-level and meta-level interpreter can be used to implement the required separation of domain and control knowledge. this architecture is shown in figure . the two-layer architecture of object-level and recta-level interpreter is extended in socrates with a third level, the scheduler, discussed below. each of the three interpreters communicates with the knowledge base in order to store and retrieve logical propositions. as described below, the knowledge base can be organized into a number of partitions, each of which can be hierarchically organized into subpartitions. the socrates architecture distinguishes between front- ends to be used by the 'end-user' and the 'knowledge engineer'*. knowledge engineers and end-users will need different tools for communication with the system. for example, an end-user, when asking for an explanation, will not want to see the entire proof tree, but rather 'edited highlights'. a knowledge engineer, on the other hand, may want to be able to monitor the reasoning *the end-user is the person who uses the expert system application built with the socrates toolkit, whereas the knowledge engineer is the person building such an application system. vol no june process much more closely and m a y require the full p r o o f tree. similarly, a knowledge engineer will need tools for adjusting the b e h a v i o u r o f the interpreters and for edit- ing the knowledge base, whereas an end-user only needs to browse in a read-only m a n n e r t h r o u g h the knowledge base. f u r t h e r m o r e , the levels o f abstraction at which the system communicates will differ between end-user and knowledge engineer. this general, application independent architecture needs to be "configured' for a particular expert systems application, given the characteristics o f the d o m a i n and the task o f that application. the use o f logic as the underlying framework for the system gives a very specific meaning to the process o f configuring the system for a particular application. r a t h e r than having to choose between a r b i t r a r y representational schemes and inference m e t h o d s (as is the case in the hybrid systems), the configuration process now consists o f three well defined stages; i.e. the declaration o f a 'logical language', the declaration o f a ' p r o o f theory" for that logical language, and the declaration o f a ' p r o o f strategy'. l o g i c a l r e p r e s e n t a t i o n l a n g u a g e d e c l a r a t i o n as the first step in the process o f configuring a system for a particular application, the knowledge engineer defines the logical language that will be used for representing the d o m a i n knowledge. in the context o f a knowledge-based system, where the logical language is to be used for representational purposes, the 'syntactic" declaration o f the language has to be augmented with a mechanism for 'storage and retrieval'. each o f these aspects will be discussed in this section. r e p r e s e n t a t i o n l a n g u a g e as the first part o f the declaration o f the logical represen- tation language, the knowledge engineer has to declare the "vocabulary' o f the language. t h e predicate symbols, function symbols and constants that are going to be part o f the representation language have to be defined. socrates uses ' m a n y - s o r t e d logics' o f the kind pro- posed by waither ~, making use o f typed quantification, where the sorts are organized in a hierarchy. each sort is used to represent a n o n - e m p t y set o f individuals belonging to that sort. in such logics the sort hierarchy is defined by a partial ordering on the set o f sorts, corres- ponding to a lattice structure with the universal sort as the maximal, and the e m p t y sort* as the minimal element in the set. this lattice is m o r e general than a tree structure, since it allows sorts to have more than one supersort, thereby increasing the expressiveness o f the sort o f system. t h e use o f m a n y - s o r t e d logics over unsorted logics brings a n u m b e r o f advantages. first, complicated unsorted expressions can be reformulated in a much simpler m a n y - s o r t e d form. f o r example, the unsorted expression: v x y [human(x) ~ h u m a n ( y ) & mother(y, x)] * t h e e m p t y s o r t , n o t a t i o n ~,~, is t h e o n l y s o r t r e p r e s e n t i n g a n e m p t y set o f individuals, a n d a s s u c h d o e s n o t fulfill a n y r e p r e s e n t a t i o n a l role. t h e r e a s o n w h y s o r t s m u s t c o r r e s p o n d t o n o n - e m p t y sets is d i s c u s s e d below. t t h e n o t a t i o n x . t is used to indicate t h a t v a r i a b l e x is o f s o r t t. can be rewritten, after the declaration o f human as a sort, as v x: h u m a n y: h u m a n [mother(y, x)]'t this reduction in complexity o f axioms not only improves the readability o f the knowledge base, but it also causes a significant decrease in the size o f the search space for proofs. c o h n -" (in the context o f a resolution based theorem prover), and davies (in the context o f a natural d ed u ct i o n system) show that this reduction can a m o u n t to as much as an o r d e r o f magnitude. second, the sort hierarchy allows the knowledge engineer to represent what clancey calls "structural knowledge', representing the taxonomical hierarchy o f the application domain. in sorted logics, the unification algorithm has to take the sort hierarchy into account. th e rules for unification in a sorted logic are: a constant c,:t~ unifies with a co n st an t c :/ if c~ = c and ti = i a variable x.'t~ unifies with a co n st an t c.'t if t = tt a variable x.'q unifies with a variable y:t if tj th e third rule results in restricting the sorts o f both variables to t = t~ c~ t . in o rd er to g u aran t ee that t can be c o m p u t e d and is itself a legal sort, it is necessary that the system knows the values o f all pairwise intersec- tions o f all sorts, and that sort-intersection is closed o v er sorts. in o t h e r words, if t is the set o f all sorts (including ~ ) , then v t~. t=:tt • t & l • t - , t~ c~ t • t f u r t h e r m o r e , in o r d e r to g u aran t ee that is uniquely defined, we require that no two sorts are used to repre- sent the same set o f individuals. i f this were allowed, the unification algorithm would no longer be able to return a unique most general unifier. th e knowledge engineer declares the sorts o f the logic by declaration first the set o f sorts, and then the hierarchy o f sorts using set-theoretic primitives, such as equality, subset, superset, intersection, union, set- difference, c o m p l e m e n t a t i o n an d disjointness. after this has been done, the system automatically checks whether the sort hierarchy satisfies the following criteria: • no two sorts are equal. • n o sort except .~,, is empty. • intersection is closed o v er sorts. if an y o f these constraints is violated, or if the constraints are not satisfied by the current declarations (because the knowledge engineer has underspecified the sort hierarchy), the system asks the knowledge engineer for different o r additional declarations. t h e sorts o f the constants o f the logical language can be declared in two ways. t h e simplest way is by simply en u m erat i n g the constants o f a particular sort ('extensional' definition o f types). h o w ev er, for some sorts e n u m e r a t i o n is not feasible, either because the set o f constants o f that sort is not k n o w n in advance, or because this set is t o o large, o r indeed infinite. f o r these cases it is possible to define constants by declaring a "recognition procedure" for the sort. every object for which the recognition p r o c e d u r e succeeds will be assumed to be a co n st an t o f t h at particular sort ('inten- sionai' definition o f types). t h e definitions o f the knowledge-based systems recognition procedures must be supplied in the imple- mentation language of the system. this allows an inter- face between the logical representation language and the computational data types supplied by the implemen- tation language o f the system. an example of a sort that can be declared via this mechanism is the sort of all integers, or as a second example, for any given sort t, the sort t l i s t consisting of all lists of elements of sort t. as part of declaring the vocabulary of the logical representation language, it is possible to exploit what weyhrauch calls 'semantic attachment'. this is done by declaring a special class of predicates called 'evaluable predicates'. when one of these predicates is encountered in a proof, it is possible (depending on the control decisions made by the meta-level interpreter, discussed below) to execute a procedure defined for this predicate to determine its truth value, and possibly provide bindings for any variables. these predicates provide an interface between the logical representation language and the computational environment of the system, enabling external systems interaction, input/output for interacting with the end-user, and the access of the facili- ties provided by the implementation language of the system. at this stage the representation language consists only of a sort hierarchy and a set of predicates, constants and functions. we still need to extend our language with 'logical connectives', such as implication, dis- junction, etc., and possibly non-standard operators*, such as modal and temporal operators. as part of the declaration of these logical connectives the knowledge engineer can declare their properties with respect to commutativity and associativity. these declarations will be used to configure a unifier for the defined logical language (see the section below on the declaration of the proof theory for a more detailed discussion of this subject). socrates distinguishes between different kinds of logical connectives: 'unary' connectives take one argument; binary connectives take two arguments, and can be divided on the basis o f their commutativity. implication ( ~ ) is an example of a 'non-commutative binary' connective, whereas equivalence ( --,) is an example of a 'commutative binary' connective. if con- nectives are both commutative and associative (such as, for instance, conjunction), they are treated as so-called 'set-connectives'. this amounts to them taking any number o f arguments in any order. furthermore, set connectives are assumed to be idempotent: all multiple occurrences of an argument can be reduced to only one occurrence. other possible connectives are not currently implemented in socrates: 'bag' operators, which are associative and commutative but not idempotent (i.e. the order of arguments does not matter, but multiple occurrences o f an argument cannot be reduced), and 'sequence' operators, where the order of the arguments is significant, and multiple occurrences cannot be reduced. s t o r a g e and retrieval mechanism now that the vocabulary of the logical language is complete we need to declare a storage and retrieval *the terms logical connective and logical operator are used synonymously mechanism for expressions in the language. this is what is typically called the 'knowledge base' of the system. the knowledge base of socrates is not a flat space of assertions in the declared logical language, but can be divided into a number of 'partitions'. each o f these partitions can have a separate logical language asso- ciated with it (but see below for a restriction on this). this facility serves a number of different purposes. first, it allows the knowledge engineer to use mixed language representations of the domain. different types of knowledge or different aspects of the domain can be represented in separate languages. a particular example of this is described below, where we discuss the control knowledge of the system. this meta-level knowledge is expressed in a different language from the object-level knowledge. nevertheless, it can be stored in the same knowledge base, using the partitioning mechanism to separate the two. second, the partition hierarchy can be used to reduce the search that needs to be done both by the retrieval mechanism and by the inference machinery. in many problem solving situations, only a subset of all the avail- able knowledge is applicable at any one time to a given problem. partitioning allows useful subsets to be applied while others are ignored. in this way, socrates can be used to model a blackboard architecture, with each of the partitions simulating the contents of a knowledge s o u r c e . an important aspect of the partitions in the knowledge base is that they can be recursively divided into sub- partitions, and that these subpartitions are organized hierarchically. partitions lower down in this hierarchy inherit all the propositions from partitions higher up in the hierarchy, but not vice versa. a particular example of this could be the use of a single working memory for a number of different subsets of domain knowledge, where the working memory would be represented in the top node, and the subsets of domain knowledge repre- sented in subpartitions. the results o f reasoning done within one subpartition can in this way be communicated to the other partitions via the working memory. again, this feature can also be used in the modelling of a black- board system in socrates, since the blackboard needs to be visible from all knowledge sources, but not vice versa. since partitions can be created dynamically at runtime, another application for the partition hierarchy in the knowledge base is hypothetical reasoning, some- times known as 'what-if' reasoning. each time a new hypothesis is generated, a new subpartition can be created containing this hypothesis, and inheriting all existing propositions from higher partitions. the inheritance mechanism puts restraints on the logical languages that can be used within subpartitions. because subpartitions inherit propositions from super- partitions, all the partitions in a partition hierarchy must be associated with the same logical representation language. (strictly speaking, it is only necessary that the language of a subpartition is an extension of the language of its superpartition, but socrates does not support this type of language inheritance.) thus, the knowledge base can be seen as a set of trees of partitions. within each tree, all partitions must have the same language, but each separate tree can be associated with a different language. a final facility supported by the knowledge base is voi n o june the a n n o t a t i o n o f p r o p o s i t i o n s . each p r o p o s i t i o n can be a n n o t a t e d with a r b i t r a r y i n f o r m a t i o n using a slot- value m e c h a n i s m . this allows us to associate extra- logical i n f o r m a t i o n with propositions. this i n f o r m a t i o n can be used for a wide range o f purposes, o f which a few e x a m p l e s are: n a t u r a l language descriptions to be used in explanations, c o n t r o l i n f o r m a t i o n to be used by the meta-level interpreter, c e r t a i n t y values, etc. this slot-value m e c h a n i s m could be used to model a truth- m a i n t e n a n c e system in socrates, by storing each derived p r o p o s i t i o n in the k n o w l e d g e base t o g e t h e r with a n n o t a - tions that c o n t a i n the premises t h a t were used in its derivation. socrates uses a discrimination net technique for retrieval o f propositions. t h e o p t i m a l criteria that the net should use for discrimination d e p e n d on the parti- cular application, a n d socrates therefore allows the knowledge engineer to adjust these discrimination criteria to suit the p a r t i c u l a r ways in which the k n o w - ledge base retrieval m e c h a n i s m will be used. f o r example, if the m a i n control regime for a p a r t i c u l a r application is s o m e f o r m o f b a c k w a r d chaining, then p r o p o s i t i o n s o f the f o r m "left-hand-side ~ r i g h t - h a n d - side' will be retrieved f r o m the knowledge base on the basis o f the p a t t e r n s in the ' r i g h t - h a n d - s i d e ' a r g u m e n t . this implies t h a t the discrimination net should first dis- c r i m i n a t e implications on the basis o f their consequent, r a t h e r than their antecedent. as a second example, we note that m a n y predicates use a certain n u m b e r o f their a r g u m e n t s as ' i n p u t s ' , while others are used as "output" a r g u m e n t s . a predicate such as ' s u f f e r s - f r o m (patient, disease)' will typically be used to associate a given patient with a disease, r a t h e r t h a n to try a n d find all patients suffering f r o m a given disease. this p a r t i c u l a r use o f a r g u m e n t s indicates t h a t the discrimination net should use the first a r g u m e n t as the m a i n discrimination criterion, r a t h e r t h a n the second a r g u m e n t . because these p r o p e r t i e s are specific to a p a r t i c u l a r application, they can only be adjusted by the k n o w l e d g e engineer w h o configures the system, r a t h e r t h a n being hardwired into the retrieval m e c h a n i s m o f the knowledge base. proof theory declaration t h e declaration o f a logical language t o g e t h e r with a storage and retrieval m e c h a n i s m allows us to represent knowledge, but in o r d e r to m a n i p u l a t e this k n o w l e d g e to derive new conclusions we need rules that tell us what the legal d e r i v a t i o n s will be. since we have chosen logic as the r e p r e s e n t a t i o n language for the system, we are c o m m i t t e d to logical d e d u c t i o n as the inference process. w e therefore need to define a p r o o f theory, i.e. a set o f inference rules* t h a t tell us h o w logical p r o p o s i t i o n s can be m a n i p u l a t e d in o r d e r to p e r f o r m a p r o o f . in socrates we follow biedsoe s in using the natural d e d u c t i o n style o f p e r f o r m i n g p r o o f s r a t h e r than, for instance, resolution. a l t h o u g h n a t u r a l d e d u c t i o n systems use a relatively large n u m b e r o f inference rules (as o p p o s e d to the single inference rule o f resolution based systems), and thereby create a p o t e n t i a l c o n t r o l *we use the term "inference rule' in the logician's sense: an inference rule is a rule that describes how true formulas can be inferred from other true formulas, and is of the form . f t . • • f , ~- g , where g is inferred from j ~ . . . f , . the term 'inference rule' is often incorrectly used to denote formulas of the form .f--, g (i.e. logical implications). p r o b l e m , a n u m b e r o f reasons can bc given in f a v o u r o f n a t u r a l deduction. t h e inference rules o f a n a t u r a l d e d u c t i o n system are m o r e intuitively meaningful than, for instance, the resolution rule. f u r t h e r m o r e , no n o r m a l f o r m s are required for the f o r m u l a s used in a proof. as a result, the p r o o f s p e r f o r m e d by a n a t u r a l deduction system are easier to follow for a h u m a n reader, thereby i m p r o v i n g the possibilities for e x p l a n a t i o n facilities. t h e naturalness o f the p r o o f d e v e l o p m e n t also m a k e s it easier to identify heuristics to c o n t r o l the p r o b l e m solving pro- cess, in the m a n n e r discussed below. w h e n expressing inference rules, the k n o w l e d g e engineer can m a k e use o f extra-logical variables that range o v e r well f o r m e d f o r m u l a s f r o m the object-level representation language. f o r instance, the rules p , p ~ q ~ q p ~ - p v q represents the rules f o r m o d u s ponens a n d disjunction i n t r o d u c t i o n . it is i m p o r t a n t to stress that these inference rules can be used in b o t h a f o r w a r d a n d a b a c k w a r d direction. f o r instance, m o d u s ponens can be used to d e t e r m i n e t h a t p a n d p -~ q have to be p r o v e d in o r d e r to p r o v e q ( b a c k w a r d use), o r the rule c a n be used to infer that q is true when we k n o w t h a t b o t h p and p --* q are true. f u r t h e r m o r e , because o f the associativity and c o m m u t a t i v i t y o f s o m e o f the logical connectives, a single inference rule c a n often be applied in m o r e t h a n one way. f o r instance, if disjunction (v) has been declared as a ' s e t - o p e r a t o r ' , disjunction i n t r o - d u c t i o n can be applied b a c k w a r d to (/'v g) in t w o differ- ent ways, binding p to e i t h e r . f or g, thereby generating either f o r g as a sub-goal for p r o v i n g "v g). h o w e v e r , which o f these possible a p p l i c a t i o n s o f an inference rule should be used is a control decision, a n d is therefore a meta-level issue, which is not decided as p a r t o f the p r o o f theory. n o t all the inference rules t h a t the system uses are declared as p a r t o f the p r o o f strategy. first o f all, there is a set o f inference rules t h a t tell the system h o w to deal with t y p e d quantification. these rules: v x:t~ p[x] ~ p[c:t ] for an a r b i t r a r y c o n s t a n t c, a n d all sorts t~ and t with t~ _~ t ~ ~' p[c:t~] ~ x:tz p[x] for an a r b i t r a r y c o n s t a n t c, a n d all sorts t~ and t with t ~ t~ d ~)* x:tt v y:t p[x,y] i'-- v y:t q x:t~ p[x,y] for all sorts t~ a n d t are taken to be o f universal validity (that is: across differ- ent a p p l i c a t i o n areas o f the system), a n d are therefore hardwired into the retrieval m e c h a n i s m . a second set o f rules, t h a t is p a r t o f the retrieval m e c h a n i s m in the knowledge base r a t h e r t h a n the p r o o f theory, are the results that deal with the c o m m u t a t i v i t y a n d associativity o f certain logical connectives. f o r a n y o p e r a t o r ~ that has been declared as " b i n a r y - c o m m u t a t i v e ' , the inference rule p d p q ~ q ~ p is hardwired into the knowledge base retrieval mechanism, as are, for every ' s e t - o p e r a t o r ' , the additional rules p f - pdp p (p dp (q dp r)) i-- ((p ~ q) dp r). *it is exactly because of this rule that in the section on logical represen- tation declaration we required all sorts except ~ to be non-empty. k n o w l e d g e - b a s e d systems by taking all these rules out o f the explicit declaration o f the p r o o f strategy and transferring them to the retrieval mechanism o f the knowledge base, we have given a limited deductive capability to the knowledge base. the retrieval mechanism comprises two sequential stages, a syntactic 'pattern matcher' and a 'unifier'. syntactic pattern matching is effected using the discrimi- nation net technique mentioned above. given a formula as a query to the knowledge base, this process will retrieve all formulas with the same pattern o f operators and predicates, subject to the commutativity and associa- tivity rules as declared for the particular language. this means that unification will be attempted only on patterns that have a strong possibility o f succeeding. the patterns that can be specified as input to the syntactic pattern maching phase are allowed to contain 'recta-logical (propositional) variables' that can match with formulas o f the logical representation language rather than with terms. these meta-logical variables will be bound to the corresponding components o f the matching expression, using pattern matching under the inference rules governing commutativity and associa- tivity. for example, the pattern (& f ?p) will match with a formula like (& g f ) , binding ?p with g after applying commutativity. to facilitate the retrieval o f expressions containing set-operators, a special version o f these meta-logical variables is available, the so-called 'segment-variables'. these segment variables do not match with formulas, but with lists o f logical formulas. for example, a formula like (& f g h) will match with a pattern like (& g ?p'segment), binding ?p with the list ( f h). because o f the meta-logical variables, the patterns sent to the knowledge base for retrieval are actually schemata, representing whole families o f queries rather than just a single query. in the second phase o f the retrieval process the unifier will construct bindings o f the logical variables occurring in the query. this is necessary since in the context o f expert systems we are interested in performing construc- tive proofs, and we therefore need the values for the existentially quantified variables in the query for which the query succeeds. in other words, for a query such as x:ttp(x), we want not only a yes/no response, but also the values o f x which can be deduced from the contents o f the knowledge base. notice that these bind- ings might only consist o f restrictions on the sort o f x, rather than o f bindings o f x to terms. the sorted unification algorithm might tell us that the query succeeds for all values o f x o f a certain sort t , with t c t,. in this way, taxonomic reasoning is performed at retrieval time. a final point to be made about the declaration o f the p r o o f theory concerns the soundness and complete- ness o f the set o f inference rules. in order to guarantee soundness o f the p r o o f theory, the knowledge engineer should not be allowed to declare arbitrary inference rules, but only to select inference rules from a predefined (and sound) set*. this selection process will o f course affect the completeness o f the system. however, the loss o f completeness in the context o f expert systems is not serious, since one does not want to infer a / / f a c t s that follow from the available knowledge, but only those facts that one is interested in. *such a selection procedure has not been provided in the current implementation o f socrates. p r o o f s t r a t e g y d e c l a r a t i o n at this point, the knowledge engineer has declared both a logical language and a corresponding p r o o f theory. from a purely logical point o f view, no further declara- tions are necessary. the combination o f language and p r o o f theory determines all the possible inferences that the system can make. however, in order to create a practical computer system, one more step has to be taken. proving statements in any non-trivial logical language is a search intensive problem. the logical language and p r o o f theory together define a search space for the p r o o f process. what remains to be done is the specification o f the strategy that the system should use to traverse this space while searching for a proof. for this task, socrates provides a declarative language for representing such a control strategy, described below. because such a declarative language has certain dis- advantages associated with it, a more procedural language has also been investigated (see below). declarative representation of control knowledge socrates allows the knowledge engineer to explicitly specify a control strategy. this control strategy provides the system with a description o f its desired behaviour, and is interpreted at run time by the meta-level inter- preter. as a result, the meta-level interpreter executes this control strategy, and thereby guides the search through the space o f all possible proofs. the language that is used to express the control strategy is a many sorted version o f horn clause logic. this language, although also a logical representation language, should be distinguished from the logical languages used to represent the domain knowledge. unlike the object-level languages, the language used at the meta-level has a fixed set o f logical connectives, namely exactly those connectives needed in horn clause logic: conjunction, implication and negation, plus disjunction. all these connectives are declared as non-commutative, non- associative. this is done because the procedural interpre- tation (i.e. the way in which the meta-level interpreter executes expressions o f the language) is also fixed. the procedural interpretation o f the language is the standard interpretation for horn clauses, the standard depth-first p r o o f procedure as found in prolog systems. the reason why socrates does not allow the knowledge engineer to change the control regime o f the meta-level interpreter (which would amount to providing a meta-meta-level interpreter~f) follows from the analysis o f models o f rationality, sketched above. as indicated, typical expert system tasks such as diagnosis, planning, monitoring, etc. are related to particular control regimes. the meta- level controls the behaviour o f the object-level interpreter according to the expert system task. the variation in control is achieved by changing the meta- level knowledge base. there is no need to change the interpreter, which always has the same task, namely con- trolling the behaviour o f the object-level interpreter by using the data in the meta-level knowledge base. the parts o f the meta-level language that are still subject to declarations made by the knowledge engineer tthe notion of a meta-meta-level interpreter should not be confused with the scheduler; the relation between meta-level interpreter and scheduler is very different from the relation between object-level and meta-level interpreter. voi i n o june are therefore the set o f constants, predicates and function symbols, the sort hierarchy, and the set o f "evaluable predicates'. figure shows an example o f a description o f a local-best-first non-exhaustive backward chaining control strategy. f o r this control strategy a sort hierarchy was defined containing the sorts formula, list- of-formulas, substitution and list-of-substitutions. t h e sort formula was further subdivided into c o m p o u n d - , evaluable- and non-evaluable-formula. a further specialization o f list-of-formulas was non-empty-list-of- formulas. in the example shown in figure , clause ( ) states that in order to prove a n o n - c o m p o u n d expression f on the basis o f the contents o f knowledge base partition p giving a substitution s as a result, the system should either try to see if the formula is a know n fact in the knowledge base, try to infer the formula on its own, or it should ask the end-user. trying to infer the formula means generating all possible inferences, select- ing some o f these possible inferences, and continuing with clause ( ). t h a t clause chooses the best o f all selected possible steps, and tries to continue the p r o o f with this selection. i f this succeeds, the p r o o f terminates (i.e. non-cxhaustive), if this fails, the p r o o f continues with the next best step. clause ( ) states the criterion used in the best-first search, while clause ( ) describes what needs to be done in order to prove a c o m p o u n d expression: prove both left- and right-hand sides o f the c o m p o u n d expression, and c o m b i n e the results. clause ( ) states that all evaluable predicates e n c o u n t e r e d in a p r o o f should be evaluated without any further control scheduling. this example shows how the different aspects o f this strategy can be changed if needed for a particular appli- cation. f o r example, the order o f the disjuncts in clause ( ) might be changed to ask the end-user for solutions before the system tries a p r o o f itself, or the ask-user (i) (v f:non-evaluable-formula, p:partition, s:substitution, next:non-empty-list-of-formulas. somenext:non-empty-list-of-formulas) [knowledge-base-lookup(f, p, s) v ( object-level-interpreter(f, p, backward. next) & ~lect-inferences(next, somenext) & infer(somenext, p, s)) v ask-uscr(f, p, s)) -oproof(f. p, s)] ( ) (v inferences:non-empty-list-of-formulas, p:partition, s:substitution, best:formula, rest:list-of-formulas) [best(inferences. best,rest) & (proof(best, p. s) v infer(rest, p, s)) -,infer(inferences, p. s)] ( ) (v list-non-empcy-list-of-formulas, best:formula, rest:list-of-formulas) [higher-certainty-value(list, best,rest) --.best(list, best,rest)l ( ) (v f:compound-formula. p:partition, s:substitution, lhs:formula, rhs:formula, lhssubst:substitution, rhssubst:substitution) [split-compound-expression(f, lhs, rhs) & proof(lhs;, p, lhssubst) & proof(rhs, p, rhssubst) &-combine(lhssubst. rhssubst, s) --* proof(f, p, s)] ( ) v f:evaluable-formula, p:partition, s:substitution. [evaluate(f. p, s) -, proof(f, p, s)] figure . declarative specij~cation of control in socrates disjunct might be deleted altogether. t h e criterion used for the best-first scheduling could be changed, or a new decision for scheduling the o rd er in which conjuncts are proved in clause ( ) could be introduced. m o r e t h o r o u g h changes to the strategy could also be made, but they would a m o u n t to writing a completely new p r o o f strategy rather than changing the one shown in this example. o f the evaluable predicates in figure (such as ask- user, knowledge-base-lookup, combine) the predicate object-level-interpreter is the most i m p o r t a n t one. this predicate encapsulates the interface between the meta- level interpreter executing the co n t ro l strategy, and the object-level interpreter handling the logical representa- tion language and the c o r r e s p o n d i n g p r o o f theory. the input o f this predicate is an object-level formula f, the name o f a knowledge base partition p and a direction in which to apply inference rules (either forward or back- ward), and returns as o u t p u t the result o f applying all inference rules specified as part o f the p r o o f theory for partition p to the input fo rm u l a f in the indicated direc- tion. in terms o f the p r o o f search space, this a m o u n t s to generating all nodes that are accessible from the current n o d e as represented by f. thus. the predicate object-level-interpreter allows the meta-level interpreter to access an explicit representation o f the object-level search space, and to choose which branches o f the object-level p r o o f tree will be ex p an d ed on the basis o f the control regime provided by the knowledge engineer. unlike m a n y logic-based meta-level architectures pro- posed in the literature, (such as silver t°, or the prolog system described inz°), socrates completely separates the languages used at the object-level from the language used at the meta-level. even when the object-level repre- sentation language h ap p en ed to be defined as sorted h o r n clause logic, the two languages would still be syntactically separate. th e meta-level and object-level languages are connected t h ro u g h a 'n am i n g relation'. th e meta-level language contains names for all object- level expressions. in socrates, the n am e o f an object-level sentence c o r r e s p o n d s to a co n st an t in the meta-level language. o t h e r meta-level constants are used to denote bindings for object-level variables. if this is required, meta-level expressions could range o v er any o f the extra- logical properties o f object-level expressions, such as truth values, certainty factors, justifications, etc. in this way, for instance, socrates could be configured to deal with certainty values by specifying as part o f the control strategy how certainty values should be used in a proof. this c o r r e s p o n d s to the a p p r o a c h suggested by shapiro ;, with the i m p o r t a n t difference that socrates makes a correct distinction between meta-level and object-level languages, whereas shapiro confuses the two and uses prolog for both. a n u m b e r o f reasons can be given why it is i m p o r t a n t for the meta-level language to be separated from the object-level languages. first, there is an epistemological reason: as argued in ~, different d o m a i n s require different representation languages, and the object-level and the meta-level o f socrates deal with widely different d o m a i n s (the object-level deals with the application d o m a i n o f the system, while the meta-level deals with the issue o f controlling the object-level). a second argu- ment concerns the m o d u l a r i t y o f the system: it should knowledge-based systems be possible to vary control knowledge and domain knowledge independently. the third argument is one o f explanation: in order to enable the system to include control knowledge explicitly in its explanations, it is important for both the human reader and the automated explanation generator that control knowledge can be syntactically distinguished from domain knowledge. procedural representation of control knowledge the approach to the specification o f a p r o o f strategy described above is based on the use of a declarative meta-level language. although the declarative style o f control o f reasoning has its attractions, there are also disadvantages. two problems in particular are caused by the use o f a declarative language, and in an attempt to overcome these problems socrates provides an alter- native, more procedural language for specifying control regimes. first, the extra layer o f interpretation that is incurred by the explicit meta-level interpreter is expen- sive, because o f the declarative nature o f the meta-ievel language. a procedural meta-level language would be much closer to the underlying implementation language and machine architecture, and therefore cheaper to execute. second, much o f the knowledge expressed in the meta-level language is procedural, rather than declarative. for example, one often wishes to apply knowledge o f the form 'try method-i before method- ', or 'in order to achieve goal-l, achieve sub-goal-i to subgoal-n'. in the declarative meta-level language this type o f procedural knowledge has to be expressed by either relying on the hardwired control regime for the meta-levei interpreter, or by using semantic attachment. neither o f these ways o f expressing procedural knowledge is very desirable, since they encode knowledge implicitly rather than represent it explicitly. a more pro- cedural language provides a more natural medium for expressing the procedural control knowledge. our approach to procedural control is therefore to provide a 'meta-level programming language' rather in the vein o f m l ° in its relationship to lcf. that is, we provide high level primitives that make the writing o f control regimes easier. it is important to note that a procedural control language is only an alternative way o f implementing the separation o f control knowledge from object-level knowledge. the procedural approach still clearly separates control knowledge from object- level knowledge. the difference is that rather than put- ting the control knowledge into a declarative knowledge base with its own interpreter, we propose implementing a special purpose meta-level interpreter that incorporates the control knowledge. the three stage process o f build- ing systems is retained. one particular point to note is that we retain the explicit declaration o f inference rules and provide primitives to apply inference rules. the procedural meta-levei language consists o f the implementation language o f the system (common lisp), extended with primitives implementing standard artifi- cial intelligence techniques that have been found useful in writing interpreters. central to the system is the use o f lazy evaluation. the procedural meta-level language provides facilities for the manipulation o f lazily evaluated lists, which form the basis for backtracking and coroutining in the control regimes. the second important component is agenda- (defun proof (goallist partition subst &aux newgoals newsubst) (if goallist (foreach (newgoals. new subst) in (or (any-of (order-inferences (evaluate (first goallist) :partition partition :subst subst)) (order-inferences (lookup (first goallist) :partition partition :subst subst)) (select-inferences (generate-backward-inferences (first goallist) :partition partition :subst subst))) (ask-user (first goallist) :partition partition :subst subst)) generate-each (proof (append newgoals (cdr goallist)) partition newsubst)) (list subst))) figure . procedural specification o f control in socrates based reasoning. this technique allows the knowledge engineer to experiment with several different control strategies, often only changing the way the agenda is handled without changing the rest o f the control regime code. the third component is a pattern matcher which provides the basis o f pattern directed invocation of p r o o f methods. it is important to realize that in this procedural approach it is still the case that the only inference rules are those declared explicitly during the declaration o f the p r o o f theory (as described above). language primi- tives are provided to apply the declared inference rules. figure shows an example o f a control regime formu- lated in the procedural meta-level language. this code is the procedural equivalent o f the declarative prover given in figure . the main function is called 'proof' and performs the same function as the predicate o f that name. being stream-based, p r o o f returns a stream o f substitutions that prove the goals in the 'goallist' argu- ment. thus, if the goallist is empty there is one such proof, given by the subst argument. if the goallist is not empty then the inference rules are applied to the first goal in goallist. the ' a n y - o f macro is a way o f lazily combining streams. thus, in this example, we generate all the possible inferences using 'evaluate' (and put them in a preferred order), then all the possible inferences using 'lookup' (again in preferred order), finally followed by a selection o f the possible inferences generated using all the inference rules. ' i f and only if' no possible inferences are generated by this procedure, then 'ask-user' is applied to obtain possible inferences. each inference rule application generates three things: a set of new goals to be proved in order to prove the goal (bound to the variable 'newgoals'); a new substi- tution, bound to the variable 'newsubst'; and a justifi- cation, which merely describes which inference rule has been applied and is effectively ignored by this prover. the 'foreach' macro itself generates a stream o f answers. thus, if at some later stage in the p r o o f there is a need to backtrack, the next inferences will be generated only then. note that, unlike the declarative prover, which generates only the first proof, the procedure ' p r o o f returns a stream o f proofs, and thus generates the set o f all proofs (lazily). it was noted at the beginning o f this section that much vol no june control knowledge seems very procedural in nature (such as the execution o f a sequence o f goals). however, some- times control knowledge is declarative in nature (e.g. a set o f criteria used in the ordering o f conjuncts). the architecture o f socrates is such that even with procedural control it is still possible to invoke a declarative meta- level interpreter that is implemented using the techniques described above. although the above features provide a basis for a procedural language for formulating control regimes, certain problems remain. first, the current language may not be powerful enough, and further additions may be needed. second, although in one sense the current proce- dural meta-level language might not be powerful enough, in a n o t h e r sense it might be too powerful. as described, the procedural meta-level language consists o f extensions to c o m m o n lisp, thereby making all the general purpose expressive power o f lisp available to the knowledge engineer when writing control regimes. it might well be the case that this provides too powerful a language, since it does not restrict the knowledge engineer in any way. the scheduler a third level in the architecture o f socrates (see figure ) is the 'scheduler'. this third level is not actually imple- mented in the current socrates architecture, but it can be added to the current system with little effort. the main notion that is treated at this level is that o f a "subtask'. as shown in reichgelt and van harmelen , m a n y expert systems perform not just one simple task, but a composite one that can be thought o f as consisting of a number o f elementary tasks ( m y c i n , ri and vm are a m o n g the systems discussed in that paper). it is unlikely that one appropriate control regime can be found that would be suitable for these composite tasks. rather, the composite task should be split up into its constituent subtasks, and a proper control regime can then be chosen for each o f the subtasks. the subtasks that would result from this decomposition process are the kind o f prototypical tasks proposed in reichgelt and van harmelen , chandrasekaran , and breuker and wielinga , like classification, monitoring, simulation, design, etc. the scheduling level o f the socrates architecture is meant to deal with this sub- division o f the major task into prototypical subtasks. each o f these prototypical subtasks can then be solved using the appropriate meta-level control strategy. (by using the knowledge base partition mechanism, it is possible to equip a socrates configuration with more than one control strategy.) f o r engineering purposes it would be easiest to equip the scheduling level with a language similar to (but again syntactically separate from) the language used to describe the control strategy at the meta-level. however, early experience indicates that the type o f knowledge to be expressed at the schedul- ing level is o f a very procedural nature (even more so than the knowledge expressed at the meta-level), and therefore a language with more conventional procedural primitives, such as sequences, conditionals, loops and subroutines, might be more appropriate. practical p r o g r e s s a n d a c h i e v e m e n t s a version o f the socrates architecture as described above has been implemented in approximately k lines o f c o m m o n lisp code. (notice the distinction between "implementation language" and "representation language'of a system: although the knowledge represen- tation and inference process o f socrates are both logic based, the system is n o t implemented in prolog or any other logic programming language.) the system currently runs in a number o f c o m m o n lisp implemen- tations on u n i x based systems. in one o f these c o m m o n lisp systems, the poplog system, a graphics- based knowledge engineer interface, has been con- structed, allowing interaction with the system via menus, browsers, graphers, etc. a substantial set o f different control strategies has been written as meta-level programs, including backward and forward chaining, exhaustive and non-exhaustive search, user guided or automatic conflict resolution, best-first, depth-first, breadth-first search, branch and bound type algorithms, generate and test procedures, elimination and confirmation strategies, etc. a number o f demonstration systems have been built using socrates, including an expert system in the domain o f personal investment advice and a route planning system. more significantly. socrates has been used to reimplement an existing expert system under the name o f docs, developed by g e c research in collaboration with westminster hospital, london. this system con- sists o f rules divided over two knowledge base partitions. the partition hierarchy is organized so that these two partitions can use data from a c o m m o n work- ing memory partition. a sort hierarchy o f over sorts was used to model the taxonomic hierarchy o f the medical domain. a generate and test control strategy was specified tbr this system, consisting o f some clauses in the meta-level knowledge base. in the area o f theorem proving socrates has been used to solve the problem described by walther ', known as 'schubert's steamroller'. this problem was originally formulated because o f the huge search space that it generates. using the sort hierarchy to model the taxonomic part o f the problem, a breadth-first search strategy, implemented using our procedural control tech- niques, resulted in the first natural deduction style p r o o f for this problem. a number o f other procedural control strategies were implemented to solve the problem, and a comparison o f these different strategies showed the importance o f the exploitation o f meta-knowledge to guide the search for a proof. this application o f socrates is described in detail by davies . current areas o f activity are: • the use o f meta-level interpretation for dealing with extra-logical issues, such as uncertainty, truth maintenance, explanation, etc. (how the slot-value a n n o t a t i o n mechanism o f socrates' knowledge base opens up these possibilities is described above). • the implementation o f modal logics through the use o f reification, as described by reichgelt ~. • the classification o f domains and subtasks, as stated by reichgelt and van harmelen s' , in order to provide the knowledge engineer with guidelines that indicate how to choose the appropriate representation language given a particular application domain, and how to choose the appropriate control regime given a particular prototypical task. • the development o f a library o f control regimes that knowledge-based systems can be executed by the system. rather than having to write a new control strategy from scratch every time, the knowledge engineer can use a library o f preprogrammed control strategies. the knowledge engineer can then either use one o f the strategies directly from the library, or use one o f the library elements as the basis for his own control strategy by making small changes to the preprogrammed strategy. o p e n p r o b l e m s o f the three stages o f the configuration process, the third one (declaring a p r o o f strategy) is by far the most problematic. an important open question here is the choice o f a good language for specifying such a control strategy. as described above, the system primarily uses a declarative logical language to do this, but a procedural language has also been investigated. apart from the type o f language used at the meta-level, a related problem is the required vocabulary o f such a language. at the moment, the vocabulary o f the system is specified by the knowledge engineer, and although this is to a certain extent inevitable, since part o f the vocabulary will be application-specific, one would hope that at least a central core vocabulary can be distinguished that can be preprogrammed into the system. the example o f a control regime discussed above suggests predicates to do with manipulating substitutions and formulas, and with generating the object-level search space, but a more extensive and more exactly defined vocabulary is needed in order to alleviate the task o f the knowledge engineer. a second problem associated with the explicit meta- level interpreter is that o f meta-level overhead. although the flexibility in defining the appropriate control strategy at the meta-level can considerably reduce the object-level search space, the price we have to pay for this is the fact that the object-level inference process is completely simulated by the meta-level interpreter. this is obviously much more expensive than an object-level interpreter that has the appropriate control strategy hardwired into it. this problem could be solved by taking the explicit formulation o f a control regime, and compiling it into an interpreter that has the particular control regime hardwired into it. this compilation process (whose first stages could be similar to that described by altman and buchanan z) has been simulated in socrates by hand coding a number o f hardwired control strategies. experi- ence with these hardwired strategies in both the docs system and in solving schubert's steamroller indicates that the meta-level overhead can indeed be reduced to an acceptably small amount. c o n c l u s i o n the work described in this paper attempts to create an environment for building expert systems based on the following principles: • an epistemological analysis o f the domain and task o f a particular application guides the choice o f the appropriate knowledge representation language and the appropriate control regime. • logic is used as the main underlying formalism. • control knowledge is represented explicitly and is separated from the domain knowledge. when configuring the socrates environment into a particular expert system, a knowledge engineer can vary the architecture along three dimensions: • the representation language: a knowledge engineer can define his own logical representation language, including first order logics (possibly many-sorted), modal logics, temporal logics, etc. • the inference rules for the logical language: the set o f rules that determine the possible inferences made in the logical language can be changed by the knowledge engineer. • the control regime under which the inference rules will be used to perform proofs in the logical represen- tation language. an implementation o f the socrates abstract architecture and a number o f applications o f the system have proved the feasibility o f this approach. due to limitations o f space, some o f the arguments and descriptions in this paper are rather terse. a long version o f this paper can be found in corlett, davies, khan, reichgelt and van harmelenl a c k n o w l e d g e m e n t s as the grant holder, peter jackson has made many con- tributions to the work described above. ray mcdowell and dave brunswick have contributed to the work done at gec research. the research was carried out as part of aivey project ikbs/ in which g e c research, the university o f edinburgh and g e c avionics were partners. the university work was supported by serc grant gr/d/ . r e f e r e n c e s aivey, p 'problems o f designing a medical expert system' rd tech. conf. british comput. soc. specialist group on expert syst. (december ) pp - bobrow, d and stef'~, m the loops manual rank xerox, u k ( ) the knowledge engineering environment intellicorps, california, usa ( ) williams, c art, the advanced reasoning tool, conceptual overview inference corporation, los angeles, california, usa ( ) breuker, j a and wiefinga, b j 'models o f expertise' proc. th europ. conf. on artif. intel. (july ) pp - chandrasekaran, b 'towards a functional architec- ture for intelligence based on generic information processing tasks' th int. joint conf. on artif. lntel. (august ) pp - corlett, r, davies, n, khan, r, reichgelt, h and van harmelen, f 'socrates: a flexible toolkit for building logic based expert systems' in jackson, p, reichgelt, h and van harmelen, f (eds) logic-based knowledge representation mit press, usa ( ) davis, r 'meta rules: reasoning about control' artif. lntel. vol no ( ) pp - bundy, a and welham, b 'using meta-level inference for selective application o f multiple rewrite rules in algebraic manipulation' artif. intel. vol n o ( ) pp - vol no june clancey, w "the advantages of abstract control knowledge in expert system design' proc. rd annual meeting o f th ~ american assoc..for a rt([~ intel. ( ) pp - aiello, l and levi, g 'the uses o f metaknowledge in ai systems' proc. european conf. on artif. intel. (september ) pp - clancey, w 'representing control knowledge as abstract tasks and metarules' in cooml~, m and bolc, l (eds) springer-verlag, f r g ( ). also: stanford knowledge systems laboratory, working paper no ksl- - (april ) claneey, w 'the epistemology of a rule-based expert system: a framework for explanation' artif. lntel. vol ( ) pp - . also: stanford heuristic programming project, memo hpp- - , stan-cs- - (november ) bundy, a, byrd, l, luger, g, mellish, c, milne, r and palmer, m 'solving mechanics problems using meta-level inference" proe. int. joint conf. on artif. lntel. (august ). also in michie, d (ed.) expert s.vstems in the micro electronic age edinburgh university press, uk ( ) pp davis, r 'teiresias: applications of meta-level knowledge' in davis, r and lenat, d b (eds) knowledge-based systems in artificial intelligence mcgraw-hill, new york, usa ( ) pp ~, pereira, l m "logic control with logic' proc. st int. logic programming conf. ( ) pp - . also in: campbell, j a (ed) implementations ~?fprolog ellis horwood, uk ( ) genesereth, m and smith, d 'an overview o f meta- level architecture' proc. rd annual meeting o f the american assoc. for art(f. intel. ( ) pp - sterling, l 'implementing problem solving strategies using the meta-level" dai research paper no , department o f artificial intelligence, university of edinburgh, uk ( ) silver, b meta-level injerence studies in computer science and artificial intelligence, north holland, amsterdam, the netherlands ( ) welham, b 'peclaratively programmable interpreters and meta-level inferences' hewlett-packard labora- tories bristol research centre technical memo no hpl-brc-tm- - (september ). also in: maes, p and nardi, d (eds) meta-h, vel architectures an'l r~/h, ction north holland. amsterdam. the netherlands ( ) waltber, c "a mechanical solution of schubert's steamroller by many-sorted resolution" proc. th annual meeting of the american assoc. art(f. lntel. ( ) pp - cohn, a g 'ol the solution of schubert's steamroller in many sorted logic" pro,'. th int. join t. con/i on art(/: lntel. (august ) pp - davies, n "schubert's steamroller in a natural deduction theorem prover" proc. th tech. conf. o f the british comput. soc. specialist group on expert syst. (december ) weyhraueh, r 'prolegomena to a theory of mechanised formal reasoning' art(/: hztel. vol no ( )pp bledsoe, w 'non-resolution theorem proving" artif. intel. vol no i ( ) pp i-- gallaire, m and lassetre, c 'meta-level control for logic programs" in clark, k and tarnlund, s (eds) logic programming academic press. uk ( ) pp shapiro, e "logic programs with uncertainties: a tool for implementing rule-based systems" proc. th int. joint conf. on art(f. lntell. (august ) pp - reichgelt, h and van harmelen, f "relevant criteria for choosing an inference engine in expert systems" proc. th tech. conf. o f the british comput. soc. specialist group on expert syst. (december ) pp -- reichgelt, h and van harmelen, f "criteria for choos- ing representation languages and control regimes for expert systems" the knowledge eng. rev. vol i no (december ) pp -- gordon, m, milner, r and wadsworth, c edinburgh lcf." a mechanized logic o f computation springer- verlag lecture notes in computer science, vol springer-verlag, f r g ( ) reichgelt, h 'semantics for a reified temporal logic" proc. a i s b conf. (april ) pp altman, r b and bwehanan, b g 'partial compilation of strategic knowledge' proc. th annual meeting o[ the american assoc. /or , rt(/~ lntel. (july ) pp - knowledge-based systems the exploration of customer satisfaction model from a comprehensive perspective the exploration of customer satisfaction model from a comprehensive perspective wen-bao lin * national formosa university, department of business administration add: , wen-hua road, huwei, yunlin, taiwan, roc abstract this study provides a model of customer satisfaction from a comprehensive perspective and tries to use the nonlinear fuzzy neutral network model to verify the assumptions of the study. samples are taken from the information and tourism industries at a proportion of : based on the population in taipei and kaohsiung cities. a total of questionnaires are returned. as the result of the empirical research shows, the interpersonal-based service encounter is better than the technology-based service encounter in functional quality, while the technology-based service encounter is better than the interpersonal-based service encounter in technical quality. the functional quality has a positive and significant effect on customer satisfaction; the service quality has a positive and significant effect on service value; the service value has a positive and significant effect on customer satisfaction. the service encounter has a positive and significant effect on relationship involvement and the relationship involvement has a positive and significant effect on customer satisfaction. � elsevier ltd. all rights reserved. keywords: customer satisfaction; service encounter; service value; relationship involvement; service quality . introduction the industrial structure in taiwan has transformed from traditional agricultural and manufacturing industries to the tertiary service industry that is booming and domi- nating the market with its high employment rate and pro- duction value. in the circumstances, it is very important for the service industry to improve their service quality. the core value provided by the service industry to consumers includes not only the uniqueness of tangible and intangible products, but also various factors involved in the process of service delivery to customers, such as physical facilities, company image, and quality of the service delivery. emphasizing customers’ perception of service quality is important to enhance the extended perception and impres- sion of consumers on products, and to increase the added value of products. with the rise of consumer awareness, requirements of customers changing from mass production to customization and quality. therefore, how to under- stand the requirements of customers and provide them with products and services they need, reduce customer costs and improve customer value, and continually trace the satisfac- tion of consumers are essential for an enterprise to march toward success. much research on the cause of service quality and cus- tomer satisfaction has been published in recent years. this research focuses on the following dimensions: ( ) discuss- ing the antecedent variable that affects customer satisfac- tion. zeithaml and bitner ( ), for example, found that customer satisfaction was affected by service quality, prod- uct quality, price, personal and situational factors. oliver and desarbo ( ) found that service quality was the antecedent variable of customer satisfaction, while ander- son, fornell, and lehmann ( ) found that customer sat- isfaction was dependent on service value; ( ) presenting theoretical models to explain the importance and necessity of customer satisfaction. anderson and suillivan ( ) pointed out that the satisfaction or dissatisfaction of cus- tomers could be measured based on the gap between the - /$ - see front matter � elsevier ltd. all rights reserved. doi: . /j.eswa. . . * tel.: + ; fax: + . e-mail addresses: e @ms .hinet.net, wblin@sunws.nfu.edu.tw www.elsevier.com/locate/eswa expert systems with applications ( ) – expert systems with applications mailto:e @ms .hinet.net mailto:wblin@sunws.nfu.edu.tw expectations of customers before purchasing and the per- formance of the purchased products or services. woodruff, cadotte, and jenkins ( ) raised the norms in models of consumer satisfaction. they found that consumers evalu- ated products of different brands based on ‘‘norms’’ and their satisfaction was determined by the conformity among these norms. meuter, ostrom, roundtree, and bitner ( ) introduced the technology infusion matrix to ana- lyze how to enhance the provision of customized service by helping staffs or consumers use hi-tech tools and ensure the service failure compensation to improve customers satisfac- tion; ( ) bringing up research dimensions and approaches for measuring customer satisfaction and service quality. for example, perkins ( ) pointed out that industries were different in target market segmentation and, thus, measurement of customer satisfaction varied depending on the characteristics of individual industries. he raised three dimensions to be taken into consideration for empir- ical research on measurement of customer satisfaction: operation dimension (availability, delivery on time, price, credibility), service dimension (sales service, technical sup- port, product line), product dimension (technical value, reliability, design). parasuraman, zeithaml, and berry ( ) referred to ‘‘encounter’’ and ‘‘feasibility’’ as key dimensions. after interviewing senior management and consumers of four service industries and conducting com- prehensive exploratory research, they raised the five-gap theory and brought up an important measurement approach. with this approach, excess of expected service (es) over perceived service (ps) was represented by es > ps, indicating that the overall service quality was not acceptable, while satisfactory service quality was repre- sented by es = ps and excess of perceived service over expected service was represented by es < ps, indicating an inclination toward ideal quality. previous research on customer satisfaction had the fol- lowing characteristics: ( ) emphasizing the discussion of performance. oliver ( ), for example, found that satis- faction was a value consumers give to a deal. the higher the service level that consumers expect to have, the more difficult for the service provider to provide, the higher the customer satisfaction. according to the conformity theory of mittal and tsiros ( ), when the expectation of con- sumers was not in conformity with the actual performance of a product, they might change their perception of the product and incline to eliminate the gap between the expec- tation and performance, and thus improve the satisfaction; ( ) emphasizing the discussion of cause and result of satis- faction. cronin and taylor ( ), for example, evaluated service quality based on performance and argued that ser- vice quality was the cause of customer satisfaction. para- suraman, zeithaml, and berry ( ) researched the gap between service quality and customer satisfaction; ( ) explaining the cause of customer satisfaction from a single perspective, such as the attribution theory of bitner ( ) and the equity theory of oliver and desarbo ( ). kotler ( ) found that customer satisfaction was closely linked to customer value and the measurement of customer satis- faction should be carried out from the perspective of value. as some previous literature shows, the topics such as ‘‘ser- vice quality’’, ‘‘service value’’, and ‘‘customer satisfaction’’ were researched separately and only a limited number of literature focused on integration research in the relation- ship between the behavior intention of consumers and the service quality, service value, and customer satisfaction as a whole (cronin, brady, tomas, & hult, ). this study provides integrated research of these variables and focuses on their relationship as the first research motive; ( ) most of the satisfaction models used in previous research empha- sized the perceptive (kotler, ) and emotional perspec- tives (price, arnould, & tierney, ) of consumers. however, what a provider of the service industry empha- sizes is not only the interaction with consumers to ensure their satisfaction, but also the method of changing the trade exchange to the value increment exchange (the rela- tionship continuum theory of day, ). therefore, this study proceeds to the research from the perspective of rela- tionship involvement as the second research motive. besides, the multivariate statistic analysis approach or case study was commonly used for empirical research (bol- ton & drew, ; cronin, michael, brady, hightower, & shemwell, ; zeithaml, ) and the nonlinear research model (such as the nonlinear fuzzy neural network model) was not used often for theory verification. the non- linear fuzzy neural network model is maturely developed and has a wide range of applications. it is not only useful for prediction and assortment, but also for uncertain behavior systems. therefore, this study uses the nonlinear fuzzy neural network model to verify the initial data of the research as the third research motive. based on the above, the purpose of this study is listed as follows: ( ) to provide a customer satisfaction model from a comprehensive perspective of relationship involve- ment, service encounter, service quality, and service value. ( ) to verify collected data with the nonlinear fuzzy neu- ral network model and discuss its meaning for management. . literature review service encounters were classified as follows in the previ- ous research: ( ) according to the level of service encounter with the staff or facilities, the service encounter was classi- fied into high service encounter (contacting staff more fre- quently or contacting customers widely) and lower service encounter (mainly contacting physical facilities). industries belonging to the former include the hotel industry, insur- ance industry and tourism industry, while industries that belong to the latter include internet banks and couriers. however, many industries of high or middle service encoun- ters contact with customers and make deals with them w.-b. lin / expert systems with applications ( ) – https://isiarticles.com/article/ shell fitting space for classification expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a shell fitting space for classification mostafa ghazizadeh ahsaee ⇑, hadi sadoghi yazdi, mahmoud naghibzadeh department of computer engineering, ferdowsi university of mashhad, mashhad, iran a r t i c l e i n f o a b s t r a c t keywords: shell fitting space distance based transformation space fitting classification - /$ - see front matter � published by doi: . /j.eswa. . . ⇑ corresponding author. e-mail addresses: ghazizadehahsaee.most@stu-m ahsaee), h-sadoghi@um.ac.ir (h. sadoghi yazdi), naghibzadeh). in this paper, a shell fitting space (sfs) is presented to map non-linearly separable data to linearly sep- arable ones. a linear or quadratic transformation maps data into a new space for better classification, if the transformation method is properly guessed. this new sfs space can be of high or low dimensionality, and the number of dimensions is generally low and it is equal to the number of classes. the sfs method is based on fitting a hyper-plane or shell to the learning data or enclosing them into a hyper-surface. in the proposed method, the hyper-planes, curves, or cortex become the axis of the new space. in the new space a linear support vector machine (svm) multi-class classifier is applied to classify the learn data. � published by elsevier ltd. . introduction classification is an important research area with a wide range of applications. nonlinear discriminant functions (ndf) are useful in training a system to recognize specific patterns and now many applications are based on this method. neural network and support vector machine are preeminent mathemat- ical tools of ndf. support vector machines (vapnik, ) are very popular and powerful in learning systems because of the utilization of kernel machine in linearization, providing good generalization properties, their ability to classify input patterns with minimized structural misclassification risk and finding acceptable separating hyper-plane between two classes in the feature space. the result of applying kernels allows the algorithm to fit the maximum-margin hyper-plane in the transformed feature space. the transformation may be non-linear and the transformed space may be high dimensional; thus though the classifier is a hyper-plane in the high-dimensional feature space it may be non-linear in the original input space. if the used kernel is a gaussian radial basis function, the corresponding feature space is a hilbert space of infinite dimension. maximum margin clas- sifiers are well regularized, so the infinite dimension does not spoil the results. of course kernel methods (kms) (abe, ; huang, kecman, & kopriva, ; scholkopf & smola, ; shawe-taylor & cris- tianini, ) map input space into a high dimensional feature (hdf) space that may be helpful in linearization. as we know some kernels were proposed for this purpose, namely polynomi- elsevier ltd. ail.um.ac.ir (m. ghazizadeh naghibzadeh@um.ac.ir (m. als, gaussians, and splines (friedman, ), but these kernels do not guaranty linearization in hdf. this problem motivates us for presentation of new space in which patterns can be clas- sified by linear classifier, we name it shell fitting space (sfs) be- cause we use the concept of shell fitting for the creation of the new space. support vector machines (svm) and its variants (abe, ; lin & wang, ; sadoghi yazdi, effati, & saberi, ; wang, ; wu, jie-chi, & lee, ) is a particular instance of kms. but it has some weaknesses as follows. slow training (compared to neural network) due to computationally intensive solution to qp problem especially for large amounts of training data ) needs special algorithms. � the kernel to be used is not deterministic and it changes for each data set and finally a large feature space is produced (with many dimensions). � slow classification for the trained svm. � generates complex solutions (normally > % of training points are used as support vectors), especially for large amounts of training data. � difficult to incorporate prior knowledge. our proposed approach is not to expand the original space into a new space with many dimensions in comparison with kernel methods. in sfs the mapping is done to an m-dimensional space; where m is the number of classes. the rest of this paper is as follows. section is devoted to the development of our method, section explains how the new method works on example datasets, section is devoted to the application of the method on real datasets. in section we test our method with a simple linear classifier on sfs data and compare its accuracy results with multi-class svm results on input space data. conclusions are made in section . http://dx.doi.org/ . /j.eswa. . . mailto:ghazizadehahsaee.most@stu-mail.um.ac.ir mailto:h-sadoghi@um.ac.ir mailto:naghibzadeh@um.ac.ir http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – . shell fitting space formulation definitions: fxij ½x ; x ; . . . ; xk ; x ; . . . ; xk ; . . . ; x j; x j; . . . ; xkj j; . . .�; j ¼ ; . . . ; mg is the ith sample with n dimensions of class j. cj is the fitted curve, hyperplane, or surrounded cortex or (shell) to the set {(xij, yj), i = , . . . , kj} of data, where yj is the jth label of the training data and kj shows the number of samples with label j. in general our mapping is done from a space with m patterns (classes) of data in n-dimensional space to m-dimensional space by the fol- lowing notation: u : x ¼fx ; x ; . . . ; xng rn ! uðxÞ rm where / is a function that depends on distance of data to each pat- tern’s fitted curve, hyperplane or surrounded cortex or (shell). this transformation can be seen in fig. , where x is a feature in the input space and d = {d , d , . . . , dm} is the feature space element and ci is the class i. . . dealing with hyperplane when a hyperplane is fitted to a set of data the distance of the point xi from that plane is calculated with the following formula: suppose that the fitted hyperplane is y in ( ) y ¼ w x i þ w x i þ���þ wnxni ð Þ so the distance from y is obtained as ( ) dðxi; cÞ¼ w x i þ w x i þ���þ wnxniffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi w þ w þ . . . þ w n q ð Þ for instance in y = w x + w x we have the shape shown in fig. a. y w x w x= + ww xwxw = + + d (a) (b) fig. . (a) distance from a hyperplane an fig. . rbf-like transformation space. . . dealing with a hypersphere when a hypersphere is fitted to a set of data it can be formu- lated as: ðx � a Þ þðx � a Þ þ���þðxn � anÞ ¼ r ð Þ the distance between xi and c is calculated by ( ) dðxi; cÞ¼ r � ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x i þ x i þ���þ x ni q���� ���� � � ð Þ where r is as follows: r ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðx � a Þ þðx � a Þ þ�� �þðxn � anÞ q ð Þ for example, consider two points (x , y ) and (x , y ) in fig. . the distances d and d are calculated as shown in fig. b. . . dealing with a polynomial in two-dimensional space when a polynomial is fitted to a set of data, points obtained by minimizing eq. ( ) are examined to see which one corresponds to the actual global minimum. to do so, eq. ( ) is evaluated for each of these points and the one that gives a lower value is selected r ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðx � x Þ þðy � y Þ q ð Þ where y is as follows: y ¼ w þ w x þ w x þ���þ wnxn ð Þ for example suppose: y ¼ w þ w x þ w x þ���þ w x then we have: rðxÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðx � x Þ þðw þ w x þ w x þ���þ w x � y Þ q ð Þ the points that minimize r are calculated by setting the derivatives of r equal to zero. generally, by setting the derivatives of r equal to zero we have simple formulas as ( ) @r=@x ¼ x � x þðwnxn þ . . . þ w x þ w � y Þ � ðnwn xn� þðn � Þwn� xn� þ���þ w Þ¼ ð Þ then for each answer xi we find r(xi) from ( ) and at the end the minimum distance is: min frðxiÞjxi fanswersgg )y-()x-( yxrd +−= )y-()x-( yxrd +−= d (b) distance from a hypersphere. m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – . . dealing with a line in n-dimensional space generally an equation for a line in n-dimensional space can be written as follows: lðx; y; . . . ; zÞ : w ¼ a x þ b y ¼ a x þ b . . . z ¼ anx þ bn >>>< >>>: ð Þ so the distance between a point (x , y , . . . , z ) and (x, y, . . . , z) on l can be calculated with the formula below: dðw; x; . . . ; zÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðw � w Þ þðx � x Þ þ�� �þðz � z Þ q ð Þ considering eq. ( ), the distance can be expressed as: dðxÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ða x þ b � w Þ þðx � x Þ þ�� �þðanx þ bn � z Þ q ð Þ once again, we like to minimize ( ). therefore, its derivatives are set to zero and the x values for which d(x) is minimized are calculated @d=@x ¼ða x þ b � w Þa þðx � x Þþ �� �þðanx þ bn � z Þan ¼ ð Þ then for each solution xi, d(xi) is evaluated to find the actual global minimum distance, that is: min fdðxiÞjxi fanswersgg . . dealing with a curve in n-dimensional space if the number of features is more than two, we can use neural network-based methods like adaptive neural-fuzzy inference sys- tem (anfis) for this purpose. detailed explanation can be followed from wu et al. ( ) and a brief description is given in the appendix. steps of transformation using anfis are: (a) learn anfis the classes of data. (b) evaluate new data by the system. (c) calculate the distance with a subtraction operation; evalua- tion result is subtracted from the main value of data as dis- tance. you can see the outputs in section . . . dealing with a voluminous classes in n-dimensional space . . . svdd one-class classification three general approaches are proposed to resolve the one-class classification problem (tax, ). the most straightforward meth- od to obtain a one-class classifier is to estimate the density of the training data and to set a threshold on this density. several distri- butions can be assumed, such as a gaussian or a poisson distribu- tion. the most popular density models are the gaussian model, the mixture of gaussians, and the parzen density (bishop, ; par- zen, ). in the second method a closed boundary around the target set is optimized. k-centers, nearest neighbor method and support vector data description (svdd) are examples of the bound- ary methods (tax & duin, ; ypma et al., ). reconstruction methods are another one-class classification method which have not been primarily constructed for one-class classification, but rather to model the data. by using prior knowledge about the data and making assumptions about the generating process, a model is chosen and fitted to the data. some types of reconstruction meth- ods are: the k-means clustering, learning vector quantization, self- organizing maps, pca, a mixture of pcas, diabolo networks, and auto-encoder networks. to obtain data description for a group of n target objects in in- put space, we try to find a sphere with minimum volume which encloses all or most of these target objects. the problem of finding the minimum hyper-sphere represented by a center ‘‘a” and radius ‘‘r” can be formulated into: fðr; aÞ¼ r þ c x i ni ð Þ s:t: kxi � ak r þ ni; i; ni p ð Þ the parameter c gives the tradeoff between the volume of the description and the errors. the free parameters, a, r and ni, have to be optimized, taking the constraints ( ) into account. con- straints ( ) can be incorporated into formula ( ) by introducing lagrange multipliers and constructing the lagrangian: lðr; a; ai; ci; niÞ¼ r þ c x i ni � x i aifr þ ni �ðkxik � a:xi þkak Þg� x i cini ð Þ with the lagrange multipliers ai p and ci p . setting partial derivatives to gives these constraints: @l @r ¼ : x i ai ¼ ð Þ @l @a ¼ : a ¼ p iai xip iai ¼ x i ai xi ð Þ @l @ni ¼ : c � ai � ci ¼ ð Þ from the last equation ai = c � ci and because ai p , ci p , la- grange multipliers ci can be removed when we demand that < ai < c ð Þ resubstituting ( )–( ) into ( ) results in: l ¼ x i aiðxi � xiÞ� x i;j aiajðxi � xjÞ ð Þ by definition, r is the distance from the center of the sphere ‘‘a” to (any of the support vectors on) the boundary. support vectors which fall outside the description (ai = c) are excluded. therefore: r ¼ðxk � xkÞ� x i aiðxi � xkÞ� x i;j aiajðxi � xjÞ ð Þ because we are able to give an expression for the center of the hypersphere ‘‘a”, we can test if a new object ‘‘z” is accepted by the description. for that, the distance from the object ‘‘z” to the cen- ter of the hypersphere ‘‘a” has to be calculated. a test object z is ac- cepted when this distance is smaller than or equal to the radius: kz � ak ¼ðz � zÞ� x i aiðz � xiÞ� x i;j aiajðxi � xjÞ r ð Þ . . . using svdd for our distance based method here we used one svdd for each class of a dataset (for example svdd for classes). then to check if a data belongs to a class, its distance from cortex of each svdd classes is calculated and this distance is used as a main criterion to classify the data. in other words, a data belongs to a class, if its distance to a voluminous class’ cortex is less than the other classes’ in kernel space. the dis- tance is calculated by ( ). as stated above, like hypersphere-but in kernel space, the transformation is done m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – d ¼ðz � zÞ� x i aiðz � xiÞ� x i;j aiajðxi � xjÞ� r ð Þ fig. . fitting a line to each of data classes. . experimental results first the proposed circle fitting space transformation method is now demonstrated using simple data as an illustrating exam- ple, and at the end, our method will be tested using some well known datasets. in the following discussion, we assume that the label for each train data is known, in advance. here our method has been implemented and tested on matlab (math- work inc.). . . operation on synthetic data we describe our approach by some examples. at first we intro- duce our synthetic data as shown in fig. a–i. in fig. a there are two classes of linear shape which are linearly separable. but in other figures (b–i) are not linearly separable and these figures are divided into three categories: ( ) sphere-based fig. . (a) two linear classes; (b) circle classes; (c) half-circle classes; (d) half-circle class (h) two classes for anfis and (i) two classes for anfis in -d space. figures (b–f), ( ) sinuous function (fig. g): more complex classes which are not easily separable. ( ) fig. h and i: are used in learn- ing anfis. es; (e) half-circle classes; (f) half-circle classes interfere each other; (g) sin-function; fig. . transformation space of fig. . m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – consider two data classes in fig. a that are linear classes. firstly a line is fitted to each class of data with least square error fitting method. the result is seen in fig. . and then to map them to a new space we use a formula like the following to compute the distance of each data (x , y ) to the lines. here d(x , l ) stands for the distance of point x = (x , y ) from line d dðx ; l Þ¼ ja x þ b y þ c jffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi a þ b q dðx ; l Þ¼ ja x þ b y þ c jffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi a þ b q we use these distances to map data to a new -dimensional space like fig. which is linearly separable. fig. . (a) fitted spheres and fig. . fitting a half-circle t as the second example, consider fig. b. these classes are examples of hypersphere but in a d-space. curve fitting is done first and the distance between each circular curve and dataset ele- ments is calculated. the results are depicted in fig. a and b. hor- izontal axis shows the distance between each dataset element and the internal circle and the other axis is the distance between each dataset element and the external circle. . . transformation of circle classes our proposed method is applied to a data series in -dimen- sional space (two classes with two features) and the results are shown in figs. – . we have tested our method on four kinds of half-circle (fig. c–f). the two half-circles may even interfere each other (fig. f). the results of curve fitting and calculating distance between each dataset element and each fitted data are depicted in figs. – . . . transformation of more complex classes data distribution may seem so complex, but this method tries to separate these two classes of data (fig. ). . . transformation using anfis other experimental results are those which use anfis to fit a curve (figs. and ). here we have used two classes in - and -dimensional space. the two anfis were used to fit curves to the data set. as shown in fig. , the result is linearly separable. (b) output of mapping. o each of data classes. fig. . fitting a half-circle to each of data classes. fig. . fitting a half-circle to each of data classes. fig. . fitting a half-circle to each of data classes. m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – . . transformation using svdd as stated before, svdd is a one-class classification used in clas- sification of voluminous data. at this point, a dataset is shown in fig. a and the result of our proposed method is depicted in fig. b. . experiments on datasets in this section our proposed method is used to classify some datasets. these datasets are from uci machine learning reposi- tory. datasets used here are breast-cancer-wisconsin dataset, iris dataset, ionosphere dataset, transfusion dataset and heart dataset with , , , , , and attributes with , , , , , and classes of data in sequence. . . circle fitting on datasets as we said in previous sections, circle fitting is used in classifi- cation of like-circle data. we show here that our method is good to classify breast-cancer-wisconsin and iris datasets and to some ex- tent to classify ionosphere dataset using circle fitting. our method calculated the distance of each data in the dataset to circles fitted fig. . fitting a curve to each of data classes. fig. . curves have been fitted to data sets of two classes and the result of transformation has been shown in right figure. fig. . fitting a curve to a data set using anfis and transform them to a new linearly separable space. m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – on data and the classes are identified. we can see the result in figs. – . we can see from figs. and that data may have interfer- ence in their distribution or may not be well fitted to a hypersphere. . . using svdd on dataset here we used one svdd for each class of iris dataset ( svdd for classes). then we calculated distances from cortex of each svdd class. the result of classification is depicted in fig. . for other datasets (breast-cancer-wisconsin dataset, ionosphere dataset, wine dataset, transfusion dataset and heart dataset), the same process is followed and the result shows, figs. – , that svdd is good to be used as a tool for our distance based classification. . accuracy of our method in comparison with multi-class svm to show the performance of our method, we first transformed the data of each dataset to the proposed shell fitting space and we then used a simple linear classification (adaline) to classify the transformed data. in this stage, % of the data in each dataset was used for the training purpose. kernel-based svm classifier was used on the original data to compare the accuracy of our classifica- tion results. amongst the available kernels the one with the best performance for each dataset was used in the training stage. both systems were tested for each test data of each dataset. the compar- ison results of the mean values of runs of both methods follow. it is worth mentioning that in each run of each dataset the data are randomly divided into training class and test class. adaline classifier uses a w, weight vector, to separate classes with a hyper-plane where w is obtained from ( ): w ¼ dxðxx Þ� ð Þ in ( ) d is the vector of the desired label for each training data of the class of training data and x is the vector of training samples. we used rbf, linear, polynomial, . . . , kernels for svm classifier and se- lected the best results for the comparison purpose. accuracy (ac) is calculated for a dataset using the following definition: ac ¼ðnumber of true classificationsÞ =ðnumber of true and false classificationsÞ or, ac ¼ number of true positives þ true negatives number of true positives þ true negatives þ false positives þ false negatives ð Þ the comparison results are shown in table . as you can see, in our method only the hyper-sphere fitting to wine dataset and heart dataset has led to a lower accuracy. this shows that, if the fitting shape is not selected well, the result might not be desirable. . conclusions in this paper we introduced a new transformation space which is easier than the other space transformations to classify input fig. . result of svddd on ionosphere dataset. fig. . result of svddd on breast-cancer-wisconsin dataset. fig. . result of circle fitting on ionosphere dataset. fig. . result of circle fitting on iris dataset. fig. . result of svdd on iris dataset. fig. . (a) input dataset and (b) result of classification using distance based method. fig. . result of circle fitting on breast-cancer-wisconsin dataset. m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – fig. . result of svddd on wine dataset. fig. . result of svddd on ionosphere dataset. table comparison of our method accuracy with multi-class svm (mean accuracy in runs). mean accuracy in runs our method (circle fitting) (%) our method (using svdd) (%) multi-class svm (%) cancer . . . heart . . wine . . iris . . . ionosphere . . . m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – data. the main idea was to use distance of data to a line, curve, or hyperplane fitted to each class’ data or shells enclosing the classes. fig. . anfis a the results show that this space transformation works well for col- lections of data. appendix. adaptive neuro-fuzzy inference system (anfis) architecture (jang, ) a typical architecture of anfis is shown in fig. for modeling of function, in which a circle indicates a fixed node, and a square indicates an adaptive node. for simplicity, we consider two inputs x, y and one output z in the fuzzy inference system (fis). the anfis used in this paper implements a first-order sugeno fuzzy model. among the many fuzzy inference systems, the sugeno fuzzy model is the most widely used due to its high interpretability and compu- tational efficiency, and built-in optimal and adaptive techniques. for example for a first-order sugeno fuzzy model, a common rule set with two fuzzy if-then rules can be expressed as ( ) and ( ) rule : if x is a and y is b ; then z ¼ p x þ q y þ r ð Þ rule : if x is a and y is b ; then z ¼ p x þ q y þ r ð Þ where ai, bi (i = , ) are fuzzy sets in the antecedent, and pi, qi, ri (i = , ) are the design parameters that are determined during the training process. as in fig. , the anfis consists of five layers. layer , every node i in this layer is an adaptive node with a node function like ( ) o i ¼ laiðxÞ; i ¼ ; o i ¼ lbiðyÞ; i ¼ ; ð Þ where x, y are the input of node i, laiðxÞ and lbiðyÞ and can adopt any fuzzy membership function (mf). in this paper, gaussian mfs which are defined by ( ) are used gaussianðx; c; rÞ¼ e� x�c rð Þ ð Þ where c is center of gaussian membership function and r is stan- dard deviation of this cluster. layer , every node in the second layer represents the ring strength of a rule by multiplying the incoming signals and forward- ing the product as ( ) o i ¼ wi ¼ laiðxÞlbiðyÞ; i ¼ ð Þ layer , the ith node in this layer calculates the ratio of the ith rule’s ring strength to the sum of all rules’ rings strengths: o i ¼ -i ¼ xi x þ x ; i ¼ ; ð Þ where -i is referred to as the normalized ring strengths. rchitecture. m. ghazizadeh ahsaee et al. / expert systems with applications ( ) – layer , the node function in this layer is represented by ( ). o i ¼ ~xizi ¼ ~xiðpix þ qi y þ riÞ; i ¼ ; ð Þ where -i is the output of layer , {pi, qi, ri} and is the parameter set. parameters in this layer are referred to as the consequent parameters. layer , the single node in this layer computes the overall out- put as the summation of all incoming signals like ( ) o ¼ x i� -i zi ¼ x z þ x z x þ x ð Þ it is seen from the anfis architecture that when the values of the premise parameters are fixed, the overall output can be expressed as a linear combination of the consequent parameters: z ¼ð~x xÞp þð~x yÞq þð~x Þr þð~x xÞp ð~x yÞq þð~x Þr ð Þ the hybrid learning algorithm (jang, ) and (jang, sun, & mizu- tani, ) combining the least square method and the back prop- agation (bp) algorithm can be used to solve this problem. this algorithm converges much faster since it reduces the dimension of the search space of the bp algorithm. during the learning process, the premise parameters in layer and the consequent parameters in layer are tuned until the desired response of the fis is achieved. the hybrid learning algorithm has a two-step process. first, while holding the premise parameters fixed, the functional signals are propagated forward to layer , where the consequent parameters are identified by the least square method. second, the consequent parameters are held fixed while the error signals, the derivative of the error measure with respect to each node output, are propagated from the output end to the input end, and the premise parameters are updated by the standard bp algorithm. references abe, s. ( ). support vector machines for pattern classification. springer-verlag london limited. abe, s. ( ). advances in pattern recognition. springer-verlag london limited. issn: - . bishop, c. ( ). neural networks for pattern recognition. walton street, oxford ox dp: oxford university press. friedman, j. h. ( ). multivariate adaptive regression splines. annals of statistics, , – . huang, t., kecman, v., & kopriva, i. ( ). kernel based algorithms for mining huge data sets. berlin, heidelberg: springer-verlag. uc irvine machine learning repository. jang, j.-s. r. ( ). anfis: adaptive-network-based fuzzy inference system. ieee transactions on systems, man, and cybernetics, ( ), – . jang, j.-s. r., sun, c. t., & mizutani, e. ( ). neuro–fuzzy and soft computing: a computational approach to learning and machine intelligence. englewood cliffs, nj: prentice-hall. lin, c.-f., & wang, s.-d. ( ). fuzzy support vector machine. ieee transactions on neural networks, ( ), – . parzen, e. ( ). on estimation of a probability density function and mode. annals of mathematical statistics, , – . sadoghi yazdi, h., effati, s., & saberi, z. ( ). the probabilistic constraints in the support vector machine. applied mathematics and computation, ( ), – . scholkopf, b., & smola, a. j. ( ). learning with kernels. massachusetts institute of technology. shawe-taylor, j., & cristianini, n. ( ). kernel methods for pattern analysis. cambridge university press. tax, d. m. j. ( ). one-class classification: concept learning in the absence of counter- examples (vol. ). netherlands: technische universiteit delft. tax, d. m. j., & duin, r. p. w. ( ). support vector data description. machine learning (vol. ). kluwer academic publishers (pp. – ). vapnik, v. ( ). the nature of statistical learning theory. new york: springer- verlag. wang, l. (ed.). ( ). support vector machines: theory and applications. berlin heidelberg: springer-verlag. issn: - . wu, c., jie-chi, y., lee, y. ( ). an approximate approach for training polynomial kernel svms in linear time. in th annual meeting of the association for computational linguistics companion volume proceedings of the demo and poster sessions, june (pp. – ). ypma, r.d. ( ). support objects for domain approximation. in proceedings of international conference on artificial neural networks (icann ). skovde, sweden. http://archive.ics.uci.edu/ml/ shell fitting space for classification introduction shell fitting space formulation dealing with hyperplane dealing with a hypersphere dealing with a polynomial in two-dimensional space dealing with a line in n-dimensional space dealing with a curve in n-dimensional space dealing with a voluminous classes in n-dimensional space svdd one-class classification using svdd for our distance based method experimental results operation on synthetic data transformation of circle classes transformation of more complex classes transformation using anfis transformation using svdd experiments on datasets circle fitting on datasets using svdd on dataset accuracy of our method in comparison with multi-class svm conclusions adaptive neuro-fuzzy inference system (anfis) architecture (jang, ) references mining changes in customer behavior in retail marketing mining changes in customer behavior in retail marketing mu-chen chen a,*, ai-lun chiu b , hsu-hwa chang c a department of business management, national taipei university of technology, no. , section , chung-hsiao e. road, taipei , taiwan, roc b sas, taipei, taiwan, roc c department of business administration, national taipei college of business, taipei, taiwan, roc abstract during the past decade, there have been a variety of significant developments in data mining techniques. some of these developments are implemented in customized service to develop customer relationship. customized service is actually crucial in retail markets. marketing managers can develop long-term and pleasant relationships with customers if they can detect and predict changes in customer behavior. in the dynamic retail market, understanding changes in customer behavior can help managers to establish effective promotion campaigns. this study integrates customer behavioral variables, demographic variables, and transaction database to establish a method of mining changes in customer behavior. for mining change patterns, two extended measures of similarity and unexpectedness are designed to analyze the degree of resemblance between patterns at different time periods. the proposed approach for mining changes in customer behavior can assist managers in developing better marketing strategies. q elsevier ltd. all rights reserved. keywords: data mining; association rules; retailing; customer behavior; change patterns . introduction customized service is actually crucial in retail markets. customized service has become a key issue in developing customer relationships. marketing managers can develop long-term and pleasant relationships with customers if they can detect and predict changes in customer behavior. in the past, researchers generally applied statistical surveys to study customer behavior. recently, however, data mining techniques have been adopted to predict customer behavior (giudici & passerone, ; song, kim, & kim, ). data mining techniques search through a database without any specific pre-determined hypothesis to obtain implicit, previously unknown, and potentially useful information including knowledge rules, constraints and regularities (chen, han, & yu, ). data mining is a stage in knowledge discovery in databases (kdd), involving the application of specific algorithms for pattern extraction (mitra, pal, & mitra, ). various successful applications have been reported in areas such as marketing, finance and banking. applications in these domains generally involve the collection and storage of large amounts of data. data mining brings various techniques together to discover patterns (rules) and to construct models from databases. currently businesses face the challenge of a constantly evolving market where customer needs are changing all the time. in such a situation, change mining can enable market analysts to better understand changes in customer needs and how those needs change. change mining is more appropriate in dynamic business environments, and involves extensive human intervention (song et al., ). retail market managers must not only provide high- quality products and services, but also must react appro- priately to changes in customer needs. data mining can be applied to identify useful customer behavior patterns from large amounts of customer and transaction data (giudici & passerone, ). as a result, the discovered information can be ascertained to support better decision-making in retail marketing. data mining techniques have mostly been adopted to generate predictions and describe behaviors. relatively little research has focused on mining changes in databases collected over time (liu, hsu, han, & xia, ). - /$ - see front matter q elsevier ltd. all rights reserved. doi: . /j.eswa. . . expert systems with applications ( ) – www.elsevier.com/locate/eswa * corresponding author. tel.: c x ; fax: c . e-mail address: bmcchen@ntut.edu.tw (m.-c. chen). http://www.elsevier.com/locate/eswa in the dynamic retail market, understanding changes in customer behavior can help managers to establish effective promotion campaigns (song et al., ). liu et al. ( ) devised a method of change mining in the context of decision trees for predicting changes in customer behavior. since decision tree is a classification-based approach, it cannot detect complete sets of changes (song et al., ). association rule extraction was widely used for analyzing the correlation between product items purchased by customers, and to support sales promotion and market segmentation (changchien & lu, ; changchien, lee, & hsu, ). song et al. ( ) employed an approach based on association rules to identify changes in customer behavior. most previous customer behavior studies applied custo- mer demographic variables to analyze customer behavior (song et al., ). however, valuable customer behavioral variables, such as recency, frequency, and monetary (rfm), can be used to differentiate customer contributions to a business (miglautsch, ; stone, ; suh, noh, & suh, ; tsai & chiu, ). researchers have observed that rfm is a widely used technique for customer behavioral analysis that can effectively investigate customer values and segment markets. customer behavioral variables, rfm, respectively, measure the recency of customer purchasing behavior, the frequency of purchasing, and the average monetary expenditure on purchasing. rfm can be trans- formed using customer and transaction databases. this study attempts to integrate customer behavioral variables (rfm), demographic variables, and transaction database to establish a method of mining changes in customer behavior. song et al. ( ) designed two measures of similarity and unexpectedness to analyze the degree of resemblance between patterns at different time periods. however, these two measures are limited to the analysis of patterns with a single attribute on the right-hand- side (consequent part) of an association rule. this study designs two modified measures of similarity and unexpect- edness to overcome the above limitations. in this study, the data required for analysis are integrated and transformed from customer, product, and transaction databases. the proposed approach for mining changes in customer behavior can assist managers in developing better marketing strategies. . mining customer behavior changes in this study, customer behavior patterns are first identified using association rule mining. following the association rules of customer behavior are discovered, the changes in customer behavior are identified by comparing two sets of association rules generated from two datasets of different periods. based on previous studies, changes in customer behavior include emerging patterns, added patterns, perished patterns, and unexpected patterns (dong & li, ; liu & hsu, ; liu, hsu, mun, & lee, ; padmanabhan & tuzhilin, ; song et al., ). the discovered change patterns can be further explained and assessed to provide a basis for formulating marketing strategies. fig. illustrates the flowchart of change mining for customer behavior. further details of the change mining procedure are discussed below. . . data pre-processing prior to analysis, data accuracy and consistency must be ensured to obtain truthful results. generally, some useful variables can be hidden in a large quantity of raw data, and thus can be obtained through data integration and transformation. customer behavioral variables (rfm) are hidden in customer and transaction databases, and can be extracted from data integration and transform- ation. since the data required for analyzing association rules must be discrete, continuous variables are trans- formed to discrete variables, and a simple - - rule is applied (han & kamber, ). in rfm, recency represents the interval between the most recent transaction time of individual customers and the evaluation time (stone, ). moreover, frequency rep- resents the number of purchases by individual customers during a specific period. additionally, monetary represents the average expenditure of a customer during a specific period. individual customers’ recency, frequency, and monetary are scored to calculate the value of the purchasing behavior of each customer. this study adopts the rfm scoring approach of miglautsch ( ) to transform the customer behavioral variables. data preprocessing and customer behavioral & demographic data transformation mining customer behavior mining change patterns: • emerging • added • perished • unexpected supporting marketing customer product transaction fig. . flowchart of mining changes for customer behavior. m.-c. chen et al. / expert systems with applications ( ) – https://isiarticles.com/article/ local-shapelets for fast classification of spectrographic measurements local-shapelets for fast classification of spectrographic measurements daniel gordona,c,∗, danny hendlera, aryeh kontorovicha, lior rokachb,c adepartment of computer science, ben-gurion university of the negev be’er sheva , israel bdepartment of information systems engineering, ben-gurion university of the negev be’er sheva , israel ctelekom innovation laboratories, ben-gurion university of the negev be’er sheva , israel abstract spectroscopy is widely used in the food industry as a time-efficient alternative to chemical test- ing. lightning-monitoring systems also employ spectroscopic measurements. the latter appli- cation is important as it can help predict the occurrence of severe storms, such as tornadoes. the shapelet based classification method is particularly well-suited for spectroscopic data sets. this technique for classifying time series extracts patterns unique to each class. a signif- icant downside of this approach is the time required to build the classification tree. in addition, for high throughput applications the classification time of long time series is inhibitive. although some progress has been made in terms of reducing the time complexity of building shapelet based models, the problem of reducing classification time has remained an open challenge. we address this challenge by introducing local-shapelets. this variant of the shapelet method restricts the search for a match between shapelets and time series to the vicinity of the location from which each shapelet was extracted. this significantly reduces the time required to examine each shapelet during both the learning and classification phases. classification based on local- shapelets is well-suited for spectroscopic data sets as these are typically very tightly aligned. our experimental results on such data sets demonstrate that the new approach reduces learning and classification time by two orders of magnitude while retaining the accuracy of regular (non- local) shapelets-based classification. in addition, we provide some theoretical justification for local-shapelets. keywords: spectrography, time series, classification, shapelets, local research highlights • we present an algorithm for classifying spectrographic measurements. • the concept of locality is introduced into an established time series algorithm. • a technique for estimating a tolerance parameter is presented. ∗corresponding author. tel: + ( ) ; fax: + ( ) ; email addresses: gordonda@cs.bgu.ac.il (daniel gordon), hendlerd@cs.bgu.ac.il (danny hendler), karyeh@cs.bgu.ac.il (aryeh kontorovich), liorrk@bgu.ac.il (lior rokach) preprint submitted to expert systems with applications november , • learning and classification times are reduced by two orders of magnitude. • accuracy levels are retained. . introduction spectroscopy is a field devoted to the study and characterization of physical systems by measuring the electromagnetic frequencies they absorb or emit (herrmann and onkelinx, ). items differing in their chemical composition or molecular bonds absorb or emit light at different wavelengths leaving a different spectroscopic fingerprint thus enabling differentiation between them. for example, al-jowder et al. ( ) used mid-infrared spectroscopy to detect meat adul- teration by comparing the spectra of adulterated meat with that of unadulterated meat. a study by briandet et al. ( ) discriminated between two different types of coffee beans (arabica and robusta) using mid-infrared spectroscopy. other methods for distinguishing between different types of food exist, which are based on wet chemical analysis (bicchi et al., ; lumley, ; sharma et al., ). the advantages of spectroscopy over wet chemical analysis are in its sim- plicity (briandet et al., ) and speed of response. spectroscopic measurements are also generated by systems monitoring lightning (eads et al., ). this application is important as relative percentages of different types of lightning can indicate the outbreak of severe storms, such as tornadoes. in addition to laboratory research, spectroscopic equipment is starting to be mass produced for every day use allowing anyone to analyze their surroundings with the aid of spectroscopic measurements (scio, ). the measurements are uploaded to a cloud service where they are analyzed and then the results of the analysis are made available. the service is cloud based, requiring algorithms with high throughput to enable a quick response to high volumes of queries by users. the outcome of the spectroscopic analysis of a physical system is a vector in which each index represents a frequency and each value is the measured intensity of that frequency. the representation of spectroscopic measures and time series are identical (ye and keogh, a), as the only explicit data are the measurements and the meaning of each measurement is defined by its location in the vector. this equivalence allows the application of time series classifica- tion methods to the field of spectroscopy. a previous experimental study (hills et al., ) showed that the shapelet based classification method is particularly suited for data sets from the field of spectroscopy, as it achieved a higher accuracy than other machine-learning classification methods. recently, ye and keogh ( a) introduced the shapelets approach for classifying time se- ries. a shapelet is a subsequence extracted from one of the time series in the data set. the shapelet is chosen by its ability to distinguish between time series from different classes. a test time series is classified based on its distance from the shapelet. in the case of multiple shapelets, these form the nodes of a classification tree. the intuition behind this approach is that the pattern best separating the classes is not necessarily an entire time series. rather, a certain subsequence may best describe the discriminating pattern. ye and keogh’s algorithm considers all possible subsequences in the training set in order to identify those shapelets that yield the optimal split. through the rest of this paper, we will refer to this algorithm as the yk-algorithm. two key advantages of classification with shapelets are the accuracy and interpretability of the induced classification model, as it supplies information on the patterns characteristic of the different classes (ye and keogh, a). a significant downside of this approach is the time re- quired for building the classification tree (hills et al., ; mueen et al., ; rakthanmanon and keogh, time series index a b so rb a n ce (a) coffee data set time series index a b so rb a n ce (b) wheat data set figure : examples of two data sets (coffee and wheat) from the field of spectroscopy. the time series of each class are vertically separated and in different colors. as shown, the examples of each class are tightly aligned, i.e., similar patterns are exhibited at similar locations along the x-axis. ). the search for the best shapelet requires examining all subsequences of all lengths from all the time series in the training set, and for each shapelet calculating its distance to each time series at all possible locations. even for small data sets, this process has a time scale of days, and for large data sets, the time scale becomes one of years. hence, the original implementation on commonly available hardware is only practical on the smallest of data sets. additionally, for high-throughput applications, the classification time may be prohibitively expensive for long time series. this is because at each node of the tree, all possible matches between the node’s shapelet and the time series to be classified need to be examined. . . our contributions our goal was to reduce both learning and classification time without impairing the accuracy of the induced model by exploiting a feature common to spectroscopic data – the localized nature of information in the time series. in the yk-algorithm, no importance is attributed to the location from which the shapelet was extracted. hence, the best match between a shapelet and a time series is searched for anywhere in the time series. we observed that for many data sets from the field of spectroscopy, time series from the same class show similar behavior patterns at similar locations along the frequency axis. fig. presents examples of two data sets which strongly support this insight. based on this insight, we propose a new property as part of the definition of a shapelet, derived from the location in the time series from which the shapelet was extracted. this property limits the scope of the search for the best match of a shapelet to a time series to the vicinity of the location from which the shapelet was extracted. the assumption of locality is justified as spectroscopic measurements of items with similar properties should have very similar spectroscopic fingerprints, especially in areas characteristic of a specimen which are not expected to be contaminated. our current implementation assumes that all time series are of equal length. although the time series are generally aligned, some allowance for misalignment is neces- sary. we therefore introduce a method for learning the misalignment characteristic of a data set. we evaluate our approach on data sets from the field of spectroscopy, and show that local- shapelets can reduce learning and classification time (especially for data sets with long time series) by over two orders of magnitude without impairing accuracy. for reproducibility, we have made all our code available online (local shapelets, ). the rest of the article is organized as follows: first we present basic definitions required for understanding the article and shortly describe the yk-algorithm in sect. . then we present related work (sect. ), followed by a description of local-shapelets and our proposed method for determining the range to examine (sect. ). next we present our experimental evaluation (sect. ) followed by a brief statistical analysis, which provides a theoretical justification for our local-shapelets approach (sect. ). finally, we summarize our results and present additional research directions to pursue (sect. ). . background here we present a number of definitions necessary for the proper understanding of this article and a short description of the original shapelet algorithm as it is the basis of our work. . . definitions definition . a time series t of length m is a series of m consecutive equally spaced measure- ments: t = t , t , ..., tm− . definition . a subsequence s of length k extracted from time series t of length m at index i such that k ≤ m is a series of k consecutive measurements: s = ti, ti+ , ..., ti+k− . definition . the euclidean distance between two time series t,r of length m is: de (t,r) = √√√ m− ∑ i= (ti − ri) . definition . the euclidean distance between a time series t of length m and a subsequence s of length k such that k ≤ m is: de (t,s ) = min i de (t [i : i + k − ],s ) i∈[ ,m−k]. this is the minimal distance between s and all subsequences of length k in t . definition . given a data set d, with c classes and n examples, each class i with ni examples, the entropy is: ent(d) =− c− ∑ i= ni n log ni n . intuitively, a data set’s entropy is a measure of its class-homogeneity. larger homogeneity corresponds to smaller entropy values. specifically, the smallest entropy value ( ) corresponds to a data set in which all members belong to the same class. definition . given a data set d with n examples, split into two subsets d and d , containing n and n examples respectively, such that d ∪ d = d and d ∩ d = ∅ the information gain (ig) is: ig(d, d , d ) = ent(d) − (n n ent(d ) + n n ent(d ) ) . intuitively, information gain is a measure of the class-homogeneity induced by a split of data set d. the larger the information gain, the larger the decrease in entropy of the split data set w.r.t. d, in turn implying better class-homogeneity. as we will soon see, each shapelet induces a data set split and its effectiveness is measured by the split’s information gain. . . the yk-algorithm for completeness, we briefly present the yk-algorithm as presented in ye and keogh ( a). first, we present the algorithm for two classes; we then extend the description to a multi-class data set. let d be a dataset with two classes and n time series. the yk-algorithm examines all possi- ble subsequences of every length (from a minimal length, usually , to a maximal length which is usually the length of the shortest time series) from every time series. for each subsequence s , the distance to each time series is calculated, as defined in definition . then, the time series are ordered by their distance from s . using this induced order, the average distance of every two adjacent time series to s is calculated. we will refer to this average distance as the splitting distance. each of the n splitting distances defines two subsets, one containing all time series with a distance to s smaller than or equal to the splitting distance, and the other containing all time series with a distance to s greater than the splitting distance. for every possible split into two subsets, the information gain is calculated (see definition ). if the current information gain is better than the best so far, the shapelet is kept along with the corresponding splitting distance. tie breaking is done by keeping the shapelet which induces a larger average distance between the two subsets which is referred to as the margin. after checking all possible subsequences, the best shapelet and the corresponding splitting distance are returned. this method can be easily extended to a multi-class problem by building a tree, with a shapelet and splitting distance in each node. a new node receives one of the two subsets created by the shapelet found by the node above it, and learns the best shapelet and splitting distance for this subset of time series. the stopping criteria for this recursive algorithm is that all the time series in the subset be of one class. two important implementation issues are that all distance calculations are computed after local normalization (goldin and kanellakis, ) and that the margin is normalized, by dividing it by the length of the subsequence. classification of a time series t is accomplished by traveling down the tree. at each node the distance of the shapelet s associated with the node to t is calculated. the node decides to which of its child nodes t should be directed, depending on whether its distance from s is smaller or greater than the splitting distance. when t reaches a leaf, it is assigned the class associated with this leaf. . . . time complexity of the yk-algorithm let m denote the length of a time series and let us assume all time series are of equal length. assuming all shapelet lengths from to m are examined, the number of different shapelets to examine in a single time series is ∑m i= i = o(m ). let n denote the number of time series in data set d. the number of shapelets to examine in the entire data set is o(nm ). when searching for the minimal distance between a shapelet s of length k and a time series t , the distance of s to all subsequences of length k in t needs to be calculated (see definition ). the time complexity of this operation is o(m ). calculating the distance of s to all time series requires o(nm ) calculations. the total number of calculations for all shapelets and time series is o(n m ) which explains the formidable time required for learning a model even for small data sets. . related work the time complexity of the yk-algorithm is formidable (see sect. . . ) leading to a large number of attempts to reduce it. as we will show in this section, none of the previous approaches utilized the location from which the shapelet was extracted to reduce the time required to learn a model. in addition, most of these approaches do not reduce the time required to classify a time series. the first attempt to reduce the time complexity of the yk-algorithm was introduced in the paper first presenting shapelets (ye and keogh, a). two optimizations were suggested. the first optimizes the distance calculation of a shapelet to a time series. the distance calculation is terminated if it exceeds the minimum distance found so far between the current shapelet and time series. this optimization was coined early-abandon. the second optimization (named entropy- pruning) checks whether the most optimistic ig possible, given the distances of a shapelet to time series already computed, can be better than the best ig found so far. if the ig cannot be improved, the shapelet under examination is discarded. as pointed out by lines et al. ( ) this optimization requires testing o( c) different possibilities (c is the number of classes in the data set), which can greatly reduce the effectiveness of this optimization when the number of classes is large. later, mueen et al. ( ) introduced additional optimizations. one optimization manages to compute the distance of a shapelet to a time series in constant time by precomputing necessary statistics. this optimization manages to reduce the time complexity to o(n m ). a major down- side is that for each two time series, a matrix of size m needs to be maintained. this leads to a total space complexity of o((nm) ) which is untenable for large data sets (gordon et al., ; rakthanmanon and keogh, ). a second optimization discards shapelets similar to shapelets that were already discarded. a disadvantage of this approach is the large time overhead when applied to data sets with a large number of classes. two recent attempts managed to dramatically reduce the time complexity for learning a shapelet based model. the first method (rakthanmanon and keogh, ) quickly picks out a small number of shapelets from each shapelet length which seem able to effectively divide the data set into its classes. then only this subset of shapelets are fully analyzed. this approach manages to reduce the time complexity to o(nm ) but requires a considerable amount of space to accommodate this reduction in time complexity. we will refer to this solution as the hashing- algorithm. a second approach (gordon et al., ), named salsa-r, randomly samples a constant number of shapelets ( , ) which are examined, reducing the time complexity to o( , × nm ) with no excess space requirements. as shown, none of the aforementioned optimizations utilize the location from which the shapelet was extracted to reduce the time complexity of the learning process. in addition none of them significantly reduces classification time. xing et al. ( ) introduced an optimization for fast classification of streaming time series with the aid of shapelets. their main idea was to prefer shapelets which appear early in a time series over shapelets which appear later. this ensures that time series can be classified quickly once initial measurements have arrived with no need to wait for further measurements. they coined this new type of shapelets as local-shapelets. although we both name our shapelets similarly, the concepts differ. the motivation of xing et al. was to classify streaming time series as early on as possible while ours is to optimize learning and classification time of non-streaming data sets. also, the implementations differ. xing et al. did not preserve the location from which the shapelet was extracted. conversely, with our method, the location from which the shapelet was extracted is exploited with no limitation on the location from which to extract a shapelet. . local-shapelets in the material that follows, the basic idea of local-shapelets is described as well as modifi- cations which make it useful in practice. definition . a shapelet is a tuple <~s ,d>. ~s is a series of consecutive measurements extracted from one of the time series in the data set and d is a cutoff distance. time series with a distance to s smaller or equal to d traverse one side of the tree while time series with a distance to s greater than d traverse the other side of the tree. definition . a local-shapelet is a tuple <~s ,d, i>. ~s and d are as in definition . i is the location from which the local-shapelet was extracted. unlike a shapelet, a local-shapelet contains information regarding the location from which it was extracted, which is utilized when calculating the distance of the local-shapelet to a time series. definition . the euclidean distance between a time series t of length m and a local-shapelet <~s ,d, i> of length k such that k ≤ m given x which defines a range around i is: de (t,s ) = min j de (t [ j : j + k − ],s ) j∈[max( ,i−x),min(i+x,m−k)]. the distance between t and a local-shapelet <~s ,d, i> adds a constraint on the subsequences of t to which the distance of subsequence ~s is calculated. instead of calculating the distance of ~s to all subsequences of length k in t , the distance of ~s is calculated only to subsequences in the vicinity of the location i from which ~s was extracted. this vicinity is defined by the constant x. thus, the time required for calculating the distance of a shapelet to a time series is reduced. . . the tolerance range a naı̈ve implementation of local-shapelets is to calculate the distance of a shapelet only to the single subsequence in the time series at the exact location i from which it was extracted. this approach may have detrimental impact on the learning process as exemplified in fig. which presents two time series from the same class of the data set lightning . as is clearly shown, the characteristic spike may appear at slightly different locations. restricting the distance calculation time series index a b so rb a n ce (a) example from class time series index a b so rb a n ce (b) example from class figure : two examples from different classes of the lightning data set illustrating the need for a tolerance range. without a tolerance range the spike characteristic of each class cannot be utilized as its location is different in different time series. of a shapelet to the exact location from which it was extracted would cause characteristic patterns to be overlooked. to accommodate this issue some tolerance, which we will refer to as the radius, needs to be added to the index i, such that the distance of a local-shapelet to a time series t is the minimum distance to all subsequences starting in the range [i − radius, i + radius] (see definition ). we will refer to this range as the tolerance range. in food spectroscopy there are many sources of noise, such as differences in the residual water content of the freeze-dried samples (briandet et al., ). this noise may cause distortions in the pattern generated but will not cause a shift in the frequencies emitted, therefore we set the tolerance range to . in lightning spectroscopy, interferences such as frequency-dependent dispersion induced by the ionosphere (moore et al., ) may lead to a shift in the frequencies measured requiring the introduction of a tolerance range greater than . we present a method for computing the value of the tolerance required, as described in pro- cedure . this is achieved by experimentally observing the tolerance required for a small number of subsequences. procedure splits each time series into ten equal consecutive and disjoint sub- sequences (line ). from each class (c) and for each subsequence location, a single subsequence (subseq) is extracted from a randomly selected time series (rand-ts) (lines - ). then the distance of subseq to all time series in c is calculated using the global method for calculating distances (see definition ). the location of the best match of subseq to each time series is recorded in locations (line ). before utilizing the information available in the list of locations, it is necessary to filter out values which are obviously non-characteristic of the data set (line ). a major motivation is that a larger tolerance leads to an increase in learning and classification time as more distance calculations are required for each shapelet. filtering out locations which are atypical should not impair accuracy significantly. for example, let us suppose we receive the following list of locations: , , , , , , , , , , , . it is quite clear that the values , are outliers procedure algorithm for computing the tolerance characteristic of a data set compute-tolerance(d) {input is a data set} : start-indices ← indices, s.t. time series will be split into subsequences which are as equal in length as possible : for each class c in d do : for each start-index i in start-indices do : subseq-length ← length of subsequence : rand-ts ← randomly selected time series from c : subseq ← rand-ts[i : i+subseq-length- ] : for each time series t in class c do : locations ← append(location-of-best-distance(t, subseq)) : locations ← outlier-filter(locations) {using iqr-filtering} : radiusc,i ← max(locations) - min(locations) : radiusc ← min(radiusc,i) : tolerance ← max(radiusc) : return tolerance and should be filtered out. without filtering, the size of the range is , while after filtering, the size of the range is only . our method for outlier filtering is based on the interquartile range (iqr). using this method, the first (q ) and third (q ) quartiles are calculated and the iqr is calculated as i qr = q − q . all values greater than q + × i qr or smaller than q − × i qr are filtered out. although it is customary to use . as the multiplication factor, we chose a multiplication factor of so as to not filter out too many values which may lead to overfitting. our tuning of the multiplication factor to a value of is a heuristic as it cannot guarantee total avoidance of overfitting, because other parameters such as the model complexity also play a part. three advantages of iqr filtering are that it is simple, that it is a-parametric and that it does not automatically drop extreme values if they are similar to the rest. once we have filtered out the outliers, the characteristic radius as reflected by this subse- quence of class c is calculated and recorded (line ). after a radius for each of the subse- quences of a class has been calculated, the minimum of all these radii is selected to represent the locality of the class (line ). we chose the minimal radius as this leads to the smallest number of distance calculations of a shapelet to a time series, which promises the best possible speedup in runtime. the last stage is the selection of a single radius as the tolerance for the data set. we chose the maximum radius from all classes (line ) as the tolerance must accommodate the most loosely localized class. otherwise, for some of the classes, the ultimate matches may reside outside of the recommended range and will not be examined. the time complexity of this phase is negligible as only subsequences per class are exam- ined and the distance of each subsequence is calculated only to time series of its class. . . random selection of shapelets for completeness, we present the procedure used by salsa-r for randomly selecting shapelets in procedure . we chose a distribution similar to the uniform distribution but simpler to imple- ment. procedure describes the method for randomly selecting the next shapelet to examine. first (line ), the time series from which to extract the shapelet is chosen. then (lines - ), the procedure algorithm for randomly selecting shapelets extract-shapelet(d,min-sh-length) {input is a data set and the minimum length of a shapelet} : num-ts ← number-of-time-series-in-data-set(d) : ts-index ← random-selection( ,num-ts) : highest-index ← times-series-length - min-sh-length + : sh-index ← random-selection( ,highest-index) : longest-possible-shapelet ← times-series-length - sh-index + : sh-length ← random-selection(min-sh-length,longest-possible-shapelet) : sh ← extract-shapelet(d,ts-index,sh-index,sh-length) : return sh index in the time series from which to extract the shapelet is randomly generated. the range of possible indices is between and the last index from which the shortest possible shapelet can be extracted. last (lines - ), the length of the shapelet is randomly selected. the upper limit on the length of the shapelet (longest-possible-shapelet) is calculated based on the location from which the shapelet is to be extracted (line ). the function random-selection(a,b) randomly selects an integer from the range [a,b- ] with a uniform distribution. . experimental results our goal is to show that for data sets from the field of spectroscopy, the usage of local- shapelets reduces the time complexity during both training and classification phases in com- parison with global-shapelets (i.e., non-local shapelets) without degrading accuracy. first, we present the data sets from the field of spectroscopy with which we evaluated local-shapelets. then, we establish the utility of our method (procedure ) for calculating the tolerance range. in the next phase of our evaluation we compare local-shapelets vs. global-shapelets within the yk- algorithm. in the last phase, we re-implement salsa-r to utilize the locality of shapelets. we compare accuracy and run-time with the original implementation of salsa-r and the hashing- algorithm. we ran all experiments on an intel xeon e computer comprising two . ghz processors with gb of ram and with a -bit version of windows server r as the operating system. . . description of data sets our experiments were conducted on a collection of data sets from the field of spectroscopy, available online (ye and keogh, b; keogh et al., ). the collection contains four data sets from the field of food spectroscopy (beef, coffee, oliveoil, wheat) and two from the field of lightning spectroscopy (lightning , lightning ). table contains information on the number of examples in the training and test sets, the number of classes, the length of the time series and the tolerance range used. all data sets were already split into train and test sets. we preserved the original division to train and test sets to allow easy reproduction of our results, as well as a fair comparison with other published results. as the test sets are very small, our initial measurements of classification times were inaccu- rate due to minor overheads which dampened the effect of locality on the outcome and due to the inaccuracy of computer time measurements at small time scales. we solved this by enlarging each test set to a size of gb. this was done by duplicating examples. to ensure that the ac- curacy obtained would be identical to that on the original data set, we duplicated each example an equal number of times. as the classifier is deterministic, the classification of an example will always be identical no matter how many times it appears. therefore, the proportion of correct classifications out of all classification examples will not change and the accuracy will remain the same. table : description of the data sets dataset train set test set num. time series tolerance size size classes length beef , coffee , lighting , lighting , oliveoil , wheat , , . . . food spectrographs beef. the beef data set (al-jowder et al., ) contains the spectral absorbance of one type of beef cut (silverside). one class contains the spectral absorbance of the beef cut without any contaminates. each of the other four classes contains the spectral absorbance of the beef cut contaminated with a different type of offal (kidney, liver, heart and tripe) which is cheaper than the declared beef cut. coffee. the coffee data set (briandet et al., ) contains the spectral absorbance of instant coffee of two different types of coffee beans arabica and robusta. coffee from arabica beans is more highly estimated as it has a finer and more pronounced taste. approximately % of world coffee production is from arabica and another % is from robusta. as the price of arabica is higher than that of robusta, it is important to be able to distinguish between them even after the long process that is required to produce instant coffee. oliveoil. olive oil samples from four different european countries (greece, italy, portugal and spain) were collected (tapp et al., ). the classification task is to be able to discern the country of origin using the spectrograph of the olive oil sample. wheat. this data set (ye and keogh, a) consists of spectrographs of wheat grown during the years - . there are a number of different types of wheat in the data set but the class was assigned based only on the year in which the wheat was grown and not on the type of wheat. . . . lightning spectrographs data on frequencies emitted during lightning events were collected and then a fourier trans- form was applied to produce spectrographs (eads et al., ). the lightning events were cat- egorized into different classes differing in the charge of the lightning (positive or negative), whether the event was gradual or abrupt and whether the event was intra-cloud or from cloud to ground. the original authors of this data set (eads et al., ) note that there is a large inter- class variation and intra-class similarity. the data set lightning contains examples of all classes, while lightning is a simpler binary problem of distinguishing between cloud to ground events and intra-cloud events. . . utility of tolerance range calculation the last column of table presents the tolerance range used during our experiments. for the four data sets of food spectroscopy, the value of the tolerance range was not calculated using procedure ; rather we set it to based on prior knowledge in this domain. a comparison of the tolerance range based on prior knowledge with those recommended by procedure shows large agreement. for three data sets (coffee, oliveoil and wheat) values are identical. for the beef data set, although the calculated tolerance range was not exactly , it was very close with a value of . our procedure also succeeds when a tolerance range greater than is required. when applied to the lightning data set for which it is clear a tolerance range larger than is required, as shown in fig. , the calculated tolerance range is . these findings show that our method for predicting the tolerance range manages to successfully estimate the required range. . . local yk-algorithm here we show that local-shapelets reduce the time consumption without impairing accuracy. we compare local and global shapelets within the yk-algorithm framework which is the initial implementation of the shapelet algorithm which examines all possible shapelets. due to the large time and space requirements of the yk-algorithm we could not collect results for the lightning data set. for these experiments, we used the original code used by ye and keogh ( a). as pointed out by hills et al. ( ), the entropy-pruning optimization (see sec. ) has an overhead which grows exponentially with the increase in the number of classes in the data set. we encountered this experimentally with the wheat data set which has classes. the original implementation did not finish examining all shapelets for the first node of the tree after days, while the same implementation without the entropy-pruning optimization finished learning the whole tree in this period of time. therefore we conducted the experiments using the original code without the entropy-pruning optimization. results of our experiments are presented in table . each two columns compare results of global-shapelets vs. local-shapelets. the first comparison is the accuracy achieved, the second is the time required to learn a model and the third is the time required to classify the test set. in all measures, the local approach outperforms the global approach. the average improvement in accuracy is %, the average speedup during the learning phase is . and during the classification phase it is . table : global yk-algorithm vs. local yk-algorithm data set accuracy (%) learning time (sec) classification time (sec) global local global local global local beef . . , , . . coffee . . , . . lighting . . , , . . oliveoil . . , , . . wheat . . , , . . . . local-salsa-r in this set of experiments we re-implemented salsa-r to use local-shapelets instead of global-shapelets. the number of shapelets randomly selected was set to , which was found to be an optimal number by gordon et al. ( ). we compared local-salsa-r with global- salsa-r and the hashing-algorithm. we repeated our experiments thirty times as each of the three methods includes an element of randomness. table presents a comparison of the average accuracy of local-salsa-r with global- salsa-r and the hashing-algorithm. we applied a friedman test (friedman, ) to test if there is a significant difference between accuracy attained by the different methods. the p-value was . , which clearly shows that there is no significant difference in the accuracy achieved by any of the methods. table : comparison of accuracy of local-salsa-r with that of global-salsa-r and the hashing-algorithm data set global-salsa-r hashing-algorithm local-salsa-r beef . . . coffee . . . lighting . . . lighting . . . oliveoil . . . wheat . . . a comparison of average learning and classification times is presented in table . a fried- man test on the learning and classification times shows that there is a significant difference be- tween the methods during both the learning (p-value = . ) and classification phases (p-value = . ). a one sided wilcoxon-test (wilcoxon, ) affirms our claim that local-shapelets are significantly faster than global-shapelets during the learning phase (p-value = . for both global-salsa-r and the hashing-algorithm) and the classification phase (p-value = . for global-salsa-r and p-value = . for the hashing-algorithm). on average, local-salsa-r reduces the time required to learn a model by a factor of and in comparison with global- salsa-r and the hashing-algorithm, respectively. during classification, local-salsa-r is times faster than global-salsa-r and times faster than the hashing-algorithm, on average. table : comparison of learning and classification time of local-salsa-r with that of global-salsa-r and the hashing-algorithm data set learning time (sec) classification time (sec) global- hashing- local- global- hashing- local- salsa-r algorithm salsa-r salsa-r algorithm salsa-r beef . . . . . . coffee . . . . . . lighting . . . . . . lighting . . . . . . oliveoil . . . . . . wheat . . . . . . analytically, it is easy to argue that the longer the time series, the greater the time saved by using local-shapelets, as the number of distance calculations avoided increases. we confirmed this experimentally as can be seen in fig. . the figure shows the ratio between the time required by global-salsa-r and local-salsa-r as a function of the length of the time series for all data sets for which the tolerance range used was (all data sets apart from lightning ). this figure clearly confirms that the ratio increases with the length of the time series. we chose to compare local-salsa-r with global-salsa-r and not with the hashing-algorithm as their implementation apart from the aspect of locality is identical, allowing isolation of locality as the only parameter influencing the outcome. time series length t im e s p e e d u p learning time classification time figure : each point is the ratio of the time required by global-salsa-r and local-salsa-r. one plot is for learning times and the second for classification times. as can be seen the ratio increases with the length of the time series. . statistical analysis the approach proposed in this paper was mainly motivated by algorithmic considerations: restricting the search to a small subset of the possible shapelet locations significantly speeds up both training and classification. in this section, we will argue that as a by-product, our approach offers statistical advantages as well. by restricting the number of features, we are constraining the complexity of the hypothesis class. as we show below, hypothesis classes of low complexity require fewer training examples to attain a certain accuracy level. the argument is made precise in the language of learning theory. a learner faced with a classification task is trying to learn a function g : x → {− , }, where x is the instance space (in our case, it is the set of all possible time series). the learner gets to observe example-label pairs (xi,yi) ∈ x× {− , } generated iid from some unknown distribution p over x× {− , }. this corresponds to the intuition that the training labels may be noisy, and indeed, there may be no “correct” classifier g : x → {− , } that achieves perfect accuracy. although there are universal approximators capable of fitting arbitrary labeled samples, if unconstrained they will necessarily overfit (devroye et al., ). hence, when choosing the learning model, its richness (i.e., hypothesis complexity) must be taken into account. the learner’s n observed labeled examples (xi,yi) constitute the training set, based on which it will produce a hypothesis g : x→{− , }. we will denote by h the collection of all admissible hypotheses (formally, h ⊂ x) and associate with every h ∈ h two key quantities: its sample (or training) error, êrr(h) = n n∑ i= {h(xi ),yi} and generalization error, err(h) = e[ {h(xi ),yi}] = p(h(x) , y ). in words, êrr(h) is the relative fraction of mistakes that h makes on the training set while err(h) is the probability that h makes a mistake on a freshly drawn (x,y ) pair — crucially, drawn from the same distribution used to generate the training set. note that while the typical goal is to guarantee a small generalization error, the latter quantity cannot be computed without knowledge of the sampling distribution. instead, the readily computable êrr(h) may be used (under certain conditions, detailed below) as a proxy for err(h). in this setting, the learner’s task is twofold: (i) algorithmic: efficiently find an h ∈ h for which êrr(h) is small, and (ii) statistical: guarantee that, with high probability, err(h) will not be much greater than êrr(h), regardless of which h ∈h the learner chooses. the foregoing sections were devoted to the algorithmic aspects, and we shall focus on the statistical one here. in the case of finite h, a particularly simple connection exists between err(h) and êrr(h): theorem (mohri et al. ( )). suppose that |h|<∞ and the learner observes a training set consisting of n examples. then, for any δ > , we have that err(h) ≤ êrr(h) + √ log |h|+ log( /δ) n ( ) holds with probability at least −δ, uniformly over all h ∈h. for simplicity, let us consider the case where h = x (i.e., h consists of all possible binary functions). in this case, |h| = |x|, and hence even a modest reduction of the instance space — by reducing the feature set, for example — can have a noticeable effect on the second term in the right-hand side of ( ), and hence yield a faster convergence rate. the basic features used in this paper are distances from a shapelet to a time series. it is precisely this feature set that gets reduced when our algorithm considers only a subset of the possible locations. this observation provides a statistical justification to our local-shapelets approach, in addition to the algorithmic speedup. . conclusions the objective of our investigation was to utilize the localization of characteristic patterns in spectrographic measurements using the shapelet algorithm, which has previously been found to be suited (hills et al., ) for this domain. our adaption to the shapelet algorithm reduces the number of distance calculations and thus shortens the time required to train a model and classify examples. as pointed out by ye and keogh ( a), one important advantage of the shapelet approach is its interpretability, i.e., the process extracts informative subsequences representative of each class. the chosen shapelets provide insights into patterns characteristic of each class. one such example can be seen in fig. . the shapelet, shown as a dashed red line, is overlaid on each of the two time series at the location from which it was extracted and represents a pattern characteristic of class . in addition to identifying the discriminative pattern, local-shapelets also identify the discriminative frequencies. we compared the frequencies found to be discriminative by the shapelet with those found to be discriminative by briandet et al. ( ) and found that they coincide. time series index a b so rb a n ce class class figure : interpretability of local-shapelets. two time series from the coffee data set are presented in black. each time series is from a different class. only part of each time series is presented so as to focus on the important details. overlaid on each of the time series is the shapelet selected by local-salsa-r as most discriminative appearing as a dashed red line. as can be seen the shapelet represents time series from class (robusta coffee beans). the algorithm we presented searches for matches of a shapelet on a time series only in the vicinity of the location from which the shapelet was originally extracted. we proposed an algo- rithm (procedure ) for calculating the exact vicinity based on properties of each class in the data set. our main result is that local-shapelets can indeed speedup both the learning and classifica- tion methods -fold when using salsa-r without impairing accuracy. we also show that our estimation of the vicinity to examine is quite accurate. this research can be extended in many ways. it may be interesting to explore possible trade- offs between speedup and accuracy as a function of the vicinity to examine. another research direction is the application of local-shapelets to domains other than spectroscopy which may also require the adaption of our algorithm for time series of different lengths. references r. herrmann, c. onkelinx, quantities and units in clinical chemistry: nebulizer and flame properties in flame emission and absorption spectrometry (recommendations ), pure and applied chemistry ( ) ( ) – . o. al-jowder, e. k. kemsley, r. h. wilson, detection of adulteration in cooked meat products by mid-infrared spec- troscopy, journal of agricultural and food chemistry ( ) ( ) – . r. briandet, e. k. kemsley, r. h. wilson, discrimination of arabica and robusta in instant coffee by fourier transform infrared spectroscopy and chemometrics, journal of agricultural and food chemistry ( ) ( ) – . c. p. bicchi, a. e. binello, m. m. legovich, g. m. pellegrino, a. c. vanni, characterization of roasted coffee by s- hsgc and hplc-uv and principal component analysis, journal of agricultural and food chemistry ( ) ( ) – . i. lumley, authenticity of meat and meat products, in: food authentication, springer, – , . n. sharma, a. srivastava, j. gill, d. joshi, differentiation of meat from food animals by enzyme assay, food control ( ) ( ) – . d. r. eads, d. hill, s. davis, s. j. perkins, j. ma, r. b. porter, j. p. theiler, genetic algorithms and support vector machines for time series classification, in: international symposium on optical science and technology, international society for optics and photonics, – , . scio, http://www.consumerphysics.com/myscio/, . l. ye, e. keogh, time series shapelets: a novel technique that allows accurate, interpretable and fast classification, data mining and knowledge discovery ( a) – . j. hills, j. lines, e. baranauskas, j. mapp, a. bagnall, classification of time series by shapelet transformation, data mining and knowledge discovery ( ) – . a. mueen, e. keogh, n. young, logical-shapelets: an expressive primitive for time series classification, in: proceedings of the th acm sigkdd international conference on knowledge discovery and data mining, acm, – , . t. rakthanmanon, e. keogh, fast shapelets: a scalable algorithm for discovering time series shapelets, in: proceed- ings of the thirteenth siam conference on data mining (sdm), siam, – , . local shapelets, ftp://www.ise.bgu.ac.il/, . d. goldin, p. kanellakis, on similarity queries for time-series data: constraint specification and implementation, in: principles and practice of constraint programming cp’ , springer, – , . j. lines, l. m. davis, j. hills, a. bagnall, a shapelet transform for time series classification, in: proceedings of the th acm sigkdd international conference on knowledge discovery and data mining, acm, – , . d. gordon, d. hendler, l. rokach, (in press) fast and space-efficient shapelets-based time-series classification, intelligent data analysis ( ). z. xing, j. pei, s. y. philip, k. wang, extracting interpretable features for early classification on time series., in: eleventh siam international conference on data mining (sdm), siam, – , . k. r. moore, p. c. blain, s. d. briles, r. g. jones, classification of rf transients in space using digital signal processing and neural network techniques, in: spie’s symposium on oe/aerospace sensing and dual use photonics, international society for optics and photonics, – , . l. ye, e. keogh, shapelet data sets, http://alumni.cs.ucr.edu/~lexiangy/shapelet.html, b. e. keogh, q. zhu, b. hu, h. y., x. xi, l. wei, c. a. ratanamahatana, the ucr time series classification/clustering homepage, www.cs.ucr.edu/~eamonn/time_series_data/, . h. s. tapp, m. defernez, e. k. kemsley, ftir spectroscopy and multivariate analysis can distinguish the geographic origin of extra virgin olive oils, journal of agricultural and food chemistry ( ) ( ) – . m. friedman, the use of ranks to avoid the assumption of normality implicit in the analysis of variance, journal of the american statistical association ( ) ( ) – . f. wilcoxon, individual comparisons by ranking methods, biometrics bulletin ( ) ( ) – . l. devroye, l. györfi, g. lugosi, a probabilistic theory of pattern recognition, vol. of applications of mathematics (new york), springer-verlag, new york, isbn - - - , . m. mohri, a. rostamizadeh, a. talwalkar, foundations of machine learning, mit press, . http://www.consumerphysics.com/myscio/ ftp://www.ise.bgu.ac.il/ http://alumni.cs.ucr.edu/~lexiangy/shapelet.html www.cs.ucr.edu/~eamonn/time_series_data/ introduction our contributions background definitions the yk-algorithm time complexity of the yk-algorithm related work local-shapelets the tolerance range random selection of shapelets experimental results description of data sets food spectrographs lightning spectrographs utility of tolerance range calculation local yk-algorithm local-salsa-r statistical analysis conclusions microsoft word - pinter _ _ --- calibrating anns by go.docx calibrating artificial neural networks by global optimization jános d. pintér , : faculty of engineering, Özyeğin university, kusbakisi caddesi, �o. , istanbul, turkey; www.ozyegin.edu.tr; janos.pinter@ozyegin.edu.tr. : pintér consulting services, inc., canada; www.pinterconsulting.com; janos.d.pinter@gmail.com. technical report, Özyeğin university, istanbul submitted for publication: july abstract an artificial neural network (ann) is a computational model − implemented as a computer program − that is aimed at emulating the key features and operations of biological neural networks. anns are extensively used to model unknown or unspecified functional relationships between the input and output of a “black box” system. in order to apply such a generic procedure to actual decision problems, a key requirement is ann training to minimize the discrepancy between modeled and measured system output. in this work, we consider ann training as a (potentially) multi-modal optimization problem. to address this issue, we introduce a global optimization (go) framework and corresponding go software. the practical viability of the go based approach is illustrated by finding close numerical approximations of (one-dimensional, but non-trivial) functions. keywords artificial neural networks, ann model calibration by global optimization, lipschitz global optimizer (lgo) solver suite, ann implementation in mathematica, mathoptimizer professional (lgo for mathematica), illustrative numerical examples. . introduction a neuron is a single nerve cell consisting of a nucleus (cell body, the cell’s life support center), incoming dendrites with endings called synapses that receive stimuli (electrical input signals) from other nerve cells, and an axon (nerve fiber) with terminals that deliver output signals to other neurons, muscles or glands. the electrical signal traveling down on the axon is called neural impulse: this signal is communicated to the synapses of some other neurons. this way, a single neuron acts like a signal processing unit; the entire neural network has a complex, massive multi-processing architecture. neural systems are essential for an organism, in order to adapt properly and quickly to its environment. an artificial neural network (ann) is a computational model − implemented as a computer program − that is aimed at emulating the essential features and operations of neural networks. anns are used extensively to model complex relationships between input and output data sequences, or to find hidden patterns in data sets when the analytical (explicit model function based) treatment of the problem is tedious or currently not possible. anns replace these unknown functional relationships by adaptively constructed approximating functions (approximators). such approximators are typically designed on the basis of training examples of a given task, such as recognising hand- written words. for in-depth expositions related to the ann paradigm consult e.g. hertz et al., ( ), cichocki and unbehauen ( ), smith ( ), golden ( ), ripley ( ), mehrotra et al. ( ), bar-yam ( ), steeb ( ), cruse ( ). the mathematical model of an ann is based on the key features of a neuron as outlined above. we shall consider first a single artificial neuron (an), and assume that a finite sequence of input signals s=(s , s , …, st) is received by its synapses: the signals st t= ,…t may be real numbers or vectors. the an’s information processing capability will be modeled by introducing an aggregator (transfer) function a that depends on a certain parameter vector x=(x , x , …, xn). subsequently, x will be chosen according to a chosen optimality criterion. the optimization is typically carried out in such a way that the sequence of model output signals m=(m , m ,…, mt) generated by the function a=a(x, s) is as close as possible to the corresponding set of observations o=(o , o ,…, ot). this generic description can be summarized symbolically as follows. s → an → m s=(s , s , …, st), mt=a(x, st), t= ,…t. ( ) minimize d(m, o) m=(m , m ,…, mt), o=(o , o ,…, ot), d is a discrepancy measure. a single an modeled as above is called a perceptron. flexible generalizations of a single perceptron based model are multiple perceptrons receiving the same input signals: such a structure is called an artificial neuron layer (anl). to extend this structure further, multilayered networks can also be defined: in these several anls are connected sequentially. it is also possible to model several functional relationships simultaneously by the corresponding output sequences of a multi-response ann. figure illustrates a multilayered feed-forward ann that includes an input layer with multiple inputs, an intermediate (hidden) signal processing layer, and an output layer that emits multiple responses. the hidden layer’s an units receive the input information, followed by certain decomposition, extraction, and transformation steps to generate the output information. hence, the ann implementation versions used in practice all share the features of adaptive, distributed, local, and parallel processing. figure schematic representation of a multilayered ann source: http://en.wikipedia.org/wiki/artificial_neural_network following the generic description ( ), the optimal parameterization strategy is summarized as follows. given the functional form of a and the sequence s of input values, we want to find a parameterization x such that the difference between the output set m=a(x, s) and the observation set o is minimal. the discrepancy d(m, o) is typically expressed by a suitable norm function: d(m, o)=||m-o||. once the parameter vector x becomes known − i.e. it has been estimated as a result of using the training i/o sequences s and o − the trained ann can be used to perform various tasks related to the assessment of further input data such as {st} for t=t+ , t+ , and so on. some important application examples of this paradigm are signal processing, data classification, and time series forecasting; further application areas will be mentioned later on. . function approximation by a��s: theoretical background and a corresponding optimization model formulation the key theoretical feature of anns is their ability to approximate (unknown) continuous functions by learning from observed data, under general analytical conditions. specifically, the approximation theorem of cybenko ( ) states that a feed-forward ann with a single hidden layer that contains a finite number of neurons equipped with a suitable transfer function can serve as a universal approximator of any given continuous real-valued function f: rk→r defined on [ , ]k. the approximation is uniform, in terms of the supremum norm of functions defined on [ , ]k. we will discuss this result below in some detail, and then introduce a corresponding optimization model. for a somewhat more complete picture of the theoretical foundations of anns, let us remark that funahashi ( ) and hornik et al. ( ) proved similar results to that of cybenko. hornik et al. ( ) proved that feed-forward neural networks with a single hidden layer of suitable transfer function units can approximate any borel measurable multivariate vector function f:rk →rk uniformly, to arbitrary prescribed accuracy on an arbitrary compact set c⊂rk . for more detailed technical discussions of these and closely related approximation results, consult also the foundations laid out by kolmogorov ( ) which, however, do not lead directly to the results cited here since the approximating formula depends on the function f. more recent discussions of related function approximation results are presented by park and sandberg ( , ), lang ( ), sutskever and hinton ( ), with further topical references. let us consider a single input signal st from r k k≥ , and the corresponding scalar output of ot = f(st) for all t= ,..,t: here f is an unknown continuous model function that we wish to approximate by a suitably defined ann. we remark that in our general model the input signals can be delivered to a single (synapse of an) an, or to several synapses of several ans at the same time: the corresponding information will be then aggregated by the ann. cybenko ( ) establishes the theoretical possibility of generating a parameterized approximating function a(x, st), using a fixed type of transfer function σ such that the approximation shown below is uniformly valid. f(st)~a(x, st):=∑ j= ,…,j αj σ(∑ i= ,…,i yji sti + θj). ( ) to relate this general result directly to an underlying ann model, in ( ) j= ,…,j denotes the index of the individual signal processing units anj in the hidden layer. for each unit anj, σ is a given function form to be specified below, αj, yj, and θj are model parameters to be determined depending on the unknown function f. specifically, σ is a continuous real valued transfer function (in principle, σ could also be a function of j); ∑ i= ,…,i yji sti is the scalar product of the synaptic weight vector yj of anj and the signal st=(st ,…, sti) component arriving from synapse i of anj, for i = ,…,i; θj is called the threshold level (offset, bias) of anj; and αj is the weight of anj in the function approximation ( ) for j= ,…,j. following cybenko ( ) and hornik et al. ( ), the transfer function σ is assumed to be monotonically non-decreasing, and it satisfies the relations lim σ(z)= as z→-∞, and lim σ(z)= as z→∞. ( ) transfer functions with these properties are called sigmoidal or squashing functions. let us remark here that other types of transfer functions can also be used in generating function approximations. suitable sets of radial basis functions (rbf) are another well-known example (park and sandberg, , ; röpke et al., ). rbfs have the general form φ(z)=φ(||z−−−−z ||): here ||.|| is a given norm function, z is the (scalar or vector) centre of the monotonically non-increasing function φ that satisfies the conditions φ( )= , and lim φ(z)= as ||z−−−−z ||→∞. we will continue the present discussion using a specific sigmoidal transfer function form. an often used function form is σ(z)= /( +e-z) z∈r. ( ) for proper theoretical and better numerical tractability, we assume and define finite box (variable range) constraints to bound the values of the parameters αj, yji and θj: αjmin≤αj≤αjmax for j= ,…,j. yjimin≤yji≤yjimax for i = ,…,i and j= ,…,j. θjmin≤θj≤θjmax for j= ,…,j. ( ) the entire set of model parameters will be denoted by x=({αj}, {yji}, {θj}) where the appropriate components are included for all i= ,…,i and j= ,…,j. the optimization problem induced for a given approximator function form is to minimize the error of the approximation ( ) for the entire training sequence (st, ot) t= ,…,t, using a selected norm function. in this study, we shall use the least squares error criterion to measure the quality of the approximation: in other words, the objective function ( ) shown below will be minimized under the box constraints ( ). e(x)=∑ t= ,…,t (f(st) − a(x, st)) ( ) the notation e(x) refers to the fact that ( ) is an error function. let us point out here that other (more general linear or nonlinear) constraints regarding feasible parameter combinations may also be present if dictated by the problem context. such situations can also be handled by our model development approach, by simply adding these constraints to the optimization model formulation. an important case in point is to attempt the exclusion of identical (symmetric) solutions to in which the component functions of the function a are in principle exchangeable. to break this symmetry, one can assume e.g. that the components of {αj} are ordered so that αj≥αj+ for all j= ,…,j- . ( ) one could also add further lexicographic constraints if dictated by the problem at hand. another possible − but not always applicable − option could be to normalize the weights {αj}so that we have ∑ j= ,…,j αj= , αj≥ for all j= ,…,j. ( ) . postulating and calibrating a�� model instances cybenko ( ) emphasizes that his cited result only guarantees a uniformly valid approximation of a continuous function f, as opposed to its exact representation. we should also keep in mind that the general approximation result expressed by ( ) does not specify the actual number of processing units: hence, the key model parameter j needs to be estimated for each problem instance separately. cybenko ( ) conjectures that j could become “astronomical” in many function approximation problems ‒ especially so in higher dimensional approximations. these points should be kept in mind, to avoid making unjustifiably optimistic claims. once we select (test or postulate) the ann structure to use in an actual case study, the remaining basic requirement is its application-specific parameterization with respect to a given training data set. the objective function in ( ) makes this a nonlinear model fitting problem that could have a multitude of local or global optima. this multimodality issue is discussed in detail e.g. by pintér ( , ) with further references; in the context of anns, consult e.g. auer et al. ( ), bianchini and gori ( ), abraham ( ). for a more complete picture, let us remark that detailed expository discussions related to global optimization models, algorithms and applications are presented e.g. in the topical handbooks edited by horst and pardalos ( ), pardalos and romeijn ( ). let us also observe here that following the selection of the function form σ, the number of real-valued parameters (decision variables to optimize) is j(i+ ), each with corresponding box constraints; furthermore, the number of optionally added linear constraints − considering ( ) and ( ) − is j. since in the general ann model setting the values of i and j cannot be theoretically specified a priori, in practical applications one can try several computationally affordable combinations. in general, anns with more components could serve as better approximators, within reason and in line with the size of the training data set that should at least satisfy t≥j(i+ ). at the same time, in order to assure the success of a suitable nonlinear optimization procedure (and a corresponding software), one should restrict i and j to manageable values. this choice will then influence the accuracy of the approximation obtained. summarizing the preceding discussion, first one has to postulate an ann structure, and the family of parameterized transfer functions to use. next, one needs to provide a set of training (input-output) data. the training is then aimed at the selection of an optimally parameterized function from the chosen family of function models, to minimize the error function. the iterative training of anns ‒ based on local optimization techniques ‒ is often referred to in the topical literature as back-propagation. due to the possible multimodality of the error function ( ), the simplest “standard” tools of unconstrained local nonlinear optimization (direct search, gradient based methods, newton-type methods) often fail to locate the best parameter combination. this issue is well-known for ann researchers: consult e.g. bianchini and gori ( ), sexton et al. ( ), abraham ( ), hamm et al. ( ). to overcome this difficulty, the various optimization strategies used in practice to train anns include experimental design, sampling by low-discrepancy sequences, theoretically exact global or local scope search approaches, as well as popular heuristics such as evolutionary optimization methods, particle swarm optimization, simulated annealing, tabu search and tunneling functions. ann model parameterization frameworks and numerical studies are presented and discussed e.g. by watkin ( ), prechelt ( ), bianchini and gori ( ), sexton et al. ( ), jordanov and brown ( ), toh ( ), ye and lin ( ), abraham ( ), hamm et al. ( ). next we will discuss a theoretically exact global optimization based approach and a corresponding software implementation that will be used subsequently to solve several illustrative ann calibration problems. similarly to most other researchers, we will study “only” the problem of estimating the parameters of given anns with a given (postulated) architecture. the problem of finding the most appropriate model form in full generality seems elusive, and it is certainly difficult − if not impossible − to address in actual numerical studies. . global optimization: a general modeling framework to match the modeling approach and notations introduced earlier, we shall use the following symbols and assumptions: x∈rn n-dimensional real-valued vector of decision variables e:rn→r continuous (scalar-valued) objective function d⊂rn non-empty set of feasible solutions, a proper subset of rn: the feasible set d is closed and bounded, defined by l∈rn, u∈rn component-wise finite lower and upper bounds on x, and (optionally) by g:rn→rm an m-vector of continuous (more general) constraint functions. we shall then consider the continuous global optimization (go) model min e(x) subject to x∈d:={x: l≤x≤u gj(x)≤ j= ,...,m}. ( ) without going into details which are not necessary for the present discussion, let us remark that the terse model formulation ( ) covers many special cases. specifically, it also subsumes the ann model calibration model statement summarized by formulas ( ) to ( ), with the optionally added constraints ( ) and ( ). additionally, the go problem statement ( ) also guarantees that the set of global solutions is non-empty, and its structure supports the application of (theoretically) globally convergent deterministic and stochastic optimization algorithms. for technical details, consult e.g. pintér ( ). needless to say, the numerical handling of go model instances of ( ) can be a tough challenge: even small-dimensional model instances can be hard to handle, and they essentially require an appropriate global scope search strategy. these general notes apply also to many ann model instances as it will be illustrated shortly (in section ). consequently, the calibration of anns ‒ as a rule ‒ requires proper go software: one such software product will be briefly described next. . the lgo software package for constrained global-local optimization let us reiterate the key “black box” characteristic of the ann modeling framework: namely, that its error function is evaluated by a numerical procedure that − in a model development and testing exercise − may be subject to changes and modifications. similar situations that require (also) global scope search often arise in the real world of optimization: their solution requires easy-to-use, robust and efficient solver technology. within the broad go software category, direct (derivative-free) global search algorithm implementations have the advantage of immediate applicability to “black box” problems. next, we shall briefly review the key features of the lipschitz global optimizer (lgo) solver suite that has been designed to directly address also “black box” problems. developed since the late ’s, lgo is currently available for a range of compiler platforms (c, c++, c#, fortran / / ), with seamless links to optimization modeling languages (aimms, ampl, gams, mpl), ms excel spreadsheets, and to the leading high-level technical computing systems maple, mathematica, and matlab. for theoretical, algorithmic and implementation details not discussed here, we refer e.g. to pintér ( , , , , , , a, b), pintér and kampas ( ), pintér et al. ( ). the overall design of lgo is based on the combination of continuous nonlinear optimization strategies, with corresponding theoretical global and local convergence properties. hence, lgo can be used for both constrained global and local optimization. let us point out here that lgo’s overall derivative-free design is different from many other nonlinear optimization software packages that require explicit analytical model information to support model parsing and function type specific operations. of course, we do not claim that one design is “always superior” to the other. a model function decomposition-parsing-bounding based approach supports the precise solution of a range of go models, assuming sufficient runtime and computational resources. all models, however, have to be defined in terms of a given set of possible component functions. in contrast to this approach, the design of lgo allows the handling of genuine “black box” problems that will remain outside of the scope of the analytical global optimization approaches referred to above. for optimization model examples with embedded computational procedures, consult e.g. pintér ( , ), pintér et al. ( ), kampas and pintér ( , ). in numerical practice (that is in hardware resource-limited and time-limited runs), lgo’s global search options generate a global solution estimate(s) that is (are) refined by the seamlessly following local search mode(s). this way, the expected result of using lgo is a global and local search based high-quality feasible solution that meets at least the local optimality criteria. (to guarantee theoretical local optimality, standard local smoothness conditions need to apply.) at the same time, one should keep in mind that no global ‒ or, in fact, any other ‒ optimization software will “always” work satisfactorily, with default settings and under resource limitations related to model size, time, model function evaluation, or to other preset usage limits. extensive numerical tests and an increasing range of widely different practical applications demonstrate that lgo and its platform-specific implementations can find high-quality numerical solutions not only when using academic go test problems, but also in often far more complicated, sizeable go models. examples of real-world optimization challenges handled by various lgo implementations are environmental systems modeling and management (pintér, ), laser design (isenor et al., ), intensity modulated cancer therapy planning (tervo et al., ), the operation of integrated oil and gas production systems (mason et al., ), vehicle component design (goossens et al., ), various packing problems with industrial relevance (castillo et al., ), fuel processing technology development (pantoleontos et al., ), and currency trading strategies (Çağlayan and pintér, ). . �umerical examples and discussion the detailed testing and comparative assessment of optimization software products is a significant research area. one needs to follow or to establish objective evaluation criteria, and to present (in principle) fully reproducible results. the software tests themselves are based on a given set of test problems and a selection of criteria such as the reliability and efficiency of the solvers used, in terms of the quality of solutions obtained and the corresponding computational effort. for examples of benchmarking studies, in the ann context we refer to prechelt ( ), hamm et al. ( ); in the go context, see e.g. ali et al. ( ), khompatraporn et al. ( ), pintér ( , ); with further topical references therein. here we will illustrate the performance of global optimization software (specifically, of an lgo implementation) in the ann calibration context, by solving merely one- dimensional, but non-trivial function approximation problems. the key features of the computational environment used and the numerical tests are summarized below: the complete set of detailed numerical results is available from the author upon request. • hardware: pc, intel™core™ duo cpu p @ . ghz, . gb ram. • operating system: windows xp pro with sp ( -bit). • global optimization software: mathoptimizer professional (mop). let us remark that mathoptimizer professional is the software product name introduced for the lgo solver implementation that is linked to mathematica (wolfram, ). for details, consult e.g. pintér and kampas ( ), kampas and pintér ( ). • ann computational models used: -j- type networks that have a single input node, a single output node, and j processing units. recall that this setup directly follows the function approximation structure expressed by ( ). the ann models considered are implemented in native mathematica code. • ann initialization by assigned starting parameter values is not necessary, since mop (i.e., lgo) uses a genuine global scope optimization approach. let us note at the same time that all lgo implementations also support the optional use of set initial values in their constrained local optimization mode, without applying first a global search mode. • a single run is done in all cases; no repetitions are needed, since mop produces identical runs under the same initializations and solver options. repeated runs with possibly changing numerical results can be easily generated by setting several options of mop. the latter are based on selecting one of the several global search operational modes, using different starting points in the local search mode, using different random seeds, and assigning different computational resources to mop runs. • in all test problems the optimized function approximations are based on a set of training points, in line with the modeling framework of section . • simulated noise structure: none. in other words, we assume that all training data are exact. this assumption could be easily relaxed, by introducing a probabilistic noise structure as done e.g. in pintér ( ). • key performance measure: standard error (the root of the mean square error, rmse) calculated for the numerical solution found. • some additional mathematica features used: documentation and model visualization capabilities. • some additional mop features used here (or elsewhere): automatic generation of optimization model code in c or fortran; and the automatic generation of input and result text files. these files are used and generated by the lgo solver core of mop. in the numerical examples presented here, our objective is to “learn” the graph of certain one-dimensional functions defined on the interval [ , ] based on a preset number of uniformly spaced training points. as earlier, we shall use the notation t for the number of training data, and j for the number of processing units in the hidden layer. example our first (easiest) test problem presented here has been studied also by toh ( ), as well as by ye and lin ( ). f(x)=sin( πx) cos( πx) ≤x≤ . ( ) applying t= , j= , and mop’s local search mode (ls), we obtain a corresponding function approximation characterized by rmse ~ . . with default mop settings, the number of error function evaluations is , ; the corresponding runtime is . seconds, implying an estimated , function evaluations per second. (let us remark that the speed estimate can vary a bit, based on the actual status of the computer used.) for comparison, the function ( ) is displayed below (see the left hand side of figure ), directly followed by list plot (right hand side) produced by the output of the trained ann at the uniformly distributed t= sample points. figure function ( ), and the list plot produced by the optimized ann. both the numerical result (the rmse obtained) and figure indicate a good fit. for a more complete picture, we also solved the same problem using lgo’s most extensive global search mode (stochastic multi-start, ms) followed by corresponding local searches from the best points found. this operational mode is referred to as msls, and it has been used in all subsequent numerical examples. in the msls mode, we assume (postulate) the variable ranges [lb, ub] = [- , ] for all ann model parameters, and set a limit for the global search steps (error function evaluations) as gss= . based on our go related numerical experience, this does not seem to be an exorbitant setting since we have to optimize parameters globally. (for a simple comparison, one can think of aiming at just % precision for each variable setting in dimensions, using a uniform grid based search: this naïve strategy would require function evaluations.) as a result of using the msls operational mode, we receive rmse ~ . which represents about . % reduction of the average error found using only ls. the corresponding runtime is . seconds. it is also worth pointing out that the optimized ann parameterizations found by ls and msls are quite different. several components of these solutions lie on the boundary of the preset interval range [- , ] indicating a possible need for extending the search region. these findings, together with our other runs for this and other test problems indicate the potentially massive multi-modality of the objective ( ) in similar or more complicated function approximation problems as well as the possible need to search on larger parameter regions. in this particular test example, we could have perhaps reduced the global search effort, or just use the local search based estimate which seems “good enough”. such experimental “streamlining” may not always work, however, and we are fully convinced that global scope search is typically needed in calibrating anns. hence, we will continue to demonstrate mop’s standard global optimization capabilities, by solving all further test problems in msls mode. the examples presented below are chosen to be increasingly more difficult, although all are “only” approximation problems related to one-dimensional functions. obviously, one could “fabricate” functions that become very difficult to match by anns. . . . . . - . - . . . - . - . . . example f(x)=sin( πx) sin( πx) sin( πx) ≤x≤ . ( ) to handle this problem, we use t= , j= , and mop’s msls search mode. notice that the number of processing units has been increased heuristically, reflecting our belief that this function could be more difficult to reproduce than ( ). this leads to a - variable go model: we keep the allocated global search effort the same (gss= ) as in example , however. we conjecture again that the induced go model is highly multi-extremal. the global search based numerical solution has an rmse ~ . ; the corresponding runtime is . seconds ( , , function evaluations, i.e. ~ , evaluations per second). for comparison, the function ( ) is displayed below, directly followed by the list plot produced by the trained ann, see figure . again, the reproduction of the target function seems to be quite satisfactory, both numerically and visually. figure function ( ), and the list plot produced by the optimized ann. example f(x)=sin( πx) sin( πx) sin( πx) log( +x) ≤x≤ . ( ) in ( ), log denotes the natural logarithm function. as above, first we use t= , j= , and mop’s msls search mode with global search steps; furthermore, we extend the search region to [- , ] for each parameter. the global search based numerical solution has an rmse ~ . which is noticeably not as good as the previous rmses; the corresponding runtime is . seconds. for comparison, the function ( ) is displayed below, directly followed by the list plot produced by the trained ann. in the corresponding figure , one can notice a more apparent discrepancy in the “low-amplitude” section of the function, estimated by simple inspection as the interval [ , . ], while the approximation is quite satisfactory in the remaining “higher-amplitude” interval [ . , ]. . . . . . - . - . - . . . . - . - . - . . . . figure function ( ), and the list plot produced by a suboptimal ann. for comparison, we also ran mop in its (only) ls mode; the resulting standard error is rmse ~ . ; the runtime is . seconds. figure , corresponding to this case, visibly becomes a much more crude approximation of ( ) than the list plot in figure . again, this finding illustrates the typical need for a global scope model fitting approach. figure function ( ), and the list plot produced by the locally optimized ann. example in several subsequent runs, we increased the required precision of the approximation to function ( ) by the ann, by increasing one or several of the following: the number of training data t, the number of processing units j, and the solver runtime allocated to mop. obviously, all the above measures could lead to improved results. without going into unnecessary further details, figure (as an example) illustrates that e.g. the combination t= , j= , gss= , , , and the additionally set runtime limit of seconds allows the approximation of the entire function rather faithfully. this finding is also indicated by the corresponding rmse ~ . . let us remark that this mop run was automatically interrupted at the set time limit: arguably, without this limit even better approximations could be found. . . . . . - . - . . . . . - . - . . . . . . . . . . - . - . . . . . - . - . . . . . figure function ( ), and the list plot produced by the globally optimized ann. to summarize our illustrative numerical experiments, we can say that the go based ann calibration approach can be used to handle non-trivial function approximation problems within reasonable computational effort (on today’s personal computers). let us recall that ‒ in light of the theoretical results cited earlier ‒ exact function representations are in general too much to ask for. our approach is directly applicable to solve at least certain types of multi-input and multi-response approximation problems. if the input sequences need to be aggregated then this will lead to increase go model dimensionality. however, if the input data sequences can be processed by independently operated hidden layers for each function output, then our approach can be applied iteratively (repetitively) to such multi-response problems. hence, in case of fully decoupled ann models, the corresponding sub- problems can be solved independently, thereby avoiding the (expected) exponential increase of computational complexity. notwithstanding the above, one should be careful to avoid overly optimistic blanket statements when handling general multiple-input multiple-output problems by anns. . conclusions due to its “universal approximator” capability, the ann computational structure flexibly supports the modeling of many important research problems and real world applications that can be described by continuous functions defined over compact sets. ann application areas includes data classification and pattern recognition (ripley, ), damage detection and earthquake simulation (pei et al., ), function approximation (toh, ; ye and lin, ), material science (bhadeshia, ), experimental design of engineering systems (röpke et al., ), nonlinear optimization (malek et al., ), polypeptide structure prediction (dorn and de souza, ), prediction of trading signals of stock market indices (tilakaratne et al., ), regression analysis (de veux et al., ), signal and image processing (watkin, ; masters, ), time series analysis and forecasting (franses and van dijk, ; kajitani et al., ). at the same time, the ann computational model itself has an obvious meta-heuristic flavor in the sense that it needs to be instantiated and calibrated for each new model type. even if such problem-specific ann design can benefit from the expertise of domain specialists, manual design can become truly difficult when problem complexity increases. . . . . . - . - . . . . . - . - . . . . . furthermore, once the ann structure is selected, the computational effort may still need adjustments to handle the application satisfactorily. these practical aspects require flexible and transparent ann implementations. our current research code set up in an interpreted computing environment (mathematica) is an illustrative example of such an implementation. after defining the function to be approximated, one needs to change only a few input data to produce a new sampling plan, to define the ann instance (within a postulated ann family), and to activate the optimization engine. the corresponding results can be then directly observed and visualized (after a few seconds or minutes in the examples presented here, using an average capability pc). a focal point of the present study has been to demonstrate the essential requirement of high-quality ann model calibration. to meet this challenge, we apply a global optimization (go) based model fitting strategy. our approach is based on theoretically sound go methodology, and the corresponding software implementations are capable of handling non-trivial function approximation problems in reasonable time. the presented illustrative examples directly indicate the applicability of the go based approach e.g. to pattern or symbol recognition, image or signal processing, and time series analysis. in accordance with the theoretically exponential complexity of go, the numerical demand of model fitting exercises can be expected to increase considerably with the size of the induced optimization model. this specifically includes ann fitting problems in higher dimensions, and the consideration of general (possibly complicated) constraints regarding feasible parameter settings. we plan to investigate these issues in forthcoming studies, in relation to practically motivated problems that can be modeled by anns. acknowledgements the author wishes to acknowledge the valuable contributions dr. frank kampas to the development of the mathoptimizer professional software product. thanks are due also to dr. maria battarra for her comments on the manuscript. references abraham, a. ( ) meta learning evolutionary artificial neural networks. �eurocomputing , - . ali, m.m., khompatraporn, c. and zabinsky, z.b. ( ) a numerical evaluation of several stochastic algorithms on selected continuous global optimization test problems. journal of global optimization , – . auer, p. herbster, m. warmuth, m. ( ) exponentially many local minima for single neurons. in: touretzky, d. et al., (eds.), advances in �eural information processing systems, vol. , pp. – . mit press, cambridge, ma. bar-yam, y. ( ). dynamics of complex systems. westview press, boulder, co. bhadeshia, h.k.d.h. ( ) neural networks in materials science, isij international, ( ), - . bianchini, m. and gori, m. ( ). optimal learning in artificial neural networks: a review of theoretical results. �eurocomputing , - . Çağlayan, m.o. and pintér, j.d. ( ) development and calibration of currency market strategies by global optimization. research report, Özyeğin university, istanbul. (submitted for publication) castillo, i., kampas, f.j., and pintér, j.d. ( ) solving circle packing problems by global optimization: numerical results and industrial applications. european journal of operational research , – . cichocki, a. and unbehauen, r. ( ) �eural �etworks for optimization and signal processing. john wiley & sons, chichester, uk. cruse, h. ( ) �eural �etworks as cybernetic systems ( nd and revised edition). brains, minds and media, bielefeld, germany. cybenko, g. ( ) approximation by superpositions of a sigmoidal function. mathematics of control, signals, and systems , - . de veux, r.d., schumi, j., schweinsberg, j., and ungar, l.h. ( ) prediction intervals for neural networks via nonlinear regression. technometrics , – . dorn, m., and de souza, o. n. ( ) a n: an artificial neural network n-gram-based method to approximate -d polypeptides structure prediction. expert systems with applications (forthcoming), doi: . /j.eswa. . . . franses, p.h., van dijk, d. ( ) �onlinear time series models in empirical finance. cambridge university press, cambridge, uk. funahashi, k. ( ) on the approximate realization of continuous mappings by neural networks. �eural �etworks , - . golden, r.m. ( ) mathematical methods for �eural �etwork analysis and design, mit press, cambridge, ma. goossens, p., mcphee, j., schmitke, c., pintér, j.d., and stahl, h. ( ) driving innovation: how mathematical modeling and optimization increase efficiency and productivity in vehicle design. technical memorandum, maplesoft, waterloo, on. hamm, l., brorsen, b.w. and hagan, m. t. ( ) comparison of stochastic global optimization methods to estimate neural network weights. �eural processing letters , – . hertz, j., krogh, a., and palmer, r.g. ( ) introduction to the theory of �eural computation. addison- wesley, reading, ma. hornik, k., stinchcombe, m., and white, h. ( ) multilayer feedforward networks are universal approximators. �eural �etworks , – . horst, r. and pardalos, p.m., eds. ( ) handbook of global optimization, volume . kluwer academic publishers, dordrecht. isenor, g., pintér, j.d. and cada, m. ( ) a global optimization approach to laser design. optimization and engineering , - . jordanov, i. and brown, r. ( ) neural network learning using low-discrepancy sequence. in foo, n., ed., ai’ , lecture �otes in artificial intelligence , pp. - . springer-verlag, berlin heidelberg. kajitani, y., hipel, k.w., and mcleod, a.i. ( ) forecasting nonlinear time series with feed-forward neural networks: a case study of canadian lynx data. journal of forecasting , – . kampas, f.j. and pintér, j.d. ( ) configuration analysis and design by using optimization tools in mathematica. the mathematica journal ( ), - . kampas, f.j. and pintér, j.d. ( ) mathoptimizer: a nonlinear optimization package for mathematica users. research report, Özyeğin university, istanbul. (submitted for publication) khompatraporn, ch., pintér, j.d. and zabinsky, z.b. ( ) comparative assessment of algorithms and software for global optimization. journal of global optimization, , – . kolmogorov, a. n. ( ). on the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition. doklady akademii �auk ssr, , - . lang, b. ( ) monotonic multi-layer perceptron networks as universal approximators. in: duch, w. et al., eds., artificial �eural �etworks: formal models and their applications - ica�� , pp. - . lecture �otes in computer science , springer-verlag berlin heidelberg. malek, a., hosseinipour-mahania, n., and ezazipoura, s. ( ) efficient recurrent neural network model for the solution of general nonlinear optimization problems. optimization methods & software ( ), – . mason, t.l., emelle, c., van berkel, j., bagirov a.m., kampas, f. and pintér, j.d. ( ) integrated production system optimization using the lipschitz global optimizer and the discrete gradient method. journal of industrial and management optimization ( ), - . masters, t. ( ) signal and image processing with �eural �etworks, john wiley & sons, new york. mehrotra, k., mohan, c.k., and ranka, s. ( ) elements of artificial �eural �ets. mit press, cambridge, ma. pantoleontos, g., basinas, p., skodras, g., grammelis, p., pintér, j.d., topis, s., and sakellaropoulos, g.p. ( ) a global optimization study on the devolatilisation kinetics of coal, biomass and waste fuels. fuel processing technology , – . pardalos, p.m. and romeijn, h.e., eds. ( ) handbook of global optimization, volume . kluwer academic publishers, dordrecht. park, j. and sandberg, i.w. ( ) universal approximation using radial-basis-function networks. �eural computation ( ), - . park, j. and sandberg, i.w. ( ) approximation and radial-basis-function networks. �eural computation ( ), - . pei, j-s., mai, e., and piyawat, k. ( ) multilayer feedforward neural network initialization methodology for modeling nonlinear restoring forces and beyond. wcscm- , th world conference on structural control and monitoring, san diego, ca, july . pintér, j.d. ( ) global optimization in action. kluwer academic publishers, dordrecht. pintér, j.d. ( ) lgo ‒ a program system for continuous and lipschitz optimization. in: bomze, i.m., csendes, t., horst, r. and pardalos, p.m., eds. developments in global optimization, pp. - . kluwer academic publishers, dordrecht. pintér, j.d. ( ) global optimization: software, test problems, and applications. in: pardalos, p.m. and romeijn, h.e., eds. handbook of global optimization, volume , pp. - . kluwer academic publishers, dordrecht. pintér, j.d. ( ) globally optimized calibration of nonlinear models: techniques, software, and applications. optimization methods and software ( ) - . pintér, j.d. ( ) nonlinear optimization in modeling environments: software implementations for compilers, spreadsheets, modeling languages, and integrated computing systems. in: jeyakumar, v. and rubinov, a.m., eds., continuous optimization: current trends and modern applications, pp. - . springer science + business media, new york. pintér, j.d. ( ) nonlinear optimization with gams/lgo. journal of global optimization , - . pintér, j.d. ( ) software development for global optimization. in: pardalos, p.m. and t. f. coleman, eds. global optimization: methods and applications, pp. - . fields institute communications volume . published by the american mathematical society, providence, ri. pintér, j.d. ( a) lgo ─ a model development and solver system for �onlinear (global and local) optimization. user’s guide. (current version) distributed by pintér consulting services, inc., canada; www.pinterconsulting.com. pintér, j.d. ( b) spreadsheet-based modeling and �onlinear optimization with excel-lgo. user’s guide. with technical contributions by b.c. Şal and f.j. kampas. (current version) distributed by pintér consulting services, inc., canada. www.pinterconsulting.com. pintér, j.d. and kampas, f.j. ( ) nonlinear optimization in mathematica with mathoptimizer professional. mathematica in education and research ( ), - . pintér, j.d., linder, d., and chin, p. ( ) global optimization toolbox for maple: an introduction with illustrative applications. optimization methods and software ( ), - . prechelt, l. ( ) proben − a set of benchmarks and benchmarking rules for neural network training algorithms. technical report / , fakultat fur informatik, universitat karlsruhe, germany, september . ripley, b.d. ( ) pattern recognition and �eural �etworks. cambridge university press, cambridge, ma. röpke, k. et al., eds. ( ) doe – design of experiments. produced with the technical collaboration of iav gmbh. verlag moderne industrie. © sv corporate media gmbh, munich, germany. sexton, r.s., alidaee, b., dorsey, r.e. and johnson, j.d. ( ) global optimization for artificial neural networks: a tabu search application. european journal of operational research ( ) - . smith, m. ( ) �eural �etworks for statistical modeling, van nostrand reinhold, new york. steeb, w-h. ( ) the �onlinear workbook. ( rd edition) world scientific, singapore. sutskever, i. and hinton, g.e. ( ) deep, narrow sigmoid belief networks are universal approximators. �eural computation , – . tervo, j., kolmonen, p., lyyra-laitinen, t., pintér, j.d., and lahtinen, t. ( ) an optimization-based approach to the multiple static delivery technique in radiation therapy. annals of operations research , - . tilakaratne, ch.d., mammadov, m.a. and morris, s.a. ( ) predicting trading signals of stock market indices using neural networks. in: wobcke, w. and zhang, m., eds., ai : advances in artificial intelligence, pp. - . lecture �otes in computer science , springer-verlag berlin heidelberg, . toh, k-a. ( ) fast deterministic global optimization for fnn training. in: proceedings of ieee international conference on systems, man, and cybernetics, tokyo, japan, pp. v - v . watkin, t. l. h. ( ) optimal learning with a neural network. europhysics letters ( ), - . wikipedia ( ) artificial neural network. http://en.wikipedia.org/wiki/artificial_neural_network. wolfram research ( ) mathematica (version . . ). distributed by wri, champaign, il. ye, h. and lin, z. ( ) global optimization of neural network weights using subenergy tunneling function and ripple search. proc. ieee iscas, bangkok, thailand, may , v - . résumatcher: a personalized résumé-job matching system a thesis by shiqiang guo submitted to the office of graduate and professional studies of texas a&m university in partial fulfillment of the requirements for the degree of master of science chair of committee, tracy hammond committee members, anxiao jiang daniel w. goldberg head of department, dilma da silva may major subject: computer science copyright shiqiang guo abstract today, online recruiting web sites such as monster and indeed.com have become one of the main channels for people to find jobs. these web platforms have provided their services for more than ten years, and have saved a lot of time and money for both job seekers and organizations who want to hire people. however, traditional information retrieval techniques may not be appropriate for users. the reason is because the number of results returned to a job seeker may be huge, so job seekers are required to spend a significant amount of time reading and reviewing their options. one popular approach to resolve this difficulty for users are recommender systems, which is a technology that has been studied for a long time. in this thesis we have made an effort to propose a personalized job-résumé match- ing system, which could help job seekers to find appropriate jobs more easily. we create a finite state transducer based information extraction library to extract mod- els from résumés and job descriptions. we devised a new statistical-based ontology similarity measure to compare the résumé models and the job models. since the most appropriate jobs will be returned first, the users of the system may get a better result than current job finding web sites. to evaluate the system, we computed nor- malized discounted cumulative gain (ndcg) and precision@k of our system, and compared to three other existing models as well as the live result from indeed.com. ii acknowledgements my greatest thanks to the members of the sketch recognition lab for their continued support and help in the research work covered in this thesis. this thesis would not have been possible without their support. in addition, i would like to give extra thanks to my advisor dr. tracy hammond, as well as to my committee members dr. anxiao jiang and dr. daniel w. goldberg for their valuable sage advice. iii table of contents page abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii table of contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv list of figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi list of tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii . introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . recommender system . . . . . . . . . . . . . . . . . . . . . . . . . . . job recommender systems . . . . . . . . . . . . . . . . . . . . . . . . . information extraction in job recommender system . . . . . . . . . . matching algorithms in job recommender systems . . . . . . . . . . . problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . problem definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . system overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . system overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . system architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . text processing stages . . . . . . . . . . . . . . . . . . . . . . . . . . . system implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . system interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . information extraction . . . . . . . . . . . . . . . . . . . . . . . iv . semantic labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . patterns for matching . . . . . . . . . . . . . . . . . . . . . . . . . . . pattern matching library . . . . . . . . . . . . . . . . . . . . . . . . . . finite-state transducer . . . . . . . . . . . . . . . . . . . . . . . matchers in the pattern matching library . . . . . . . . . . . . . implementation of the pattern matching library . . . . . . . . model similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . similarity of major and academic degree . . . . . . . . . . . . . . . . . similarity of job title . . . . . . . . . . . . . . . . . . . . . . . . . . . similarity of skills . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ontology construction and similarity . . . . . . . . . . . . . semantic similarity in jrss . . . . . . . . . . . . . . . . . . . . . . . . ontology construction . . . . . . . . . . . . . . . . . . . . . . . . . . . ontology-based semantic similarity . . . . . . . . . . . . . . . . . . . . . path-based approaches . . . . . . . . . . . . . . . . . . . . . . . feature-based measures . . . . . . . . . . . . . . . . . . . . . . . content-based measures . . . . . . . . . . . . . . . . . . . . . . statistical-based ontology similarity measure . . . . . . . . . . . . . . evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . experiments of information extraction . . . . . . . . . . . . . . . . . . experiments of ontology similarity . . . . . . . . . . . . . . . . . . . . evaluation of the system . . . . . . . . . . . . . . . . . . . . . . . . . . comparing with the keyword searching . . . . . . . . . . . . . . . . . conclusion and future work . . . . . . . . . . . . . . . . . . . . conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v list of figures figure page . search result of indeed.com . . . . . . . . . . . . . . . . . . . . . . . . matching the jobs with résumé . . . . . . . . . . . . . . . . . . . . . . latent aspect model . . . . . . . . . . . . . . . . . . . . . . . . . . . . system architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . job description process pipeline . . . . . . . . . . . . . . . . . . . . . job description list . . . . . . . . . . . . . . . . . . . . . . . . . . . . upload resume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . résumé job matching . . . . . . . . . . . . . . . . . . . . . . . . . . . zero or one nfa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . single token nfa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . concatenation nfa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . alternation nfa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zero or one nfa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zero or more nfa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . one or more nfa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . finite automata transducers . . . . . . . . . . . . . . . . . . . . . . . procedure of finding technical terms . . . . . . . . . . . . . . . . . . interface of protege . . . . . . . . . . . . . . . . . . . . . . . . . . . . . part of ontology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi list of tables table page . résumé example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . portions of resume and job description . . . . . . . . . . . . . . . . . example of job description . . . . . . . . . . . . . . . . . . . . . . . . all words mean bachelors . . . . . . . . . . . . . . . . . . . . . . . . . semantic labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . more labels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . labeled sentence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . patterns to match degree sentences . . . . . . . . . . . . . . . . . . . . matchers of our library . . . . . . . . . . . . . . . . . . . . . . . . . . examples of matcher . . . . . . . . . . . . . . . . . . . . . . . . . . . example sentences in job descriptions . . . . . . . . . . . . . . . . . . patterns to extract terms . . . . . . . . . . . . . . . . . . . . . . . . . an example sentence matches the pattern . . . . . . . . . . . . . . . . some sentences of job descriptions . . . . . . . . . . . . . . . . . . . . similarities of skills list . . . . . . . . . . . . . . . . . . . . . . . . . similarities of skills list . . . . . . . . . . . . . . . . . . . . . . . . . information extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . similarities of skill list . . . . . . . . . . . . . . . . . . . . . . . . . . javascript similarity evaluation . . . . . . . . . . . . . . . . . . . . . html similarity evaluation . . . . . . . . . . . . . . . . . . . . . . vii . precision of job ranking . . . . . . . . . . . . . . . . . . . . . . . . . ndcg of job ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . comparison of the two approaches - java developer . . . . . . . . . . comparison of the two approaches - python developer . . . . . . . . comparison of the two approaches - average . . . . . . . . . . . . . viii . introduction . motivation currently one of the main channels for job seekers is online job finding web sites, like indeed or monster etc, that make the job finding process easier and decrease the recruitment time. but most such web sites only allow users to use keywords to search the jobs, which makes job searching tedious and blind task. for example, i used keyword “java” to search jobs with location restriction mountain view, ca on the job searching site indeed.com, the web site returned about , jobs (figure . ). the number of results of job searching is huge but not well ranked, so the job seeker has to review every job description. since no one has enough time to read all the jobs in the searching result, the actual quality of job searching service is low. this is a classic problem of information overflow. the reason for such a result is because current job searching web sites use the same information retrieval technology like “inverted index” [ ] as the common search engines, which just use keywords to map all the stored documents. modern search engines all have some ranking algorithms to sort the search result, like page rank [ ], so the top results always be the most related ones. but such algorithms are unavailable to the job search systems, because the criteria of how to rank the job searching result is very personalized. a great job opening for one job seeker maybe looks not good to the other, because the goodness of a job to a specular job seeker is heavily depend on his personal background, like his education or professional experience etc. since the people’s résumés contain the most important background information, we believe the content of the résumé could be used to rank the job openings. we give figure . : search result of indeed.com an example of résumés in table . . in this thesis, we created a web system which uses the résumés of job seekers to find the jobs that match their profiles best. the main idea is to calculate the similarity between the résumé model and job models, which should be generated from résumés and job descriptions. we want to transfer the job searching task from key word searching to candidate model matching. the matching result should be sorted by the matching score, higher matching score means a better matching. the matching algorithm does not only help job seekers to find the appropriate jobs, but also offers priority to them [ ]. the job with higher matching score means the job is more appropriate to the job seeker, and if he applies to the job, the chance of getting the interview will be higher as well. figure . shows how this approach works. table . : résumé example ryan richman work experience web developer fabuso/advanced brain technologies - ogden, ut - february to present created dynamic custom web applications for e-commerce and b b clients. designed and edited audio-visual content for many different online applications. spearheaded migration of largest client’s website from joomla platform into java code base. built dynamic event pages, document viewers, training course applications, shopping carts, and more. utilized advanced e-mail standards and best practices, sql database queries, and google analytics. it representative advanced brain technologies - ogden, ut - april to february provided internal software/hardware support for employees both in-house and remote. designed using wireframes, tested, and debugged web pages. constructed dynamic projects and graphic designs in coordination with senior developers. created html-optimized emails for hundreds of campaigns. maintained and upgraded hardware for + workstations company-wide. founder/head technician teton media services, llc - ogden, ut - october to january created and developed websites for personal and small business customers sold high-speed cable and satellite internet access on the phone, online and in person. installed and serviced high-speed internet access hardware in residential and commercial properties. designed and implemented networking solutions for homes and businesses. education computer science weber state university - ogden, ut to additional information technical skills adept in the use of html, css, jquery, javascript, sql, php, json, windows, win- dows server, mac, and photoshop. figure . : matching the jobs with résumé . contribution we make the following contributions in this work: . we proposed a résumé - job matching system. . we proposed a finite state transducer based matching tool to extract informa- tion from unstructured data source, which is a lightweight and flexible library, and can be extended in very easy ways. . we proposed a semi-automatic approach, which can collect technical terms from hr data sources, and by which we created a domain specific ontology for recruitment. . we proposed statistical-based ontology similarity measure, which can measure the similarities between technical terms . . organization the subsequent chapters are organized as follows: we first describe what has been done in terms of prior work. we introduce some basic conception of recommender systems, and how to apply recommender technologies into job recommender sys- tems. some previous job recommender systems will be introduced, their advantages and limits will be discussed as well. two import problems of content-based job rec- ommender systems, information extraction and similarity calculation, will be fully explained. then we introduce our work, résumatcher, a the personalized résumé-job match- ing system. first, we give an overview of the system, which includes the architecture and the interfaces. then we explained details of how we resolve the problems of in- formation extraction and model similarity calculation. we propose a finite state transducer library which can match patterns in sentence, and extracts related infor- mation. ontology plays an important role in this system. we will present how to construct the domain specific ontology for recruitment. we also give a brief review of different ontology similarity measures, and explain the statistical-based ontology similarity measure we used in this system. finally, we evaluate the accuracy of our information extraction approach. we used ndcg to evaluate the accuracy of statistical-based ontology similarity mea- sure. to evaluate the performance of the system, we compared our algorithm to some classical information retrieval approaches by precision@k and ndcg. we created a data set of job descriptions as documents, and use résumés as query to retrieval doc- uments. the result shows the ranking performance is better than other information retrieval approaches. . background some scholars found that current boolean search and filtering techniques can- not satisfy the complexity of candidate-job matching requirement [ ]. they hope the system can understand the job requirement, determine which requirements are mandatory and which are optional, but preferable. so they moved to use recom- mender systems technique to address the problem of information overflow. recom- mender systems are broadly accepted in various areas to suggest products, services, and information items to latent customers. . recommender system job searching, which has been the focus of some commercial job finding web sites and research papers is not a new topic in information retrieval. usually scholars called them job recommender systems (jrs), because most of them used technolo- gies from recommender systems. wei et al. classified recommender systems into four categories [ ]: collaborative filtering, content-based filtering, knowledge- based and hybrid approaches. some of these techniques had been applied into jrs; zheng et al. [ ] and alotaibi et al. [ ] summarized the categories of existing on- line recruiting platforms and listed the advantages and disadvantages of technical approaches in different jrss. the categories include: . content-based recommendation (cbr). the principle of a content-based rec- ommendation is to suggest items that have similar content information to the corresponding users, like prospect [ ]. . collaborative filtering recommendation (cfr). collaborative filtering recom- mendation finds similar users who have the same taste with the target user and recommends items based on what the similar users, like casper [ ]. . knowledge-based recommendation (kbr). in the knowledge-based recommen- dation, rules and patterns obtained from the functional knowledge of how a specific item meets the requirement of a particular user, are used for recom- mending items, like proactive [ ]. . hybrid recommender systems combine two or more recommendation techniques to gain better performance, and overcome the drawbacks of any individual one. usually, collaborative filtering is combined with some other technique in an attempt to avoid the ramp-up problem. . job recommender systems rafter et al. began to use automated collaborative filtering (acf) in their job recommender system, “casper” [ ]. in the system user profiles are gotten from server logs, that includs: revisit data, read time data, and activity data. all these factors are viewed as measure of relevance among users. the system recommend jobs in two steps: first, the system finds a set of users related to the target user; second, the jobs that related users liked will be recommend to the target user. the system use cluster-based collaborative filtering strategy. the similarity between users are based on how many jobs they both reviewed, or applied. casper also allows users to search jobs by a query which is a combination of some fields: like location, salary, skill and so on. the system uses such query to find jobs, and the returned jobs are ranked with the collaborative filtering algorithm. in their paper, the authors do not give a detailed description on how to detect the related fields they need and how to the transfer semi-structured job description to the structured data. the shortages of collaborative filtering: first, since the number of search results is huge, and the results are sorted randomly, the probability of two similar users reviewing the same jobs is low, which causes the sparsity problem of collaborative filtering. the authors also noticed the sparseness problem caused by few in users profile, so they try to user cluster-based solution to resolve this problem. second, because recommended jobs are from others users’ search results, since the quality of current searching result are low, the quality of recommendation cannot be high. färber et al. [ ] presented a recommender system built on a hybrid approach. the system integrate two methods, content-based filtering and collaborative filtering, which tries to overcome the problem of rating data sparsity by leveraging a combined model. in the system, the data source is synthetic resumes. the latent aspect model is shown in figure . . figure . : latent aspect model in malinowski et al. [ ], they classified the job recommender systems into t- wo categories, cv-recommender, which recommends cvs to recruiter, and the job- recommender, which recommends jobs to job seekers. the system collects the users’ profile data by asking input their profiles to the web form based interface field by field. the input data collected are: . demographic data (e.g. date of birth, contact information) . educational data (e.g. school courses, grades, university, type of degree, inter- mediate and final university examinations, postgraduate studies) . job experience (e.g. name of the company, type of employment, industry group, occupational field) . language skills (e.g. language, level of knowledge) . it skills (e.g. type of skill, level of knowledge) . awards, scholarships, publications, others the system also asked the users to upload their resumes, but the resumes were only for facilitating the human judgment. in malinowski’s study, the latent aspect model is a statistical model, which needs to be trained before applied to recommendation. the system uses the users’ search results as the training data to train the model for recommendation, so the system needs a a long time training for each user. . information extraction in job recommender system big it companies met the similar problem of information overflow when they received many resumes for one job opening. the recruiter had to screen the all the applications manually, but this task was also tedious and time consuming. for this reason these companies tried to build systems to help screen the resumes. amit et al. in ibm presented a system, “prospect” [ ], to aid shortlisting candidates for jobs. the system uses a résumé miner to extract the information from résumés, which use a conditional random field (crf) model to segment and label the résumés. the crf model used three kinds of features, they are: lexicon, visual, and named entity. the paper compared some algorithms to ranked the candidates applicants, such methods include: okapi bm , kl, lucene scoring, and lucene scoring + skillboost. hp also built a system to solve the similar problem, which is introduced in gon- zalez et al.’s paper [ ]. the system also pays a lot of attention to information extraction. the dictionaries which are used to tag entities need to be updated often since there always new terms appears. so an adaptive learning module is used to achieve two objectives: use semantic data to enhance the information extraction and to discover new terms. a domain-oriented ontology is used to represent knowledge, inference rules are defined based on the ontology knowledge base. when a detected term found, the system will search in external knowledge base, like dbpedia. the résumés are also classified to different categories like “web technology” and “no web technology” by naive bayes classifier. the company can allocate appropriate employees to required positions with the system. the goal of the systems built by ibm and hp is to help the companies to select good applicants, but cannot help job seekers to find appropriate jobs. yu et al. [ ] used a cascaded ie framework to get the detailed information from the résumés. in the first stage, the hidden markov modeling (hmm) model is used to segment the résumé into consecutive blocks. based on the result, a svm model is used to obtain the detailed information in the certain block, the information include: name, address, education etc. celik duygua and elci atilla proposed a ontology-based rsum parser (orp) [ ], which uses ontology to assistant the information extraction process. the system processes a résumé in following steps: converting the résumé file into plain text, separating the text into some segments, using the ontology knowledge base to find the concepts in the sentences, normalizing all the terms, and classifing the sentences to get the wanted terms. but the personal information the system retrieved like name and addresses is not the information that the recruiters care about. the recruiters want some information that relate to the job opening, and can help them to judge the competence of job applicants. . matching algorithms in job recommender systems lu et al. [ ] used latent semantic analysis (lsa) to calculate similarities between jobs and candidates, but they only selected two factors “interest” and “education” to compare candidates. xing et al. [ ] used structured relevance models (srm) to match résumés and jobs. drigas et al.[ ] presented a expert system to match jobs and job seekers, and to recommend unemployed to the positions. the expert system used neuro-fuzzy rules to evaluate the matching between user profiles and job openings. the system uses a relation matrix to represent the fuzzy relation between these specialities. the system needs the training data to train that neuro-fuzzy network. both résumé data and job opening data were manually input into the system. daramola et al.[ ] also proposed a fuzzy logic based expert system(fes) tool for online personnel recruitment. in the paper, the authors assumed that the information already be collected. the system uses a fuzzy distance metric to rank candidates’ profiles in the order of their eligibility for the job. yao et al. [ ] also presented a hybrid recommender system which integrated content-based and interaction-based relation. in content-based part, relations be- tween job-job, job-job seeker, and job seeker-job seeker can be identified by their similarity of profiles. there two approaches are used to calculate the similarities: for the structured data, like age and gender, the weight sum values will be returned; for the unstructured data, like similarity between job and user profile, the latent semantic analysis is applied in the system. in summary, there are some problems exist in previous job recommender sys- tems: . most systems can only process the structured data of résumés and job descrip- tions, but in reality both them are in unstructured formats, such as text files or html web pages. . the systems that have information extraction modules are designed for re- cruiters to select applicants, not for job seekers to select jobs. . in the systems the information fields to match résumés and job descriptions are coarse-grained. to improve the quality of recommendation, we need to improve the granularity of the information fields. . problem . problem definition the basic problem in this thesis is how to find appropriate job descriptions by user’s résumé, which means we need calculate the similarity between the users rsum and the job. if we take the résumé as a query and the job descriptions as documents, we need to build an information retrieval model to get the most relevance documents. the résumatcher will parse the job descriptions to the job models, and store them in the database. when a user searches the jobs by their résumé in the system, the system will compare the similarity values between the résumé and the job models, and return the jobs sorted by their similarity values. the core idea of our algorithm is calculate similarity between the résumé model and job model. we give a formal definition of our problem. all of the notations will be used frequently throughout the thesis. we use r to denote the user’s résumé model, which has some features ri like their academic degree, their major, their skills and so on. the symbol j is the set of job models stored in the database, and j is a job model in the set j. the similarity function sim(r,j) gives the similarity values between résumé r and job j. the return list of search function search(r,j) will calculate all the similarity value in the database, and the result of the function will be the job description list ranked by their similarity values. the equation of how to calculate similarity value is given below: sim(r,j) = n∑ i= simfuni(ri,ji) ×wi the value of sim(r,j) is the summation of the similarity values of different fields times their corresponding weights. different fields like major and skills, may have different functions to calculate their similarity values. we will describe the similarity functions of individual fields in later parts. . challenges there are two challenges exist in our system. the first one is how to extract models of jobs and résumés. to calculate the similarity between a job and a resume, the résumatcher system needs structured digital models of each document. to get the structured data, some jrss ask the job seekers input their profiles in forms field by field, and the recruiter input their job descriptions in the same way. however, as we discussed in chapter , the users are reluctant to take the tedious process [ ]. job seekers prefer upload their resumes directly, and recruiters prefer to post the whole job descriptions to web sites. so we need extract the structured information from un-structured data source, like résumés and job descriptions. the other challenge is how to compute the similarity between rsum and job models. we observed that simple keyword matching is not a good similarity measure, because job descriptions and résumés both contain richer and more complex words that cannot be described simply by keywords. in these documents, some concepts can be written in different ways, and other concepts can have close relationships. for example, table . shows portions of a résumé and a job description: if just looking at the text, we can find that the résumé has very few common words with the job description. but from the view of an experienced engineer, the candidate is closely matches the job: the two relational databases oracle and mysql are very similar, ooa/ood is the same meaning of many years of java and c++ experience, and tomcat and jboss are both java web applications servers. if we table . : portions of resume and job description resume portion job description portion b.s. degree in computer science + years java + year c++ some experience in oracle database other experience like: hibernate, jboss, junit, tomcat etc. bs degree above + years java some experience of python mysql, ms-sql java web application server ooa/ood use keyword matching, the system does not provide a strong matching result in very common cases such as this. so we need a better approach to calculate the similarity between different technical concepts. . system overview . system overview the system uses information extraction technique to parse job descriptions and résumés, and it gets information such as skills, job titles and education background. the information is used to create the models of job openings and job seekers. a domain specific ontology is used to construct the knowledge base, which includes the taxonomies that support résumé-job matching. the models of résumé includes job seekers’ specialties, working experience and education background, and all the fields are extracted from their résumés. the job models are extracted from job descriptions, and they have the same information fields as the résumé models. when a job seeker searches the jobs by their résumé, the system calculates the similarity between the résumé model and job models, then gives every job model a similarity value. . system architecture figure . shows the architecture of the whole system, which includes such mod- ules: . the web crawler can access and download all new it job opening web pages from indeed.com everyday. . the job parser can parse the job opening web pages, extract the information and create the job models. . the resume parser is much like the job parser; it parses the résumés and creates the résumé models. . all the job descriptions and job models are stored in the database. . when a user searches the jobs with their résumé, the ontology matcher calcu- lates the similarity values of jobs in the database and returns the jobs ranked by their similarity values. figure . : system architecture . text processing stages information extraction is the task of automatically extracting structured infor- mation such as entities, relationships between entities, and attributes describing entities from unstructured sources [ ]. the ie framework in our system uses six stages in order to extract the information from job descriptions: html parsing, segmenting, preprocessing, tokenizing, labeling and pattern matching, which is show in figure . . ) the html parsing will parse the web pages that contain job descriptions, which are obtained from web crawler. the parser uses html tag template to extract attributes of the jobs, like job title, location, company name, content and so on. a job will be saved as a record with these attributes in the database. in the record, the content field contains the text part of the job description, which will be processed in later stages. ) in the segmenting stage, the content field of the job description is be sep- arated into paragraphs according html tags. then paragraphs are separated into sentences by either html tags or punctuation, and after this step, all html tags will be removed. ) web pages of job description are created in different character sets, (e.g. utf and iso - ), and almost always contain some unreadable characters. in the prepossessing stage, characters in the sentences are converted to ascii characters, unreadable characters will be deleted, and some punctuation will be replaced by spaces (e.g. / and -). ) in the tokenizing stage, the sentences will be tokenized into arrays of tokens by nltk [ ]. ) in the labeling stage, the sentences will be given two layers of labels by a dictionary matching approach. the labels in the first layer are the semantic value of the text, and the labels in the second layer are the ontology hypernym of the labels in the first layers. ) in the pattern matching stage, the fst library is used to matching the labels of the labeled sentences. if a layered sentence match any pre-defined pattern, the information will be extracted and added to the job model. after every sentence of a job description has be processed, a job model will be created and saved in the database. more details about matching will follow in section c. figure . : job description process pipeline . system implementation we will describe some implementation details here. the whole system is imple- mented in python and uses some third party libraries and frameworks. we used flask, a lightweight web framework, to build the web application. we used rdflib as the web ontology language (owl) file parser, python lex-yacc (ply) as the token regular expression compiler, whoosh as the inverted index builder and beauti- ful soup as the html parser. all the jobs retrieved by the web crawler are stored in the mongodb nosql database. for the natural language processing procedure, we used natural language toolkit (nltk), a natural language processing library, to extract and tokenize the sentences. . system interface the system provides some interfaces to end users. the most important interfaces are the web pages like: reviewing all the jobs in the database, searching the jobs by keyword (figure . ), uploading users’ résumés (figure . ) and matching the jobs with a résumé (figure . ). figure . : job description list figure . : upload resume figure . : résumé job matching . information extraction in this chapter we will explain how the information extraction (ie) module of our system extracts information from these unstructured data source. an example of job description is shown in table . . the ie framework will be introduced by example of processing the job descriptions. the finite-state transducer(fst) library, which is used as pattern matching tools, will be introduced as well. . semantic labeling in this section, we will introduce why and how we add two layers of labels to the tokenized sentences. in natural language, a single concept often has multiple expressions to represent it. for example, the simple concept bachelor’s degree, can be expressed in many ways in job descriptions, e.g. b.s., ba/bs, -years-degree, and so on. table . shows the words that if followed with word “degree” have the semantic value of “bachelor’s degree”. to add labels to a sentence, we use regular expression over tokens. a regular expression over token transfer a patten to a finite-state transducer (fst), and every token of that will be transferred to an edge of fst. if we use all the expressions of a semantic value to create a pattern, the pattern will be very large, and there are too many states in the fst. for example, if we use some words in table . to create the pattern of semantic value “bachelor’s degree”, the pattern will like below: ( baccalaureate | bachelors | bachelor | b.s | bs | ba ) degree if all words in table . are added to the pattern, the fst will have too many edges, and the matching process will be very slow because of the problem of combinatorial table . : example of job description senior/principal software engineer richrelevance - san francisco, ca richrelevance powers personalized shopping experiences for the worlds largest and most innovative retail brands, including target, sears, marks & spencer, john lewis and oth- ers. founded and led by the e-commerce expert who helped pioneer personalization at amazon.com, richrelevance helps retailers increase sales and customer engagement by recommending the most relevant content to consumers regardless of the channel they are shopping. richrelevance has delivered more than $ . billion in attributable sales for its retail clients to date, and is accelerating these results with the introduction of a new form of digital advertising called shopping media which allows manufacturers to en- gage shoppers where it matters most – in the digital aisles on the largest retail sites in world. richrelevance is headquartered in san francisco, with offices in new york, seattle, boston, reading and malmho, and has been twice recognized as one of the best places to work in the bay area. richrelevance is looking for a senior/principal software engineer to join our growing team! primary responsibilities: • working with large scale distributed systems • work with hadoop ecosystem (technologies like hive, impala, hbase) • algorithmic development with primary focus machine learning • working with rapid and innovative development methodologies like: kanban, continuous integration and daily deployments • unit testing with junit, performance testing and tuning minimum requirements: • bs/ms in cs, electrical engineering or foreign equivalent plus relevant software development experience • at least + years of software development experience • expert in java, scala or any other object oriented language • proficient in sql concepts (hiveql or postgres a plus) • additional language skills for scripting and rapid application development desired skills and experience: • working with large data sets in the pbs • familiarity with unix (systems skills a plus) • working in a distributed environment and has dealt with challenges around scaling and performance • mobile development for android or ios. richrelevance is an equal opportunity employer and does not discriminate against any applicant on the basis of race, color, religion, national origin, gender, marital status, age, disability, sexual orientation, military/veteran status, or any other status protected by federal or state law or local ordinance. table . : all words mean bachelors ”baccalaureate”,”bachelors”, ”bachelor” ,”b.s.”, ”b.s”,”bs”,”ba”,”ba/bs”, ”babs”, ”bsba”, ”b.a.” ,” -year”,” -year”, ” year”, ”four year”,”college”,”undergraduate” , ”university” explosion. to resolve this problem, we proposed an approach to use the patterns to match the labels of the tokens, not the the original text. in the system, we don’t care what words the sentences really use, but want to extract the semantic value of the tokens which match the pattern. the details of the approach is described below. at first, we created two dictionaries, which are used to label the tokens. in the first dictionary, the keys are the tokens, like words in table . , and the values are the symbols for semantic values, like “bs-level” for “bachelor’s degree”, or “ms-level” for “master’s degree”. the values of the the second dictionary are the ontology hypernym of their keys, like keys “bs-level” and “ms-level” both have value “de-level”, which means that bachelor’s degree and master’s degree are both one kind of degree level. we show the dictionaries for degree information in table . . there are also some words in the dictionaries that have the same first layer and second layer labels, which is shown in table . . with the two dictionaries, we can label the tokens with two layers. table . shows how the sentence “bachelors degree in computer science or information sys- tems.” is labeled. the pattern “de-level degree in major or major ” can match the sentence above, and the output of the matching process is “bs-level” for bachelor’s degree, “major-cs” and “major-info” for two majors mentioned in sentence. in our system, most patterns match the labels in second layer. with this approach, table . : semantic labeling original text layer layer bachelors bs-level de-level bachelor b.s. baccalaureate master ms-level ms m.s. phd phd-level ph.d doctorate table . : more labels ”be”, ”be”, ”is”, ”are”, ”am” be be ”a”, ”a”, ”an”, ”an”, ”the”, ”the” de de ”mba”, ”bscs”, ”bsee”, ”mscs” major-degree major-degree ”msee”, ”msce”,”mph” ”practical experience”, ”work experience” experience experience ”professional experience”, ”experience ”preferred”, ”required”, ”desired” prefer-vbd prefer-vbd ”a plus”, ”mandatory”, ”desirable” prefer-jj prefer-jj ”similar”, ”related”, ”relevant” degree-jj degree-jj ”equivalent”, ”based” the size of the fst for the pattern will be minimized, so speed of matching process can be improved. . patterns for matching as we explained in section b, we mentioned matching tokens in the second layer to patterns we defined. to match the labels in sentences to our patterns, we proposed a library that support matching pattern over tokens. the difference between this library and traditional regular expression is that the basic unit to be matched is token, not character. some patterns used to match degree phrases are in table . . table . : labeled sentence layer de-level degree in major or major . layer bs-level degree in major-cs or major-info . words bachelors degree in computer science or information systems . the patterns looks like regular expression, but they use tokens as the basic units. table . : patterns to match degree sentences de-level, de-level, or de-level degree de-level degree ( in | of ) dt major major-degree , major-degree or major de-level (, de-level)* (or de-level)? be? perfer-vbd . pattern matching library in section c we introduced how we use the library of pattern matching over tokens to match the sentences. in this section we will introduce more details of this library, including its advantages and implementation details. . . finite-state transducer finite-state transducers [ ] have been used as a tool to match patterns and extract information for more than years. this approach has been demonstrated to be very effective in extracting information from text like circus [ ] and fas- tus [ ]. in the widely used nlp toolkit gate [ ], the semantic tagger jape (java annotations pattern engine) could describe patterns that are used to match and annotate tokens. jape adopts a version of cpsl (common pattern specifica- tion language) [ ], which provides fst over annotations. chang et al. presented cascaded regular expressions over tokens [ ], which proposed a cascaded pattern matching tool over token sequences. after studying these tools, we found most of them to be powerful and complex, but not very flexible. one reason is that developers need to learn some domain specific languages (dsls) like cpsl. the other reason is the extra effort and time required to integrate the pattern matching tool into the system. so here we proposed a more flexible and lightweight fst framework, which can do regular expression matching over labeled tokens. we give the definition of finite-state transducer here. a finite-state transducer is a -tuple (Σ , Σ ,q,i,f,e) where: • Σ is a finite alphabet, called the input alphabet. • Σ is a finite alphabet, called the output alphabet. • q is a finite set of states. • i ∈ q is the initial state. • f ⊂ q is the set of final states. • e ⊂ q× Σ∗ × Σ∗ ×q is the set of edges. for example, the fst td = ({ , },{ , },{ , , },ed ) where ed = { ( , , , ), ( , , , ), ( , , , ), ( , , , ), ( , , , ), ( , , , ) } is shown in figure . . regular expressions can be converted to automata [ ], and fst is also an au- tomata. to convert a regular expression over token to a fst we need two steps: the first is parsing the expression to a tree of matchers, the second is transfer the tree of matchers to the fst. we will introduce these two steps in next. figure . : zero or one nfa . . matchers in the pattern matching library in our library, a “matcher” could be a token to be matched, or a composition of other matchers. our library supports syntax used in traditional regular expressions over strings. we list the syntax that the library supports in table . . the first column is the names of the matchers, the second column is the explanation of the function of the matchers, and third column is the their counterpart syntaxes of traditional regular expression. the regexmatcher in our library is constructed with a regular expression, and the matcher matches any string that matches the regular expression in the matcher. we give examples of the syntax of these matchers in table . . table . : matchers of our library matcher name function counter part of regex unitmatcher token is matches the it character in regex sequencematcher a list of matcher sequence of characters questionmatcher one or more of the preceding token ? starmatcher zero or more of the preceding token * plusmatcher zero or one of the preceding token + dotmatcher any token . regexmatcher any token matches the regular ex- pression n/a table . : examples of matcher matcher name example unitmatcher degree sequencematcher de-level degree questionmatcher de-level (or de-level)? degree starmatcher de-level (, de-level)* degree plusmatcher degree in major + dotmatcher has . degree regexmatcher r“d-d” years the framework supports three styles of creating patterns: regular expression style, operator style and object style. the second and third styles are flexible be- cause developers can create their own matcher class to extend the feature of the library. we use examples to show how the three styles work. the most common style is defining pattern expression in a string, which is much like traditional regular expression. the pattern is: de-level degree ( in | of ) dt? major the code is: seqmatcher =parser.parse("de-level degree ( in | of ) dt? major") the second style is using algebraic operators to connect matchers, which can help developer reuse previous patterns when the new patterns include old ones. it is shown in follows: the pattern is: ”de-level degree (in | of) major” the code is: seqmatcher = unitmatcher("de-level") + unitmatcher("degree") + ( unitmatcher("in") | unitmatcher("of" ) ) + unitmatcher("major") we also could create a complex matcher in object-oriented programming style. the pattern is: ”de-level degree (in | of) major” the code is: matcher = unitmatcher("de-level") matcher = unitmatcher("degree") matcher = unitmatcher("in") matcher = unitmatcher("of") matcher = unitmatcher("major") matcher = alternatematcher([matcher ,matcher ]) seqmatcher = seqmatcher([matcher , matcher , matcher , matcher ]) the flexibility of the tool also comes from the fact that developers can determine which layer of the array should be matched, the original text, or labels in the first layer or labels in the second layer. developers can assign a lambda expression to the matcher’s catching function, which defines how to get the matching input strings from the sentences, as well as an out function, which defines what should be outputed. for example, the labeled sentence is a sequence of arrays, each array includes the original text token and its labels in the other two layers, which is shown in table . . to match the labeled sentence, we set the lambda expression for catching function to “lambda x:x[ ]”, and the out function to “lambda x:x[ ]”, which make the matcher match the label in second layer, and output the the value of semantic value in the first layer. . . implementation of the pattern matching library we use ply(python lex-yacc) as the grammar parser, which is a pure-python implementation of the popular compiler construction tools lex and yacc. we defined syntaxes of regular expression over tokens in the parser, which can parse the token regular expression to the tree structure of matchers. we use the algorithm proposed by thompson and ken [ ] to construct the fst from the tree of matchers, which is shown in algorithm . the algorithm accesses the inner structure of a matcher recursively, and create state for the matcher. each state is a partial nondeterministic finite automaton (nfa), which has one or more dangling arrows. the algorithm builds the whole nfa by connecting the arrows of the partial nfas. different operator will have different structures of nfa, which is shown below: the nfas for matching single token is shown in figure . . figure . : single token nfa the nfa for the concatenation e e connects the final arrow of the e machine to the start of the e machine, as shown in figure . . figure . : concatenation nfa the nfa for the alternation e | e adds a new start state with a choice of either the e machine or the e machine, which is shown in figure . . figure . : alternation nfa the nfa for e? alternates the e machine with an empty path, which is shown in figure . . figure . : zero or one nfa the nfa for e* uses the same alternation but loops a matching e machine back to the start, which is shown in figure . . figure . : zero or more nfa the nfa for e+ also creates a loop, but one that requires passing through e at least once, which is shown in figure . . figure . : one or more nfa with the above rules, we can convert the expression “dl (, dl)* (or dl)? degree” to an fst in figure . . figure . : finite automata transducers comparing our method to other state of art machine learning-based information extraction algorithms, our method has such advantages: . easy to implement. according to the pseudo code in , even some undergrad- uate students without strong machine learning background can implement our algorithm in a few days. . there is no need for data labeling. supervised machine learning algorithms always need a large labeled training data set, but data labeling is tedious work input: matcher, nfa output: nfa switch matcher do case unitmatcher state = compiletonfa(matcher]) ; connect( nfa.last, [state] ) ; nfa.last = [state] ; break; case sequencematcher for (i = ; i < len(matcher.list); i + +) do state = compiletonfa(matcher.list[i]) ; connect( nfa.last, [state] ) ; nfa.last = [state] ; end break; case alternatematcher state = compiletonfa(matcher.list[ ]) ; state = compiletonfa(matcher.list[ ]) ; connect( nfa.last, [state , state ] ) ; nfa.last = [state , state ] ; break; case questionmatcher state = compiletonfa(matcher.list[ ]) ; connect( nfa.last, [state , nfa.last] ) ; nfa.last = [state , nfa.last] ; break; case plusmatcher state = compiletonfa(matcher.list[ ]) ; connect( state , state ) ; connect( nfa.last, state ) ; nfa.last = [state ] ; break; case starmatcher state = compiletonfa(matcher.list[ ]) ; connect( state , state ) ; connect( nfa.last, state ) ; nfa.last = [state , nfa.last] ; break; endsw algorithm : compiletonfa for users. with our pattern matching algorithm, we avoid the work manual data labeling. . the accuracy of information extraction can increase monotonously as the num- ber of patterns increase. so with enough patterns, the accuracy becomes quite high. . high speed for labeling data. the time complexity of pattern matching is o(n) [ ], which is smaller than some complex machine learning based approach- es. one example is conditional random fields(crfs), which uses viterbi algorithm [ ] to label the sequence, the time complexity of it is o(n t). in this chapter we have introduced how we extracted the information from the resumes and job descriptions, and the implementation details of the pattern matching library, regular expression over tokens. we can get the models of resumes and job descriptions through the procedure described in this chapter. in the next chapter, we will discuss how our system searches and ranks job models by resume models. . model similarity the similarity value between a job model and a résumé model is the summation of weighted similarity values of different fields. the equation is given below: sim(r,j) = n∑ i= simfuni(ri,ji) ×wi the value of sim(r,j) is the summation of similarity values of different fields times their corresponding weights. simfuni(ri,ji) is the similarity function of the ith field of the model. in our system, the résumé model and job model both have four fields: job title, major, academic degree and skills. the similarity value between a résumé model and job model is the sum of the productions of similarity values of all the fields pairs and their weights. we will introduce how to calculate the similarity value for each field in this chapter. . similarity of major and academic degree in the simplest case, if the majors in the résumé model and job model are the same, the similarity value is . if they are different, we can check whether the major in the résumé model is in the list of related majors for the major in the job model. if it is, the similarity value is . ; otherwise the similarity value is . the equation is shown below: majorsim(r,j) =   , rmajor = jmajor . , rmajor ∈ related(jmajor) , otherwise   there are five kinds of academic degrees in the system: high school, associate, bachelor, master, and ph.d., which are mapped to the integer values form to . if the degree value in the résumé model is less than that in the job model, which means that the job seeker’s education background cannot satisfy the requirement of the job, the similarity value in this case is . if the degree value in the résumé model is equal to the job model and no more than above, the similarity value is . in some cases, the degree value in the résumé model is greater than that of the job model, and the difference is greater than , which means that the job seeker’s degree is much higher than the requirement of the job. the situation is also a kind of relative matching, so the similarity value here is . . the equation is shown below: degreesim(r,j) =   , rdegree < jdegree , < rdegree − jdegree . , rdegree − jdegree >   . similarity of job title another field of needs similarity calculation is the job titles in the models. a job title can be parsed into some sub fields: job role, level, platform, programming language. the value of job roles includes: developer, manager, administrator and so on. there are levels values in such roles: such as junior, senior and architect. the platforms: web, mobile and cloud are used by the very roles. the similarity value between two titles is the sum of all the similarity values of these fields. the similarity value of each sub filed ranges from to , and we also normalized the similarity summation value to by dividing the number of sub fields. if the job seeker has some working experience, there may be some job titles in their résumé. when calculating the similarity value between a résumé model and a job model, the system calculates the similarity values of the title of job model to all the titles in the résumé model and returns the maximum one. . similarity of skills the job model usually has requirements of some skills, and the résumé model lists the skills the job seeker has as well. the similarity value of skills field is the normalized summation of all the similarity values of skills in the job model. skillsetsim(sj,sr) = ∑ sji∈sj skillsim(sji,sr) |sj| for every skill in the job model, the similarity value is the maximum value it can get from the skills in the résumé model. the equation is shown below: skillsim(sji,sr) =   , sji ∈ sr max(sim(sji,rjk)), sji /∈ sr   in the equation, sji is the ith skill in the job model, and sr is the skill set of the résumé model. if there is the skill sji in the skill set sr, the similarity value for sji is , otherwise the system chooses the maximum similarity value from all the similarity values between skill sji and the skills in the the résumé model. we introduced how to calculate similarity values for three fields in résumé and job models. in the next chapter, we will introduce how to use a domain specific ontology to calculate similarity values between the skills fields of the two models. . ontology construction and similarity . semantic similarity in jrss after getting job models by the information extraction module, users can search for jobs in the system. in previous studies of jrss, ontology is used as a knowledge base to store knowledge and rules, which could help compare the similarity between different concepts. liu and dew [ ] used resource description framework (rdf) to represent and store the expertise of experts, and they used a rdf-based expertise matcher to retrieve the experts whose expertise included the required concept. proactive [ ] used two kinds of ontology, job category and company information. the system used an ontology checker to classify the job information, stored the domain knowledge and calculated the weight value in recommendations. fazel [ ] used a hybrid approach to match job seekers and job postings, which takes advantage of the benefits of both logic-based and ontology-based matching. in his paper the description logics (dl) are used to represent the candidate and job opening, and the ontology is used to organize the skills in a taxonomy. the paper provides an equation to calculate the matching degree: sim (p,j) = ∑ xij ×u(dsi) where xji is the boolean variable indicating whether desire i is satisfied by appli- cant aj in the set of all qualified applications. kumaran et al. [ ] also used an ontology to calculate the similarity between the job criteria and candidates’ résumé in their system [ ]. the similarity equation they used is: m (i , i ) = ∑n k= sim (p i k ,p i k ) ∗w i k∑n k= w i k the similarity function sim(p ,p ) is defined as follows: sim(p ,p ) =   , if similarity of p and p > t , otherwise   . ontology construction before calculating the similarity between concepts, we need to construct the on- tology first. semantic web has been a popular research topic in previous years, and at the same time thousands of domain ontologies have been created [ ]. a paradig- matic example is wordnet [ ], which is a general purpose thesaurus, and contains more than , general english concepts. acm has created a poly-hierarchical ontology that can be utilized in semantic web applications [ ], but it is mostly used in academic areas. dbpedia [ ] provides structured information from wikipedia and make this information available on the web, but its coverage is huge, and most of them is not related to job finding. currently, there is no domain specific technology ontology built for recruiting purpose. the domain specific technology ontology for recruiting should include a lot of technical terms, like programming language, programming library, commercial prod- ucts and so on. furthermore, there are new techniques invented everyday, so new it terms will appear continuously. ding et al. [ ] gave a survey of current ontol- ogy generation approaches such as manual, semi-automatic, and automatic. some aspects of the approaches were discussed in the paper, like the source data, concept extraction methods, ontology representation, and construction tools. inspired by this paper, we propose a semi-automatic approach to construct the it skill ontology, which use a pattern matching approach to collect possible technical terms, and use dbpedia to verify the them. from the observation, we found that sentences with skill requirements in job descriptions always list several skills in the sentence, which is shown in table . . based on this character, we propose a bootstrap approach to collect it terms in job descriptions. first, we manually collect about fifty terms from job descriptions, and add them to the term list. then we use our pattern match library to find the sentences that matching the pattern in table . from a set of job descriptions. an example of a sentence which matches the pattern is shown in table . . we extract the tokens which match the star symbol from the sentences; these tokens have high probability to be technical terms. then we could check the tokens in dbpedia to see whether they are under the categories like software, programming language or any other technical related ones. if they are, we could classify them as terms, and add them to the terms list. after scanning all the sentences in the job description set, the term list will be larger, and we can use the larger term list to start a new iteration of scaning. this process stops when the number of found new terms is below a threshold. the process is shown figure . . table . : example sentences in job descriptions . a high-level language such as java, groovy, ruby or python; we use java and groovy extensively . html /css /javascript, web standards, jquery or frameworks like an- gularjs would be great . html css and javascript a must . experience with ajax, xml, xsl, xslt, css, javascript, jquery, htm- l and web services table . : patterns to extract terms term , * , *, term term , * , *, and term table . : an example sentence matches the pattern experience with term , * , * , term , and * experience with ajax , xml , xsl , xslt , and css for example, we extract the token ”xsl”, which currently is not in the terms list. we check the word on dbpedia by accessing the url:http://dbpedia.org/page/xsl. if we can get the xml formatted description of xsl, and any element in “dc- terms:subject” section has the value which is a technical category, like “program- ming languages”, “markup languages” and so on, we can indicate that the word is a technical term, and add it to the term list. figure . : procedure of finding technical terms but not all the extracted terms can be verified in dbpedia, because some terms have multiple meanings in english, and the urls of their dbpedia pages are unpre- dictable. for example, the word “python” could be an animal name or a program- ming language. the meaning of the programming language has the dbpedia url http://dbpedia.org/page/python (programming language), which is difficult to pre- dict. in this case, we have to check the term manually. after getting all terms, we use protege [ ], an open source ontology editor, to edit the domain specific ontology, and saved it in rdf format. the interface of protege is shown in figure . . part of the technical ontology is shown in figure . . figure . : interface of protege . ontology-based semantic similarity sánchez et al. [ ] summarized ontology-based similarity assessment into three kinds and gave both advantages and disadvantages of each approach. the three kinds figure . : part of ontology of categories are: edge-counting approaches, feature-based measures, and measures based on information content. . . path-based approaches in path-based approaches, the ontology is viewed as a directed graph, in which the nodes are the concepts, and the edges are taxonomic relation (e.g. is-a). rada, et al. [ ] measure the similarity by the distance of two nodes in the graph. therefore, the semantic distance of two concepts a and b will be: disrad(a,b) = min|pathi(a,b)| wu and palmer [ ] realized that the depth in the taxonomy will impact the similarity measure of two nodes, because the deeper of the nodes are in the tree, the semantic distance is smaller. therefore they gave a new measure of ontology: simw&p(a,b) = ×n n + n + ×n n and n is the numbers of is-a links from each term to their least common subsumer(lcs), n is the number of is-a links of the lcs to the root of the ontology. based on the same idea, leacock and chodorow [ ] also proposed a similarity measure that combined distance np between terms a and b and the depth d of the taxonomy. siml&c(a,b) = − log(np/ d) there are some limitations of path-based approaches. first, it only considers the shortest path between concept pairs. when they meet a complex situation like multiple taxonomical inheritance, the accuracy of them will be low. another problem of the path-based approaches is that they assume that all links in the taxonomy have uniform distance. . . feature-based measures feature based approaches assess the similarity between concepts as a function of their properties. they consider the degree of overlapping between sets of ontological features, like tversky’s model [ ], which subtracts the non-common features from common features of two concepts. simtve(a,b) = α ·f (Ψ(a) ∩ Ψ(b)) −β ·f (Ψ(a) \ Ψ(b)) −γ ·f (Ψ(b) \ Ψ(a)) where f is salience of a set features, and α,β and γ are weights of the contribution of each component. rodŕıguez and egenhofer [ ] computed similarity by summing the weighted sum of similarities between synsets, features, and neighbour concepts. simre(a,b) = w ·ssynsets(a,b) + u ·sfeatures(a,b) + v ·sneighborhoods(a,b) the feature-based methods consider more semantic knowledge than path-based methods. but only big ontologies/thesauri like wordnet [ ] have this kind of in- formation. ding et al. [ ] revealed that domain ontologies very occasionally model any semantic feature apart from taxonomical relationship. . . content-based measures other approaches want to overcome the limitations of edge-counting approach- es are content-based measures. resnik [ ] proposed a similarity measure, which depends on the amount of shared information between two terms: simres(a,b) = ic(lcs(a,b)) lcs is the least common subsumer of terms in a ontology, and ic is information content, which is the negative log of its probability of occurrence, p(a). lin [ ] and jiang and conrath [ ] extended resnik’s work. they also considered the ic of each of the evaluated terms, and they proposed that the similarity between two terms should be measured as the ratio between the amount of information needed to state their commonality and the information needed to fully describe them. simlin(a,b) = ×simres(a,b) (ic(a) + ic(b)) the are also two disadvantages of the content-based measures. first, the approaches cannot compute the concepts of leave nodes, because they don’t have subsumers. second, if the concepts do not have enough common subsumers, their similarities are hard to be calculated. . statistical-based ontology similarity measure in this thesis, we proposed a new statistical-based ontology similarity measure. in most job descriptions, they list many skills the positions required. from observation, we found that related skills always exist in the job description simultaneously, and the positions of them are always close, e.g. html and css are always required together, and appear in the same sentence. we could see this phenomenon in table . , which include some skill requirement sentences from some job descriptions. we can see from the table . , the closely related concepts are always have short distance. based on such observation, we give a new statistical-based ontology table . : some sentences of job descriptions . a high-level language such as java, groovy, ruby or python; we use java and groovy extensively . html /css /javascript, web standards, jquery or frameworks like an- gularjs would be great . html css and javascript a must . experience with ajax, xml, xsl, xslt, css, javascript, jquery, htm- l and web services similarity measure. if two concepts a and b have the same direct hypernym or one is the hypernym of the other, the similarity between them is given: s(a,b) = na∩b/na∪b avg(log (mindis(di,a,b) + )) the numerator is the ratio of the number of documents in which the two terms exist together (na∩b) and the number of documents have a least one of them (na∪b). the denominator is the average log value of minimum distance mindis(doc,a,b) of the two terms in documents that have them both. we set the restriction on the position of the two concepts in the ontology, because the position of the concepts in the ontology are based on their technical similarity to others. similar techniques will be assigned into the same category, so they should share the same hypernym, and one could be an alternative to the other. for example, we put ejb and hibernate in the same category, because they are both j ee persis- tence layer technologies, and both have the o/r mapping concept. if the applicant is familiar one of them, they can master the other very quickly. another example is grail and django, they are both web frameworks and share same web design philoso- phies, but one of them is designed for java web application and the other is created for python web application. if a developer has some some experience with one of them, he/she still need to spend a lot of time to learn the other to overcome the gap between programming languages. the algorithm to calculate the similarity of two concepts is in algorithm . input: docs term , term output: similarity total = ; hastwo = ; dislist = [ ]; for i = ; i ≤ len(docs); i + + do if docsi has at least one term then total + = ; if docsi has both terms then hastwo + = ; mindis = minimium distance (docsi, term , term ) ; dislist. add (log (mindis + )) ; end end end factor = hastwo / total ; factor = avg(dislist) ; return factor / factor ; algorithm : calculating statistical-based similarity the matrix in table . show the similarity values among of some skills, which is gotten from job descriptions. for example the skill html, the most relevant skills in order are css, javascript, and jquery, which is the same from the perspective of experienced developers. the other example is java, the most relevant skill in the matrix is jsp, which is also agree with the general technical knowledge. table . : similarities of skills list term java jdbc spring hibernate mysql oracle java . . . . . jdbc . . . . . spring . . . . . hibernate . . . . . mysql . . . . . oracle . . . . . table . : similarities of skills list term javascript jquery html css java python ruby jsp javascript . . . . . . . jquery . . . . . . . html . . . . . . . css . . . . . . . java . . . . . . . python . . . . . . . ruby . . . . . . . jsp . . . . . . . . evaluation in this section we will evaluate the performance of the key components of the system, the information extraction module, the ontology similarity module, and the overall system resulting from the combination of the components. . experiments of information extraction to evaluate the performance of the information extraction module, we extract sentence types through the use of sentence filters. to explain the process of our experiment, we use the sentences whose content pertains to the applicant’s college degree information. in the experiment, we selected sentences from existing job descriptions, and the content of these sentences were requirements of candidate degree and major. we labeled the values for ”degree” and ”major” manually. we use some content patterns that we can identify from these sentences to match and extract the degree information. when we used patterns, the accuracy of ”degree” became %. we also compared our pattern matching method to the conditional random fields (crfs) model [ ], which is a state of art machine learning model for sequence labeling. we used labelled sentences to train the crf model, and the features of the crf model are words in the sentences and part of speech tags of the words. the accuracies of information extraction of the three fields with our two methods, pattern matching, and the application of the crf model are shown in table . . . experiments of ontology similarity we selected some common skills from job descriptions; table . shows simi- larity values between these skills. higher values correspond to greater similarities, so table . : information extraction field pattern num accuracy of pattern matching accuracy of crf degree . . major . . our pattern matching approach can get higher accuracy. table . : similarities of skill list term java jdbc spring hibernate mysql oracle java . . . . . jdbc . . . . . spring . . . . . hibernate . . . . . mysql . . . . . oracle . . . . . the similarity between one skill and itself is . we selected one concept and ranked the other concepts by their similarity values to this concept. human judges helped rank these concepts by assigning them ”relevance scores” so that we can use ndcg to evaluate the effectiveness of our approach. we use the normalized discounted cumulative gain(ndcg) [ ] to evaluate the statistical-based similarity. ndcg is an important measure to evaluate the ranked retrieval results, which is the ratio of discounted cumulative gain ( dcg ) to ideal discounted cumulative gain ( idcg ). ndcg = dcg idcg dcg is the measure of how documents are ranked according to their relevance scores, and idcg is the dcg value that the documents are strictly sorted by their table . : javascript similarity evaluation term similarity value position relevance jquery . html . css . java . python . ruby . jsp . table . : html similarity evaluation term similarity value position relevance javascript . jquery . css . java . python . ruby . jsp . relevance values. dcg = p∑ i= reli − log (i + ) table . shows how we evaluate the similarity between the concept “javascript” and other concepts. the first column is a skill name, the second column is its simi- larity value to “javascript”, the third column is its position ranked by the similarity value, and the fourth column is its relevance value given by the judges. the nd- cg value for concept javascript is . , and in table . , ndcg value for concept html is . . . evaluation of the system in a traditional information retrieval systems, precision and ndcg are widely used measures [ ]. precision (p) is the fraction of retrieved documents that are relevant. precision = #(releveant items retrieved) #(retrieved items) we first used precision@k to compare the performance of our approach to the classical information retrieval models: okapi bm [ ], kullback-leibler divergence, and the tf-idf. precision@k is the proportion of relevant documents in the first k positions and is given below: p@k = k m∑ i= li (r(i) ≤ k) where is the indicator function: (a) = if a is true, otherwise. to evaluate a job finding, we compare the results of the system with three classi- cal information retrieval models: kullback-leibler divergence [ ], tf-idf [ ] and okapi bm [ ]. we give the definition of these measures below. kullback-leibler divergence is a non-symmetric measure of the difference between two probability distributions p and q. the score of a document d with respect to query q is given by: s(d,q) = −d(θq ‖ θd) = − ∑ ω∈v p(ω | θq) log p(ω|θq) p(ω|θd) = ∑ ω∈v p(ω | θq) log p(ω | θd) − ∑ ω∈v p(ω | θq) log p(ω | θq) ( . ) in the equation θq is the language model for a query, and θd is the language model for a document. tf-idf is a vector space model, which calculates the cosine similarity between the vectors of the query q and the document d. in the equation, tf is the term frequency, and idf is the inverse document frequency. the tf-idf weighting scheme assigned to term t a weight in document d given by: tf-idft,d = tft,d × idft the similarity between the query q and the document d given by: score(q,d) = ∑ t∈q tf-idft,d okapi bm is a bag-of-words retrieval model that ranks a set of documents based on the query terms appearing in each document. given a query q, the bm score of a document d is: score(q,d) = ∑ t∈q idft tft,d · (k + ) tft,d + k · ( − b + b · |d| avgdl ) in the formula, idft is idf value of term t, tf is the term frequency; |d| is the length of the document d; avgdl is the average length of all the documents, and k and b are the free parameters. in the evaluation phase, we created a data set of job descriptions that includes several kinds of jobs such as web developers, server back-end developers, mobile developers and so on. we used candidate résumés and retrieved the top jobs. the relevance value of the job descriptions to each résumé were set manually by ten human judge. we created a query q from the résumé, treated the text of the job descriptions as documents d, and applied standard ad-hoc retrieval techniques to rank the jobs. we intended to return jobs that better matched the candidates’ résumés at the top. the result of precision@k is in table . . table . : precision of job ranking k okapi bm kl tf-idf résumatcher . . . . . . . . . . . . résumatcher get the highest precision. the other measure we used is ndcg, which is explained in last section. table . shows the ndcg value get from different information retrieval models. the result shows that ontology similarity performs the best. table . : ndcg of job ranking k okapi bm kl tf-idf résumatcher . . . . . . . . . . . . our sytem get the highest ndcg value. . comparing with the keyword searching we also made experiments to compare the quality of search results quality be- tween our system and the keyword searching approach. in one experiment, we first selected a résumé, and picked a related keyword e.g. “java” to search jobs in in- deed.com. we used this process to simulate how a job seeker searches jobs on in- deed.com. the top jobs returned by indeed.com were saved in our system, then we used the résumé to match the jobs. we compared the quality of the approaches by calculating the measures precision@k and dcg. we use dcg not ndcg because we compare the two approaches on the same data set, and the idcg is the same. five judges gave the relevance values to the résumés and jobs, and we used average values. table . shows comparison between the two approaches by using keyword “java” and the résumé of a java developer. the comparison for keyword “python” and the résumé of a python developer is shown in table . , and the average comparison for five keywords and five résumé is shown in table . . we can see from the result that our system can get much better results than keyword searching approach. table . : comparison of the two approaches - java developer k precision@k dcg indeed résumatcher indeed résumatcher . . . . . . . . . . . . résumatcher can get better result. table . : comparison of the two approaches - python developer k precision@k dcg indeed résumatcher indeed résumatcher . . . . . . . . . . . . résumatcher can get better result. table . : comparison of the two approaches - average k precision@k dcg indeed résumatcher indeed résumatcher . . . . . . . . . . . . résumatcher can get better result. . conclusion and future work . conclusion in this thesis we presented résumatcher, a personalized job-résumé matching system that can help job seekers find appropriate jobs faster and more accurately by using their résumé contents. the key components of the system are the information extraction procedure and the ontology matching module. in the system, job descriptions and résumés are parsed into job models and résumé models by the information extraction module. when searching the jobs by a résumé, similarity values between the résumé model and job description models are calculated in the ontology matching module. the result is sorted by the ontology similarity scores, which are the sum of similarities of different fields multiplied by their weights. we made such contributions in our works: . we developed a résumé - job matching system. . we developed a finite state transducer based matching tool to extract informa- tion from unstructured data source, which is a lightweight and flexible library, and can be extended in very easy ways. . we developed a semi-automatic approach, which can collect technical terms from hr data sources, and by which we created a domain specific ontology for recruitment. . we developed statistical-based ontology similarity measure, which can measure the similarities between technical terms . in the experiment phase, we evaluated the accuracy of information extraction. we calculated the ontology similarity with the ndcg. finally, we also tested the performance of résumé job matching algorithm via precision@k and ndcg, which showed that our algorithm can achieve a better searching result than other informa- tion retrieval models like tf-idf and okpai bm . we also compared our system with the commercial job search engine www.indeed.com, and the results showed that our system can return jobs with a better ranking. . future work finding a job is a complex process, affected by both explicit and implicit factors. our work establishes the validity of using information extraction techniques to create a more personalized job matching system, with ample potential for improvement in the future. first we can introduce a more complex job and rsum model to improve perfor- mance of the system. in the résumé model, we can consider hiring history and project experience of the job seekers. to improve the job description model, job responsi- bilities and company characteristics (size, dress code, etc.) should be considered as well. second, to improve searching speed of our system, we can reduce the the number of comparison by filtering out jobs that are clearly not related to résumés. the system can classify the jobs into some different subsets, when searching jobs, the system only need to calculate the similarity between the résumé and according subset of jobs. résumatcher is a content based recommendation system that is mostly focused on comparing the similarities between the résumé and a relevant job description. in future work, we could introduce a hybrid recommendation system that would take advantage of other recommendation algorithms such as collaborative filtering. future work on this system would place greater consideration on job seeker’s personal preference like job location, career development plan, and company background. references [ ] acm. acm computing classification system, . [ ] alfred v aho and jeffrey d ullman. foundations of computer science, volume . computer science press new york, . [ ] shaha t al-otaibi and mourad ykhlef. a survey of job recommender systems. international journal of the physical sciences, ( ): – , . [ ] douglas e appelt and boyan onyshkevych. the common pattern specification language. in proceedings of a workshop on held at baltimore, maryland: october - , , pages – . association for computational linguistics, . [ ] steven bird. nltk: the natural language toolkit. in proceedings of the col- ing/acl on interactive presentation sessions, pages – . association for computational linguistics, . [ ] christian bizer, jens lehmann, georgi kobilarov, sören auer, christian becker, richard cyganiak, and sebastian hellmann. dbpedia-a crystallization point for the web of data. web semantics: science, services and agents on the world wide web, ( ): – , . [ ] duygu çelik and atilla elçi. an ontology-based information extraction approach for résumés. in pervasive computing and the networked world, pages – . springer, . [ ] angel x chang and christopher d manning. tokensregex: defining cascaded regular expressions over tokens. technical report, technical report cstr - , department of computer science, stanford university, . [ ] hamish cunningham, diana maynard, kalina bontcheva, and valentin tablan. a framework and graphical development environment for robust nlp tools and applications. in acl, pages – , . [ ] j olawande daramola, olufunke o oladipupo, and ag musa. a fuzzy expert system (fes) tool for online personnel recruitments. international journal of business information systems, ( ): – , . [ ] li ding, tim finin, anupam joshi, rong pan, r scott cost, yun peng, pavan reddivari, vishal doshi, and joel sachs. swoogle: a search and metadata engine for the semantic web. in proceedings of the thirteenth acm international conference on information and knowledge management, pages – . acm, . [ ] ying ding and schubert foo. ontology research and development. part -a review of ontology generation. journal of information science, ( ): – , . [ ] athanasios drigas, stelios kouremenos, spyros vrettos, john vrettaros, and dimitris kouremenos. an expert system for job matching of the unemployed. expert systems with applications, ( ): – , . [ ] frank färber, tim weitzel, and tobias keim. an automated recommendation approach to selection in personnel recruitment. . [ ] maryam fazel-zarandi and mark s fox. semantic matchmaking for job recruit- ment: an ontology-based hybrid approach. in proceedings of the th interna- tional semantic web conference, . [ ] christiane fellbaum. wordnet. wiley online library, . [ ] tere gonzalez, pano santos, fernando orozco, mildreth alcaraz, victor zal- divar, alberto de obeso, and alan garcia. adaptive employee profile classifica- tion for resource planning tool. in srii global conference (srii), annual, pages – . ieee, . [ ] hal g gueutal and dianna l stone. the brave new world of ehr. john wiley & sons, . [ ] jerry r hobbs, douglas appelt, john bear, david israel, megumi kameyama, mark stickel, and mabry tyson. fastus: a cascaded finite-state transducer for extracting information from natural-language text. finite-state language processing, page , . [ ] jay j jiang and david w conrath. semantic similarity based on corpus statistics and lexical taxonomy. arxiv preprint cmp-lg/ , . [ ] v senthil kumaran and a sankar. towards an automated system for intelli- gent screening of candidates for recruitment using ontology mapping (expert). international journal of metadata, semantics and ontologies, ( ): – , . [ ] john lafferty, andrew mccallum, and fernando cn pereira. conditional ran- dom fields: probabilistic models for segmenting and labeling sequence data. . [ ] claudia leacock and martin chodorow. combining local context and word- net similarity for word sense identification. wordnet: an electronic lexical database, ( ): – , . [ ] danielle h lee and peter brusilovsky. fighting information overflow with per- sonalized comprehensive information access: a proactive job recommender. in autonomic and autonomous systems, . icas . third international con- ference on, pages – . ieee, . [ ] wendy lehnert, claire cardie, david fisher, ellen riloff, and robert williams. university of massachusetts: description of the circus system as used for muc- . in proceedings of the rd conference on message understanding, pages – . association for computational linguistics, . [ ] dekang lin. an information-theoretic definition of similarity. in icml, vol- ume , pages – , . [ ] ping liu and peter dew. using semantic web technologies to improve expertise matching within academia. proceedings of i-know (graz, austria, pages – , . [ ] yao lu, sandy el helou, and denis gillet. a recommender system for job seek- ing and recruiting website. in proceedings of the nd international conference on world wide web companion, pages – . international world wide web conferences steering committee, . [ ] jochen malinowski, tobias keim, oliver wendt, and tim weitzel. matching people and jobs: a bilateral recommendation approach. in system sciences, . hicss’ . proceedings of the th annual hawaii international confer- ence on, volume , pages c– c. ieee, . [ ] christopher d manning, prabhakar raghavan, and hinrich schütze. introduc- tion to information retrieval, volume . cambridge university press cambridge, . [ ] george a miller. wordnet: a lexical database for english. communications of the acm, ( ): – , . [ ] natalya f noy, michael sintek, stefan decker, monica crubézy, ray w ferg- erson, and mark a musen. creating semantic web contents with protege- . ieee intelligent systems, ( ): – , . [ ] lawrence page, sergey brin, rajeev motwani, and terry winograd. the pager- ank citation ranking: bringing order to the web. . [ ] roy rada, hafedh mili, ellen bicknell, and maria blettner. development and application of a metric on semantic nets. systems, man and cybernetics, ieee transactions on, ( ): – , . [ ] rachael rafter, keith bradley, and barry smyth. personalised retrieval for online recruitment services. in proceedings of the nd annual colloquium on information retrieval. ., . [ ] philip resnik. using information content to evaluate semantic similarity in a taxonomy. arxiv preprint cmp-lg/ , . [ ] stephen robertson and hugo zaragoza. the probabilistic relevance framework: bm and beyond. information retrieval, ( ): – , . [ ] stephen e robertson, steve walker, susan jones, micheline m hancock- beaulieu, mike gatford, et al. okapi at trec- . nist special publication sp, pages – , . [ ] emmanuel roche and yves schabes. finite-state language processing. mit press, . [ ] m andrea rodŕıguez and max j. egenhofer. determining semantic similarity among entity classes from different ontologies. knowledge and data engineering, ieee transactions on, ( ): – , . [ ] david sánchez, montserrat batet, david isern, and aida valls. ontology-based semantic similarity: a new feature-based approach. expert systems with appli- cations, ( ): – , . [ ] sunita sarawagi. information extraction. foundations and trends in databases, ( ): – , . [ ] amit singh, catherine rose, karthik visweswariah, vijil chenthamarakshan, and nandakishore kambhatla. prospect: a system for screening candidates for recruitment. in proceedings of the th acm international conference on information and knowledge management, pages – . acm, . [ ] zheng siting, hong wenxing, zhang ning, and yang fan. job recommender systems: a survey. in computer science & education (iccse), th in- ternational conference on, pages – . ieee, . [ ] ken thompson. programming techniques: regular expression search algorithm. communications of the acm, ( ): – , . [ ] amos tversky. features of similarity. . [ ] andrew j viterbi. error bounds for convolutional codes and an asymptotical- ly optimum decoding algorithm. information theory, ieee transactions on, ( ): – , . [ ] kangning wei, jinghua huang, and shaohong fu. a survey of e-commerce recommender systems. in service systems and service management, in- ternational conference on, pages – . ieee, . [ ] zhibiao wu and martha palmer. verbs semantics and lexical selection. in proceedings of the nd annual meeting on association for computational lin- guistics, pages – . association for computational linguistics, . [ ] xing yi, james allan, and w bruce croft. matching resumes and jobs based on relevance models. in proceedings of the th annual international acm sigir conference on research and development in information retrieval, pages – . acm, . [ ] kun yu, gang guan, and ming zhou. resume information extraction with cas- caded hybrid model. in proceedings of the rd annual meeting on association for computational linguistics, pages – . association for computational linguistics, . [ ] chengxiang zhai. statistical language models for information retrieval. syn- thesis lectures on human language technologies, ( ): – , . [ ] justin zobel and alistair moffat. inverted files for text search engines. acm computing surveys (csur), ( ): , . using possibility theory in expert systems by prakash p. shenoy school of business, university of kansas, lawrence, ks - , usa tel: ( )- - , fax: ( )- - , email: pshenoy@ku.edu abstract this paper has two main objectives. the first objective is to give a characterization of a qualitative description of a possibility function. a qualitative description of a possibility function is called a consistent possibilistic state. the qualitative descrip- tion and its characterization serve as qualitative semantics for possibility functions. these semantics are useful in representing knowledge as possibility functions. the second objective is to describe how zadeh’s theory of possibility fits in the frame- work of valuation-based systems (vbs). since vbs serve as a framework for managing uncertainty and imprecision in expert systems, this facilitates the use of possibility theory in expert systems. key words: possibility theory, consistent possibilistic state, valuation-based systems, expert systems . introduction the main goal of this paper is to describe how zadeh’s theory of possibility can be used for man- aging uncertainty and imprecision in expert systems. we give a characterization of a qualitative de- scription of a possibility function. this is a useful first step in representing knowledge as possibil- ity functions. also, we describe how possibility theory fits in the framework of valuation-based systems (vbs). since vbs serve as a framework for managing uncertainty and imprecision in ex- pert systems, this facilitates the use of possibility theory in expert systems. the theory of possibility was first described by zadeh [ , ] and further developed by dubois and prade [ ]. in possibility theory, the basic representational unit is called a possibility function. the two main operations for manipulating possibility functions are called projection and particularization [zadeh ]. in this paper, we are concerned with two issues related to using possibility theory in expert systems. the first issue is in representing knowledge as possibility functions. to do this effec- tively, we need semantics for possibility functions. what does it mean, for example, to say that appeared in: fuzzy sets and systems, vol. , issue , dec. , pp. -- . using possibility theory in expert systems possibility of a proposition is . ? in probability theory, for example, the probability of a proposi- tion can be interpreted as a relative frequency or as a betting rate. with this goal in mind, we first define a qualitative description of a possibility function called a consistent possibilistic state. this is analogous to the definition of a consistent epistemic state [shenoy a] in spohn’s theory of epistemic beliefs [spohn ]. next, we give a characterization of a consistent possibilistic state in terms of what we call the content of a consistent possibilistic state. again, this characterization is analogous to spohn’s characterization of a consistent epistemic state in terms of its content [spohn ]. this is only one step toward developing semantics for possibility functions. semantics for the quantitative aspects of possibility functions have been studied by, for example, dubois et al. [ ] and ruspini [ ]. the second issue related to using possibility theory in expert systems is making the correspon- dence between concepts in possibility theory and concepts in expert systems. if we interpret pos- sibility functions as knowledge, what do the projection and particularization operations mean in the context of expert systems? how do we use these operations to make inferences in a knowledge- based system? also, since expert systems typically deal with many propositions, there is a compu- tational issue of the tractability of using possibility theory to solve large problems. to help answer these questions, we describe how possibility theory fits in the framework of vbs. the framework of vbs was first defined by shenoy [ ] as a general language for incorpo- rating uncertainty in expert systems. it was further elaborated in [shenoy c] to include axioms that permit local computation in solving a vbs, and a fusion algorithm for solving a vbs using lo- cal computation. vbs encode knowledge using functions defined on frames of variables. the functions are called valuations. vbs include two operators called combination and marginalization that operate on valuations. combination corresponds to aggregation of knowledge. marginalization corresponds to coarsening of knowledge. the process of reasoning in vbs can be described sim- ply as finding the marginal of the joint valuation for each variable in the system. the joint valuation is the valuation obtained by combining all valuations. in systems with many variables, it is compu- tationally intractable to explicitly compute the joint valuation. however, if combination and marginalization satisfy certain axioms, it is possible to compute the marginals of the joint valuation without explicitly computing the joint valuation. the framework of vbs is general enough to represent many domains such as bayesian prob- ability theory [shenoy c], dempster-shafer’s theory of belief functions [dempster , shafer , shenoy d], spohn’s theory of epistemic beliefs [spohn , shenoy c], discrete optimization [shenoy b], propositional logic [shenoy a, b], constraint satisfaction [shenoy and shafer ], and bayesian decision theory [shenoy c, d]. saffiotti and umkehrer [ ] describe an efficient implementation of vbs called pulcinella. the correspondence between possibility theory and vbs is as follows. the particularization operation in possibility theory corresponds to the combination operation in vbs. and the projec- tion operation in possibility theory corresponds to the marginalization operation in vbs. this cor- introduction respondence was first pointed out explicitly by dubois and prade [ ]. they also state that the combination and marginalization operations in possibility theory satisfy the three axioms that en- able the use of local computation in computing marginals of the joint possibility function. in this paper, we reiterate this correspondence in greater detail. an outline of this paper is as follows. section describes the framework of vbs. section contains an axiomatic definition and a characterization of a consistent possibilistic state. section describes the main features of possibility theory in terms of the framework of vbs. section de- scribes a small example illustrating the use of possibility theory for managing uncertainty and im- precision in expert systems. section contains some concluding remarks. finally, section con- tains proofs of all results in the paper. . valuation-based systems in this section, we will sketch the basic features of vbs. also, we describe three axioms that permit the use of local computation, and describe a fusion algorithm for solving a vbs using local computation. . . the framework this subsection describes the framework of valuation-based systems. in a vbs, we represent knowledge by functions called valuations. valuations are functions that assign values to the ele- ments of frames for sets of variables. we make inferences in a vbs using two operators called combination and marginalization. we use these operators on valuations. variables and configurations. we use the symbol wx for the set of possible values of a variable x, and we call wx the frame for x. we assume that one and only one of the elements of wx is the true value of x. we are concerned with a finite set x of variables, and we assume that all the variables in x have finite frames. given a nonempty set s of variables, let ws denote the cartesian product of wx for x in s; ws = ×{wx | x∈s}. we call ws the frame for s. we call the elements of ws configurations of s. valuations. given a subset s of variables, there is a set vs. we call the elements of vs val- uations for s. let v denote the set of all valuations, i.e., v = ∪{vs | s⊆x}. if σ is a valuation for s, then we say s is the domain of σ. valuations are primitives in our abstract framework and as such require no definition. but as we shall see shortly, they are objects which can be combined and marginalized. intuitively, a val- uation for s represents some knowledge about the variables in s. nonzero valuations. for each s⊆x, there is a nonempty subset ps of vs whose elements are called nonzero valuations for s. let p denote ∪{ps | s⊆x}, the set of all nonzero valuations. intuitively, a nonzero valuation represents knowledge that is internally consistent. the notion of nonzero valuations is important as it enables us to define combinability of valuations, and it using possibility theory in expert systems allows us to constrain the definitions of combination and marginalization to meaningful operators. an example of a nonzero valuation is a possibility function. combination. we assume there is a mapping ⊗:v×v → v, called combination, such that (i) if ρ and σ are valuations for r and s, respectively, then ρ⊗σ is a valuation for r∪s; (ii) if either ρ or σ is not a nonzero valuation, then ρ⊗σ is not a nonzero valuation; and (iii) if ρ and σ are both nonzero valuations, then ρ⊗σ may or may not be a nonzero valuation. we call ρ⊗σ the combination of ρ and σ. intuitively, combination corresponds to aggregation of knowledge. if ρ and σ are valuations for r and s representing knowledge about variables in r and s, respectively, then ρ⊗σ represents the aggregated knowledge about variables in r∪s. for possibility functions, combination is pointwise multiplication [zadeh ]. marginalization. we assume that for each s⊆x, and for each x∈s, there is a mapping ↓(s−{x}): vs → vs–{x}, called marginalization to s–{x} such that (i) if σ is a valuation for s, then σ↓(s–{x}) is a valuation for s–{x}; and (ii) σ↓(s–{x}) is a nonzero valuation if and only if σ is a nonzero valuation. we call σ↓(s–{x}) the marginal of σ for s–{x}. intuitively, marginalization corresponds to coarsening of knowledge. if σ is a valuation for s representing some knowledge about variables in s, and x∈s, then σ↓(s–{x}) represents the knowl- edge about variables in s–{x} implied by σ if we disregard variable x. in the case of possibility functions, marginalization from s to s–{x} is maximization over the frame for x [zadeh ]. in summary, a valuation-based system consists of a -tuple {{σ , ..., σm}, ⊗, ↓} where {σ , ..., σm} is a collection of valuations, ⊗ is the combination operator, and ↓ is the marginalization operator. valuation networks. a graphical depiction of a valuation-based system is called a valuation network. in a valuation network, variables are represented by circular nodes, and valuations are represented by diamond-shaped nodes. also, each valuation node is connected by an undirected edge to each variable node in its domain. figure shows a valuation network for a vbs that consists of valuations σ for {w}, σ for {w, x}, σ for {x, y}, and σ for {y, z}. making inference in vbs. in a vbs, the combination of all valuations is called the joint valuation. given a vbs, we make inferences by computing the marginal of the joint valuation for each variable in the system. valuation-based systems figure . a valuation network for a vbs consisting of valuations σ for {w}, σ for {w, x}, σ for {x, y}, and σ for {y, z}. w x y z σ σ σ σ if there are n variables in the system, and each variable has two configurations in its frame, then there are n configurations of all variables. hence, it is not computationally feasible to com- pute the joint valuation when there are a large number of variables. in section . , we describe an algorithm for computing marginals of the joint valuation without explicitly computing the joint val- uation. so that this algorithm gives us the correct answers, we require combination and marginal- ization to satisfy three axioms. the axioms and the algorithm are described in the next two subsec- tions, respectively. . . axioms in this section, we state three simple axioms that enable local computation of marginals of the joint valuation. these axioms were first formulated by shenoy and shafer [ ]. the axioms are stated slightly differently here. axiom a (commutativity and associativity of combination): suppose ρ, σ, and τ are valuations for r, s, and t respectively. then ρ⊗σ = σ⊗ρ, and ρ⊗(σ⊗τ) = (ρ⊗σ)⊗τ. axiom a (order of deletion does not matter): suppose σ is a valuation for s, and suppose x , x ∈ s. then (σ↓(s–{x }))↓(s–{x ,x }) = (σ↓(s–{x }))↓(s–{x ,x }). axiom a (distributivity of marginalization over combination): suppose ρ and σ are valuations for r and s, respectively. suppose x∉r, and suppose x∈s. then (ρ⊗σ)↓((r∪s)–{x}) = ρ⊗(σ↓(s–{x})). one implication of axiom a is that when we have multiple combinations of valuations, we can write it without using parenthesis. for example, (...((σ ⊗σ )⊗σ )⊗...⊗σm) can be written simply as ⊗{σi | i = , ..., m} or as σ ⊗...⊗σm, i.e., we need not indicate the order in which the combinations are carried out. using possibility theory in expert systems if we regard marginalization as a coarsening of a valuation by deleting variables, then axiom a says that the order in which the variables are deleted does not matter. one implication of this axiom is that (σ↓(s–{x }))↓(s–{x ,x }) can be written simply as σ↓(s–{x ,x }), i.e., we need not indi- cate the order in which the variables are deleted. axiom a is the crucial axiom that makes local computation possible. axiom a states that the computation of (ρ⊗σ)↓((r∪s)–{x}) can be accomplished without having to compute ρ⊗σ. the combination operation in ρ⊗σ is on the frame for r∪s whereas the combination operation in ρ⊗(σ↓(s–{x})) is on the frame for (r∪s)–{x}. in the next subsection, we describe a fusion algo- rithm that applies this axiom repeatedly resulting in an efficient method for computing marginals. . . a fusion algorithm in this subsection, we describe a fusion algorithm for making inferences in a vbs using local computation. suppose {{σ , ..., σm}, ⊗, ↓} is a vbs with n variables and m valuations. suppose that combination and marginalization satisfy the three axioms stated in section . . suppose we have to compute the marginal of the joint valuation for variable x, (σ ⊗...⊗σm) ↓{x}. the basic idea of the fusion algorithm is to successively delete all variables but x from the vbs. the variables may be deleted in any sequence. axiom a tells us that all deletion sequences lead to the same answers. but, different deletion sequences may involve different computational costs. we will comment on good deletion sequences at the end of this section. when we delete a variable, we have to do a “fusion” operation on the valuations. consider a set of k valuations ρ , ..., ρk. suppose ρi is a valuation for ri. let fusx{ρ , ..., ρk} denote the collection of valuations after fusing the valuations in the set {ρ , ..., ρk} with respect to variable x. then fusx{ρ , ..., ρk} = {ρ ↓(r−{x})}∪{ρi | x∉ri} where ρ = ⊗{ρi | x∈ri}, and r = ∪{ri | x∈ri}. after fusion, the set of valuations is changed as follows. all valuations that bear on x are combined, and the resulting valuation is marginalized such that x is eliminated from its domain. the valuations that do not bear on x remain unchanged. we are ready to state the main theorem that describes the fusion algorithm theorem [shenoy c]. suppose {{σ , ..., σm}, ⊗, ↓} is a vbs where σi is a valuation for si, and suppose ⊗ and ↓ satisfy axioms a –a . let x denote s ∪...∪sm. suppose x∈x, and suppose x x ...xn– is a sequence of variables in x–{x}. then (σ ⊗...⊗σm) ↓{x} = ⊗{fusxn– {... fusx {fusx {σ , ..., σm}}}}. to illustrate theorem , consider the vbs shown in figure . suppose we need to compute the marginal of the joint for z, (σ ⊗...⊗σ ) ↓{z}. consider the deletion sequence wxy. after valuation-based systems fusion with respect to w, we have (σ ⊗σ ) ↓{x} for {x}, σ for {x, y}, and σ for {y, z}. after fusion with respect to x, we have ((σ ⊗σ ) ↓{x}⊗σ ) ↓{y} for {y}, and σ for {y, z}. finally, after fusion with respect to y, we have (((σ ⊗σ )↓{x}⊗σ )↓{y}⊗σ )↓{z} for {z}. theorem tells us that (((σ ⊗σ )↓{x}⊗σ )↓{y}⊗σ )↓{z} = (σ ⊗...⊗σ )↓{z}. thus, instead of doing combinations on the frame for {w, x, y, z}, we do combinations on the frame for {w, x}, {x, y}, and {y, z}. the fusion algorithm is shown graphically in figure . figure . the first valuation network shows the initial vbs. the second network is the result after fusion with respect to w. the third network is the result after fusion with respect to x. the fourth network is the result after fusion with respect to y. w x y z σ x y z y z z σ σ σ σ σ σ (σ ⊗σ ) ↓{x} ((σ ⊗σ ) ↓{x}⊗σ ) ↓{y} (((σ ⊗σ )↓{x}⊗σ )↓{y}⊗σ )↓{z} if we can compute the marginal of the joint valuation for one variable, then we can compute the marginals for all variables. we simply compute them one after the other. it is obvious, however, using possibility theory in expert systems that this will involve much duplication of effort. shenoy and shafer [ ] describe an efficient algorithm for simultaneous computation of all marginals without duplication of effort. regardless of the number of variables in a vbs, we can compute marginals of the joint valuation for all vari- ables for roughly three times the computational effort required to compute one marginal [shenoy and shafer ]. deletion sequences. different deletion sequences may involve different computational ef- forts. for example, consider the vbs in the above example. in this example, deletion sequence wxy involves less computational effort than, for example, xyw, as the former involves combi- nations on the frame for two variables only whereas the latter involves combination on the frame for three variables. finding an optimal deletion sequence is a secondary optimization problem that has shown to be np-complete [arnborg et al. ]. but, there are several heuristics for finding good deletion sequences [kong , mellouli , zhang , kjærulff ]. one such heuristic is called one-step-look-ahead [kong ]. this heuristic tells us which variable to delete next. as per this heuristic, the variable that should be deleted next is one that leads to combination over the smallest frame. for example, in the vbs described above, if we as- sume that each variable has a frame consisting of two configurations, then this heuristic would pick w over x and y for first deletion since deletion of w involves combination on the frame for {w, x} whereas deletion of x involves combination on the frame for {w, x, y}, and deletion of y in- volves combination on the frame for {x, y, z}. after w is deleted, for second deletion, this heuristic would pick x over y. thus, this heuristic would choose deletion sequence wxy. . consistent possibilistic states in this section, first we define axiomatically a consistent possibilistic state. a consistent pos- sibilistic state serves as a qualitative description of a possibility function. the definition of a con- sistent possibilistic state is analogous to the definition of a consistent epistemic state in spohn’s theory of epistemic beliefs [shenoy a, spohn ]. next, we give a characterization of a consistent possibilistic state. again, this characterization is analogous to the characterization of a consistent epistemic state given by spohn [ ]. before we define a consistent possibilistic state, we need to establish our notation. consider a variable x with frame wx. suppose wx is defined such that the propositions re- garding x that are of interest are precisely those of the form ‘the true value of x is in s,’ where s is a subset of wx. thus the propositions regarding x that are of interest are in a one-to-one corre- spondence with the subsets of wx [shafer , p. ]. the correspondence between subsets and propositions is useful since it translates the logical notions of conjunction, disjunction, implication, and negation into the set-theoretic notions of inter- section, union, inclusion, and complementation [shafer , pp. - ]. thus, if r and s are two subsets of wx, and r' and s' are the corresponding propositions, then r∩s corresponds to the consistent possibilistic states conjunction of r' and s', r∪s corresponds to the disjunction of r' and s', r⊆s if and only if r' im- plies s', and r is the set-theoretic complement of s with respect to wx (written as r = ~s) if and only if r' is the negation of s'. notice also that the proposition that corresponds to ∅ is false, and the proposition that corresponds to wx is true. henceforth, we will simply refer to a proposition by its corresponding subset. the set of subsets of wx will be denoted by wx. in a possibilistic state for x, propositions are said to be either possibly true (or simply, possi- ble) or possibly false (or simply, not possible). thus a possibilistic state for x is an assignment of labels ‘possible’ and ‘not possible’ to each proposition in wx. logical consistency requires that a possibilistic state satisfy certain conditions (axioms). a definition of a consistent possibilistic state is as follows. definition . a possibilistic state for x is said to be consistent if the following four axioms are satisfied: b . for any proposition s, one and only one of the following conditions holds: (i) s is possible. (ii) s is not possible. b . wx is possible, and ∅ is not possible. b . if s is not possible and r⊆s, then r is not possible. b . if r and s are not possible, then r∪s is not possible. some simple consequences of definition are as follows. proposition . the following conditions always hold in any consistent possi- bilistic state: b . if s is possible and r⊇s, then r is possible. b . if s is not possible, then ~s is possible. b . if s is possible, then this does not necessarily imply that ~s is not possible; ~s could also be possible. it is clear from axiom b that we can specify a consistent possibilistic state simply by listing all propositions that are not possible. then axiom b tells us that the remaining propositions are pos- sible. let i denote the set of all propositions (subsets) that are not possible. theorem gives a characterization of i. theorem . suppose i denotes the set of all propositions that are not possible in some possibilistic state for x. then the possibilistic state is consistent if and only if there exists a unique proper subset c of wx such that i = {s∈ wx | s ⊆ c}. subset c in theorem is called the content of the consistent possibilistic state. note that the content constitutes a complete specification of a possibilistic state. thus a simple corollary of using possibility theory in expert systems theorem is that in a frame wx consisting of n elements, there are exactly n− distinct possible consistent possibilistic states (corresponding to each proper subset of w as the content). the con- tent c represents the largest subset (proposition) that is not possible. example . (consistent possibilistic states) consider a frame w = {x, y, z}. table lists all seven possible consistent possibilistic states. possibilistic state represents a state of complete ig- norance in which every proposition (except the null proposition) is possible. possibilistic states , , and represent a state of complete information where one and only one element of the frame is possibly true. possibilistic states , , and are states of partial information where one element of the frame is not possible table . the seven consistent possibilistic states for a variable with frame w = {x,y,z}. state content c possible not possible i ∅ {x}, {y}, {z}, {x, y}, {y, z}, {x, z}, {x, y, z} ∅ {x} {y}, {z}, {x, y}, {x, z}, {y, z}, {x, y, z} ∅, {x} {y} {x}, {z}, {x, y}, {x, z}, {y, z}, {x, y, z} ∅, {y} {z} {x}, {y}, {x, y}, {x, z}, {y, z}, {x, y, z} ∅, {z} {x, y} {z}, {x, z}, {y, z}, {x, y, z} ∅, {x}, {y}, {x, y} {x, z} {y}, {x, y}, {y, z}, {x, y, z} ∅, {x}, {z}, {x, z} {y, z} {x}, {x, y}, {x, z}, {x, y, z} ∅, {y}, {z}, {y, z} . possibility theory in this section, we describe the main features of zadeh’s theory of possibility in terms of the framework of vbs described in the previous section, i.e., we will use the terminology of vbs in- stead of the terminology used in [zadeh ]. thus the operation zadeh calls particularization we call combination, and the operation zadeh calls projection we call marginalization. the basic unit of knowledge representation is called a possibility function. definition . a possibility function π for p is a function π: wp → [ , ] such that p . there exists a configuration w∈wp such that π({w}) = ; p . for any r∈( wg − {∅}), π(r) = max{π({w}) | w∈r}; and p . π(∅) = . possibility theory note that although a possibility function is defined for the set of all subsets of wp, it is com- pletely specified by its values for each singleton subset of wp. a possibility function is a complete representation of a consistent possibilistic state. to see what propositions are possible and what propositions are not possible in state π, consider the sub- set c = {w∈wp | π({w}) < }. by condition p in definition , c is always a proper subset of wp. c represents the content of the possibilistic state π, i.e., r is not possible in state π iff r ⊆ c. thus r is possible in state π iff π(r) = , and r is not possible in state π iff π(r) < . a possibility function consists of more than a representation of a consistent possibilistic state. it also includes degrees to which proposition are possible and degrees to which propositions are not possible. π(r) can be interpreted as the degree to which proposition r is possible, and −π(r) can be interpreted as the degree to which proposition r is not possible, i.e., r is more possible than t if π(r) > π(t) and conversely, r is more impossible than t if π(r) < π(t) < . consider the following possibility function for s: π({w}) = for all w∈ws. this means that the only proposition that is not possible is ∅ (and all other propositions are possible). we shall call such a possibility function vacuous. it represents a state of complete ignorance. proposition . suppose π is a possibility function for p. then p . for each r ∈ wp, either π(r) = or π(~r) = or both. p . for each r,t∈ wp, π(r∪t) = max{π(r), π(t)} extension of subsets. by extension of a subset of a frame to a subset of a larger frame, we mean a cylinder set extension. if r and s are sets of variables, r⊆s, and r is a subset of wr, then the extension of r to s is r×ws−r. we will let r↑s denote the extension of r to s. for example, consider three variables r, p, q with frames wr = {r, ~r}, wp = {p, ~p}, and wq = {q, ~q}, respectively. then the extension of {(r,~p), (~r,p)} (which is a subset of w{r,p}) to {r,p,q} is {(r,~p,q), (r,~p,~q), (~r,p,q), (~r,p,~q)}. note that the propositions corresponding to r and r↑s are logically equivalent. marginalization. suppose π is a possibility function for p. suppose q ⊆ p. we may be in- terested only in propositions about variables in q. in this case, we would like to marginalize π to q. the following definition of marginalization is motivated by the fact that each proposition q∈ wq about variables in q can be regarded as a proposition q↑p ∈ wp about variables in p. definition (marginalization). suppose π is a possibility function for p, and suppose q ⊆ p. the marginal of π for q, denoted by π↓q, is a possibility function for q given as follows: π↓q(q) = π(q↑p) = max{π({(x,y)}) | x∈q, y∈wp−q} ( ) for all q∈ wq. using possibility theory in expert systems in particular, if q is a singleton subset, i.e., q = {x} for some x∈wq, then ( ) simplifies to: π↓q({x}) = max{π({(x,y)}) | y∈wp−q}. note that if π is a possibility function for p, and x , x ∈ p, then (π ↓(p–{x }))↓(p–{x ,x }) = (π↓(p–{x }))↓(p–{x ,x }). in words, if we regard marginalization as reduction of π by deletion of variables, then the order in which the variables are deleted makes no difference in the final answer. example . (possibility function and marginalization) we would like to determine whether a stranger (about who we know nothing about) is a pacifist or not depending on whether he is a republican or not and whether he is a quaker or not. consider three variables r, q and p. r has two configurations: r (for republican), ~r (not republican); q has two configurations: q (quaker) and ~q (not quaker); and p has two configurations: p (pacifist) and ~p (not pacifist). our knowledge that most (at least percent) republicans are not pacifists and that most (at least percent) quakers are pacifists is represented by the possibility function π for {r,q,p} shown in table (the construction of this possibility function will be explained later in this section—see example ). note that the marginal of π for r is the vacuous possibility function for r, i.e., π↓{r}({r}) = π↓{r}({~r}) = . thus in possibilistic state π, we have no knowledge whether the stranger is a re- publican or not. similarly, notice that the marginals of π for {q} and {p} are also vacuous. the marginal of π for {r, p} is as follows: π↓{r,p}({r,p}) = . , π↓{r,p}({r,~p}) = π↓{r,p}({~r,p}) = π↓{r,p}({~r,~p}) = . thus pacifist republicans are not possible (to degree . ). similarly, notice that the marginal of π for {q, p} is as follows: π↓{q,p}({q,~p}) = . , π↓{q,p}({q,p}) = π↓{q,p}({~q,p}) = π↓{q,p}({~q,~p}) = . thus non-pacifist quakers are not possible (to degree . ). finally note that the marginal of π for {r, q} is as follows: π↓{r,q}({r,q}) = . , π↓{r,q}({r,~q}) = π↓{r,q}({~r,q}) = π↓{r,q}({~r,~q}) = . thus republican quakers are not possible (to degree . ). possibility theory table . a possibility function π for {r,q,p}. w{r,p,q} π r p q . r p ~q . r ~p q . r ~p ~q . ~r p q . ~r p ~q . ~r ~p q . ~r ~p ~q . next, we state a rule for combining possibility function. this rule is called particularization by zadeh [ ]. zadeh [ ] has given two rules for combining possibility functions—minimization and multiplication. axiom b (wx is possible) requires that we include normalization in the definition of combination. since minimization followed by normalization is not associative, multiplication followed by normalization is associative, and associativity of combination is a requirement in vbs (axiom a ), we use multiplcation followed by normalization as our definition of combination. definition . suppose π and π are possibility functions for p and p , respectively. suppose k = max{π ({w↓p }) π ({w↓p }) | w∈wp ∪p }. the combination of π and π , denoted by π ⊗π , is the function for p ∪p given by k– π ({w↓p }) π ({w↓p }) if k ≠ (π ⊗π )({w}) =  ( )  if k = for all w∈wp ∪p . if k = , then π ⊗π is not a possibility function. in this case, π ⊗π is not a nonzero valuation. this means that the knowledge in π and π are inconsistent. if k ≠ , then k is a normalization constant that ensures that π ⊗π is a possibility function. thus, combination consists of pointwise multiplication followed by normalization (if normalization is possible). example . (the rule of combination) consider two pieces of knowledge as follows: . most republicans (at least percent) are not pacifists. . most quakers (at least percent) are pacifists. using possibility theory in expert systems if π is a possibility function representation of the first piece of knowledge, and π is a possibility function representation of the second piece of knowledge, then π ⊗π will represent the aggrega- tion of these two pieces of knowledge. in particular, suppose π is a possibility function for {r,p} as follows: π ({(r,p)}) = . , π ({(r,~p)}) = π ({(~r,p)}) = π ({(~r,~p)}) = (i.e., pacifist republicans are not possible to degree . ), and suppose π is a possibility function for {q,p} as follows: π ({(q,~p)}) = . , π ({(q,p)}) = π ({(~q,p)}) = π ({(~q,~p)}) = (i.e., non-pacifist quakers are not possible to degree . ). then the possibility function π ⊗π = π, say, shown in table , represents the aggregate knowledge. table . the combination of π and π . w{r,p,q} π π π ⊗π = π r p q . . . r p ~q . . . r ~p q . . . r ~p ~q . . . ~r p q . . . ~r p ~q . . . ~r ~p q . . . ~r ~p ~q . . . proposition . the rule of combination described in ( ) has the following prop- erties: c . (commutativity) π ⊗π = π ⊗π . c . (associativity) (π ⊗π )⊗π = π ⊗(π ⊗π ). c . if π is vacuous, then π ⊗π = π . c . in general, π ⊗π ≠ π . the possibility function π ⊗π represents the aggregation of knowledge contained in possibil- ity functions π and π if π and π are independent. like bayesian probability theory, dempster- shafer’s theory of belief function, and spohn’s theory of epistemic beliefs, possibility theory requires that the possibility functions π and π be independent. this is because of property c —double-counting of knowledge may lead to erroneous results. we have already shown that axioms a and a are valid for possibility functions. theorem below states that axiom a is also satisfied. this result is stated in [dubois and prade ]. possibility theory theorem . suppose π and π are possibility functions for p and p , respectively. suppose x∉p , and suppose x∈p . then (π ⊗π ) ↓((p ∪p )–{x}) = π ⊗(π ↓(p −{x})). since all three axioms required for local computation of marginals are satisfied, the fusion al- gorithm described in section . can be used for reasoning from knowledge expressed as possibil- ity functions. in the context of possibility theory, [chatalic et al. ] describes a local computa- tional method for reasoning with possibility functions. the next section describes a small example to illustrate the use of the fusion algorithm to find marginals. . an example in this section, we describe an example in complete detail to illustrate the use of possibility theory in managing imprecision in expert systems. is dick a pacifist? consider the following items of evidence. most republicans (at least percent) are not pacifists. most quakers (at least percent) are pacifists. dick is a republican (and we are more than percent certain of this). dick is a quaker (and we are more than per- cent certain of this). is dick a pacifist or not? we will model the four items of evidence as possibility functions π for {r,p}, π for {q, p}, π for {r}, and π for {q}, respectively, as displayed in table . figure shows the valuation network for this example. if we apply the fusion algorithm using deletion sequence rq, we get (π ⊗π ⊗π ⊗π ) ↓{p} = (π ⊗π ) ↓{p}⊗(π ⊗π ) ↓{p}. the details of the computations are shown in table . the conclusion is that the proposition dick is a pacifist is possible (to degree ), and the proposition dick is not a pacifist is not possible (to degree . ). using possibility theory in expert systems table . the possibility functions π , π , π , and π in the is dick a pacifist? example. w{r,p} π w{q,p} π w{r} π w{q} π r p . q p r q r ~p q ~p . ~r . ~q . ~r p ~q p ~r ~p ~q ~p figure . the valuation network for the is dick a pacifist? example. r p q π π π π table . the computation of (π ⊗π ) ↓{p}⊗(π ⊗π ) ↓{p}. w{r,p} π ⊗π w{q,p} π ⊗π r p . q p . r ~p . q ~p . ~r p . ~q p . ~r ~p . ~q ~p . wp (π ⊗π ) ↓{p} (π ⊗π ) ↓{p} (π ⊗π ) ↓{p} ⊗(π ⊗π ) ↓{p} p . . . ~p . . . proofs . conclusions we have given an axiomatic definition of a qualitative description of a possibility function called a consistent possibilistic state. we have also given a characterization of a consistent possibilistic state in terms of its content. our objective here is to help develop semantics for possibility theory. our contribution in this paper is only one step in this direction. we have also described how possibility theory fits in the framework of valuation-based sys- tems. this was first described by dubois and prade [ ]. we have expanded on this topic by providing a brief yet complete description of the vbs framework, and by describing the main features of possibility theory in terms of the vbs framework. we have also described a small ex- ample in complete detail to illustrate the use of possibility theory in managing uncertainty and im- precision in expert systems. . proofs in this section, we give proofs for all results in the paper. first we state and prove a lemma needed to prove theorem . lemma . suppose {{σ , ..., σm}, ⊗, ↓} is a vbs where σi is a valuation for si, and suppose ⊗ and ↓ satisfy axioms a –a . let x denote s ∪...∪sm. suppose x∈x. then (σ ⊗...⊗σm}) ↓(x–{x}) = ⊗fusx{σ , ..., σm}. proof of lemma . suppose σi is a valuation for si, i = , ..., m. let s = ∪{si | x∈si}, and let r = ∪{si | x∉si}. let ρ = ⊗{σi | x∉si}, and σ = ⊗{σi | x∈si}. note that x∈s, and x∉r. then (σ ⊗...⊗σm) ↓(x–{x}) = (ρ⊗σ)↓((r∪s)–{x}) = ρ⊗(σ↓(s–{x})) (using axiom a ) = (⊗{σi | x∉si})⊗(σ ↓(s–{x})) = ⊗fusx{σ , ..., σm}. ■ proof of theorem . by axiom a , (σ ⊗...⊗σm) ↓{x} is obtained by sequentially marginaliz- ing all variables but x from the joint valuation. a proof of this theorem is obtained by repeatedly applying the result of lemma . at each step, we delete a variable and fuse the set of all valuations with respect to this variable. using lemma , after fusion with respect to x , the combination of all valuations in the resulting vbs is equal to (σ ⊗...⊗σm) ↓(x–{x }). again, using lemma , after fusion with respect to x , the combination of all valuations in the resulting vbs is equal to (σ ⊗...⊗σm) ↓(x–{x ,x }). and so on. when all variables but x have been deleted, we have the result. ■ using possibility theory in expert systems proof of proposition . to show b , suppose s is possible, and suppose r⊇s. assume r is not possible. then from b , s is not possible contradicting our earlier supposition. therefore r must be possible. to show b , suppose s is not possible. if ~s is also not possible, then from b , s∪~s = wx is not possible contradicting b . therefore ~s must be possible. to show that b is true, consider a possibilistic state in which the only proposition that is not possible is ∅ (and all other propositions are possible). clearly, this possibilistic state satisfies ax- ioms b to b . ■ proof of theorem . (sufficiency). assume that the possibilistic state is consistent. consider the proposition ∪i. it follows from b that ∪i is not possible, and follows from b that ∪i is a proper subset of wx. define c = ∪i. we need to show that i = {s∈ wx | s ⊆ c}. suppose r∈i. then r ⊆ ∪i = c, i.e., r∈{s∈ wx | s ⊆ c}. now suppose r∈{s∈ wx | s ⊆ c}, i.e., r ⊆ c. then by axiom b , r∈i. thus we have shown that i = {s∈ wx | s ⊆ c}. (necessity). suppose i = {s∈ wx | s ⊆ c} for some proper subset c of wx. we will show that this possibilistic state satisfies axioms b to b . first note that b is satisfied by definition. b is satisfied since c is a proper subset of wx. to show b , suppose that s is not possible and suppose r ⊆ s. since s is not possible, by hypothesis, s ⊆ c. therefore r ⊆ c. therefore r is not possible. to show b , suppose that r and s are not possible, i.e., r and s belong to i. by hypoth- esis, r ⊆ c, and s ⊆ c. hence, r∪s ⊆ c. therefore r∪s ∈ i, i.e., r∪s is not possible. ■ proof of proposition . p follows from condition p in definition , and p follows from condition p in definition . ■ proof of proposition . all four properties follow trivially from the definition of combination. ■ proof of theorem . note that p ∪p = (p –p )∪(p ∩p )∪(p –p –{x})∪{x}, p = (p –p )∪(p ∩p ), and p = (p ∩p )∪(p –p –{x})∪{x}. suppose w∈wp −p , u∈wp ∩p , and v∈wp −p −{x}, x∈wx. then (w,u)∈wp , and (u,v,x)∈wp . first, note that the normalization factor in the combination on the left-hand side, say k , is the same as the normalization factor in the combination on the right-hand side, say k , i.e., k = k , as shown below. k = max{π ({(w,u)}) π ({(u,v,x)}) | (w,u,v,x)∈wp ∪p } = max{π ({(w,u)}) max{π ({(u,v,x)}) | x∈wx} | (w,u,v)∈w((p ∪p )−{x})} = max{π ({(w,u)}) (π ↓(p −{x}))({u,v}) | (w,u,v)∈w((p ∪p )−{x})} = k = k, say. next, proofs (π ⊗π ) ↓((p ∪p )–{x})({(w,u,v)}) = = max{(π ⊗π )({(w,u,v,x)}) | x∈wx} = max{k– π ({(w,u)}) π ({(u,v,x)}) | x∈wx} = k– π ({(w,u)}) max{π ({(u,v,x)}) | x∈wx} = k– π ({(w,u)}) (π ↓(p −{x}))({(u,v)}) = (π ⊗(π ↓(p −{x})))({(w,u,v)}). ■ acknowledgements this work was supported in part by the national science foundation under grant iri- . i am grateful to didier dubois, serafin moral, and leen-kiat soh for comments. references arnborg, s., d. g. corneil, and a. proskurowski ( ), “complexity of finding embeddings in a k-tree,” siam journal of algebraic and discrete methods, , – . dempster, a. p. ( ), “upper and lower probabilities induced by a multivalued mapping,” annals of mathematical statistics, , - . dubois, d., j. lang, and h. prade ( ), “automated reasoning using possibilistic logic: semantics, belief revision and variable certainty weights,” proceedings of the fifth workshop on uncertainty in artificial intelligence, – , windsor, ontario. dubois, d. and h. prade ( ), possibility theory: an approach to computerized processing of uncertainty, plenum press, new york, ny. dubois, d. and h. prade ( ), “inference in possibilistic hypergraphs,” in bouchon-meunier, b., r. r. yager and l. a. zadeh (eds.), uncertainty in knowledge bases, lecture notes in computer science no. , – , springer-verlag, berlin. chatalic, p., d. dubois, and h. prade ( ), “a system for handling relational dependencies in approximate reasoning,” proceedings of the third international expert systems conference, – , london, uk. kjærulff, u. ( ), “triangulation of graphs—algorithms giving small total state space,” technical report r - , institute for electronic systems, university of aalborg, denmark. kong, a. ( ), “multivariate belief functions and graphical models,” ph.d. dissertation, department of statistics, harvard university, cambridge, ma. mellouli, k. ( ), “on the propagation of beliefs in networks using the dempster-shafer theory of evidence,” ph.d. thesis, school of business, university of kansas, lawrence, ks. ruspini, e. h. ( ), “on the semantics of fuzzy logic,” international journal of approximate reasoning, ( ), – . saffiotti, a. and e. umkehrer ( ), “pulcinella: a general tool for propagating uncertainty in valuation networks,” in d’ambrosio, b., p. smets, and p. p. bonissone (eds.), uncertainty in artificial intelligence: proceedings of the seventh conference, – , morgan kaufmann, san mateo, ca. shafer, g. ( ), a mathematical theory of evidence, princeton university press, princeton, nj. using possibility theory in expert systems shenoy, p. p. ( ), “a valuation-based language for expert systems,” international journal for approximate reasoning, ( ), – , . shenoy, p. p. ( a), “valuation-based systems for propositional logic,” in ras, z. w., m. zemankova, and m. l. emrich (eds.), methodologies for intelligent systems, , – , north-holland, amsterdam. shenoy, p. p. ( b), “consistency in valuation-based systems,” working paper no. , school of business, lawrence, ks. shenoy, p. p. ( c), “valuation-based systems for bayesian decision analysis,” working paper no. , school of business, university of kansas, lawrence, ks. to appear in in operations research, ( ). shenoy, p. p. ( d), “a new method for representing and solving bayesian decision prob- lems,” working paper no. , school of business, university of kansas, lawrence, ks. to appear in in hand, d. j. (ed.), artificial intelligence and statistics iii, chapman & hall, london, england. shenoy, p. p. ( a), “on spohn’s rule for revision of beliefs,” international journal of approximate reasoning, ( ), – . shenoy, p. p. ( b), “valuation-based systems for discrete optimization,” in bonissone, p. p., m. henrion, l. n. kanal, and j. lemmer (eds.), uncertainty in artificial intelligence, , – , north-holland, amsterdam. shenoy, p. p. ( c), “valuation-based systems: a framework for managing uncertainty in ex- pert systems,” in zadeh, l. a. and j. kacprzyk (eds.), fuzzy logic for the management of uncertainty, john wiley & sons, new york, ny. shenoy, p. p. ( d), “using dempster-shafer’s belief-function theory in expert systems,” working paper no. , school of business, university of kansas, lawrence, ks. to appear in proceedings of the spie conference on appplications of artificial intelligence x, orlando, fl, april . shenoy, p. p. and g. shafer ( ), “constraint propagation,” working paper no. , school of business, university of kansas, lawrence, ks. shenoy, p. p. and g. shafer ( ), “axioms for probability and belief-function propagation,” in shachter, r. d., t. s. levitt, j. f. lemmer and l. n. kanal (eds.), uncertainty in artificial intelligence, , – , north-holland, amsterdam. reprinted in shafer, g. and j. pearl, eds. ( ), readings in uncertain reasoning, – , morgan kaufmann, san mateo, ca. spohn, w. ( ), “ordinal conditional functions: a dynamic theory of epistemic states,” in harper, w. l. and b. skyrms (eds.), causation in decision, belief change, and statistics, , – , d. reidel publishing company, dordrecht. spohn, w. ( ), “a general non-probabilistic theory of inductive reasoning,” in shachter, r. d., t. s. levitt, j. f. lemmer and l. n. kanal (eds.), uncertainty in artificial intelligence, , – , north-holland, amsterdam. zadeh, l. a. ( ), “fuzzy sets,” information and control, , – . zadeh, l. a. ( ), “fuzzy sets as a basis for a theory of possibility,” fuzzy sets and systems, , – . zadeh, l. a. ( ), “a theory of approximate reasoning,” in ayes, j. e., d. mitchie, and l. i. mikulich (eds.), machine intelligence, , – , ellis horwood, chichester, u.k. zhang, l. ( ), “studies on finding hypertree covers of hypergraphs,” working paper no. , school of business, university of kansas, lawrence, ks. about the author prakash p. shenoy is the director of the doctoral program and professor in the school of business at the university of kansas at lawrence where he has been since . before that, he was with the mathematics research center at the university of wisconsin at madison for one year. he received a b.tech. in mechanical engineering from the indian institute of technology at bombay, india in , and a m.s. and a ph.d. in operations research from cornell university in and , respectively. his doctoral dissertation titled “on coalition formation and game theory” was written with professor william f. lucas as his dissertation advisor. in the summer of , he attended a six-week faculty development institute on information systems at the university of minnesota organized by the american association of collegiate schools of business. during – , he was an intra-university visiting professor in the department of computer science at the university of kansas. subsequently, he served as a faculty intern in the mis division of hallmark cards, inc., in kansas city, missouri, for six months from june to december, . his research interests are in the areas of artificial intelligence and operations research. he has published many articles on management of uncertainty in expert systems, decision analysis, opti- mization, and the mathematical theory of games. his articles have appeared in journals such as the international journal of approximate reasoning, ieee expert, operations research, management science, and the international journal of game theory. he has received several research grants from the database and expert systems program of the national science foundation, the research opportunities in auditing program of the peat marwick main foundation, the higher education academic development donations program of apple computer, inc., and the general research fund program of the university of kansas. he served as a guest-editor for a special issue of the international journal of approximate reasoning on the subject of dempster-shafer’s theory of belief functions in artificial intelligence. he serves on the editorial board of uncertainty in knowledge bases: an international journal, and as an ad-hoc referee for over journals and conferences in artificial intelligence and operations research. he also serves on the program committees of several conferences in artificial intelligence. his teaching interests are in the areas of optimization, game theory, decision analysis, informa- tion systems, and uncertain reasoning. he has taught graduate-level courses on linear program- ming, non-linear programming, game theory, management information systems, decision support systems, and management of uncertainty in expert systems. he has served on doctoral dissertation committees of eighteen ph.d. students in management science, marketing, accounting, economics, computer science, and geography, two as chairman. in spring , he was awarded the mentor award by the doctoral students in the school of business at the university of kansas. selected working papers unpublished working papers are available from the respective authors at: school of business university of kansas summerfield hall lawrence, ks - usa no. . “propagating belief functions with local computations,” prakash p. shenoy and glenn shafer, february . appeared in ieee expert, ( ), - . no. . “propagating belief functions in qualitative markov trees,” glenn shafer, prakash p. shenoy, and khaled mellouli, june . appeared in international journal of approximate reasoning, ( ), - . no. . “on the propagation of beliefs in networks using the dempster-shafer theory of evidence,” khaled mellouli, april . available as ph.d. dissertation from university microfilms, inc., ann arbor, mi. no. . “auditor’s assistant: a knowledge engineering tool for audit decisions,” glenn shafer, prakash p. shenoy, and rajendra srivastava, april . appeared in auditing symposium ix: proceedings of touche ross/university of kansas symposium on auditing problems, – , school of business, university of kansas, lawrence, ks. no. . “studies on finding hypertree covers of hypergraphs,” lianwen zhang, april . no. . “an axiomatic framework for bayesian and belief-function propagation,” prakash p. shenoy and glenn shafer, july . appeared in proceedings of the fourth workshop on uncertainty in artificial intelligence, minneapolis, mn, - . no. . “probability propagation,” glenn shafer and prakash p. shenoy, august . appeared in annals of mathematics and artificial intelligence, ( ), - . no. . “local computation in hypertrees,” glenn shafer and prakash p. shenoy, august . no. . “a valuation-based language for expert systems,” prakash p. shenoy, august . appeared in international journal of approximate reasoning, ( ), - . no. . “an evidential reasoning system,” debra k. zarley, november . no. . “constraint propagation,” prakash p. shenoy and glenn shafer, november . no. . “axioms for probability and belief-function propagation,” prakash p. shenoy and glenn shafer, november . appeared in uncertainty in artificial intelligence, ( ), - . reprinted in readings in uncertain reasoning, glenn shafer and judea pearl, eds. ( ), – , morgan kaufmann, san mateo, ca. no. . “the unity of probability,” glenn shafer, february . appeared in von furstenberg, g. m., ed., acting under uncertainty: multidisciplinary conceptions, - , kluwer, . no. . “macevidence: a visual evidential language for knowledge-based systems,” yen-teh hsia and prakash p. shenoy, march . a condensed version of this paper appeared as: hsia, y-t. and p. p. shenoy, “an evidential language for expert systems,” in ras, z. w., ed., methodologies for intelligent systems, ( ), - . no. . “audit risk: a belief-function approach,” rajendra srivastava, prakash p. shenoy and glenn shafer, april . no. . “on spohn’s rule for revision of beliefs,” prakash p. shenoy, july . appeared in international journal of approximate reasoning, ( ), – . no. . “consistency in valuation-based systems,” prakash p. shenoy, february , revised may . no. . “perspectives on the theory and applications of belief functions,” glenn shafer, april . appeared in international journal of approximate reasoning, ( ), - . no. . “valuation-based systems for bayesian decision analysis,” prakash p. shenoy, april , revised may . to appear in operations research, ( ). no. . “valuation-based systems for discrete optimization,” prakash p. shenoy, june . appeared in bonissone, p. p., m. henrion, l. n. kanal, and j. f. lemmer, eds., uncertainty in artificial intelligence, ( ), – , north-holland, amsterdam. no. . “a new method for representing and solving bayesian decision problems,” prakash p. shenoy, september , revised february . to appear in in hand, d. j., ed., artificial intelligence and statistics iii, chapman & hall, london, england. no. . “why should statisticians be interested in artificial intelligence?” glenn shafer, november, . appeared in proceedings of the fifth annual conference on making statistics more effective in schools of business, , – , lawrence, ks. no. . “valuation-based systems: a framework for managing uncertainty in expert systems,” prakash p. shenoy, march, . revised july . to appear in fuzzy logic for the management of uncertainty, zadeh, l. a. and j. kacprzyk, eds. ( ), john wiley and sons, new york, ny. no. . “valuation networks, decision trees, and influence diagrams: a comparison,” prakash p. shenoy, june . revised october . no. . “the early development of mathematical probability,” glenn shafer, august . to appear in grattan-guiness, i., ed., encyclopedia of the history and philosophy of the mathematical sciences, routledge, london. no. . “what is probability?,” glenn shafer, august . to appear in hoaglin, d. c. and d. s. moore, eds., perspectives on contemporary statistics, mathematical association of america. no. . “can the various meanings of probability be reconciled?,” glenn shafer, august . to appear in keren, g. and c. lewis, eds., methodological and quantitative issues in the analysis of psychological data, lawrence erlbaum, hillsdale, nj. no. . “response to the discussion of belief functions,” glenn shafer, august . to appear in international journal of approximate reasoning in . no. . “an axiomatic study of computation in hypertrees,” glenn shafer, august . no. . “using possibility theory in expert systems,” prakash p. shenoy, september . to appear in in fuzzy sets and systems. no. . “using dempster-shafer’s belief-function theory in expert systems,” prakash p. shenoy, september . to appear in in the proceedings of the spie conference on applications of artificial intelligence x, orlando, fl. no. . “potential influence diagrams,” pierre ndilikilikesha, september . no. . “independence in valuation-based systems,” prakash p. shenoy, september . revised november . doi: . /j.eswa. . . an integrated approach for developing e-commerce applications francisco garcı́a-sánchez, rafael valencia-garcı́a, rodrigo martı́nez-béjar* departamento de ingenierı́a de la información y las comunicaciones, facultad de informática, universidad de murcia, espinardo (murcia), spain abstract this paper presents a framework that merges various advanced information technologies for developing electronic commerce (e-commerce) applications. the use of e-commerce utilities provides several advantages to businesses. intelligent agents can be used to facilitate some tasks from those that take place in a commercial transaction moving to a second generation of e-commerce applications. we present a prototype which integrates a multiagent system (composed by buyer and seller agents) with a web application. q elsevier ltd. all rights reserved. keywords: electronic commerce; intelligent agents; natural language processing; ontologies . introduction several definitions have been given to the term ‘e-commerce’ or ‘electronic commerce’. the one given by the electronic commerce association is: ‘electronic commerce covers any form of business or administrative transaction or information exchange that is executed using any information and communication technology’. we limit our approach to covering commercial activities conducted on the internet. e-commerce offers opportunities to dramatically improve the way that businesses interact with both their customers and their suppliers, that is, to make business negotiations faster, cheaper, more personalized, and/or more agile. the number of web users who shop for or buy products online is continuously increasing (silverman, bachann, & al-akharas, ). however, searching and buying pro- ducts via on-line can be frustrating due to the lack of help or decision support given to the user. nowadays e-commerce applications are being improved from a first generation stage where buyers are humans who browse through a catalog of commodities (e.g. books, computer components, films) and make purchases, often by means of a credit card transaction, to a second generation with a greater degree of - /$ - see front matter q elsevier ltd. all rights reserved. doi: . /j.eswa. . . * corresponding author. tel.: c ; fax: c . e-mail addresses: fgs @alu.um.es (f. garcı́a-sánchez), valencia@ um.es (r. valencia-garcı́a), rodrigo@dif.um.es (r. martı́nez-béjar). http://www.theeca.org/. automation on both the buyer’s and the seller’s side (wooldridge, ). the aim of this work was to develop a technology for facilitating ‘business-to-consumer’ (b c) and ‘business- to-business’ (b b) processes. for it, various methodologies including (multi)agents (which allow for the interaction among several manufacturers), decision support (used for rating and differentiation processes between the products found) and natural language processing (nlp) have been used. the previous objective can be decomposed into the following sub-objectives: ( ) generation of a (computer) product ontology, ( ) design of a database containing all relevant information about the products in question, ( ) design and implementation of two agent models, one playing a buyer’s role and another playing a seller’s role, ( ) integrating the agents into the foundations for intelligent physical agents-open source (fipa-os) framework so that communication among the agents registered at the platform may take place, and ( ) design and implementation of a web application that makes use of agents to offer e-commerce services (this will be the prototype). the solution described in this work is built on top of an integrated system elaborated on topics belonging to a number of disciplines, including knowledge management, nlp, decision support theory and multiagent technology. the experiments done with the system up to now make us to believe that its future is promising in helping solve some problems we can find in actual e-commerce applications. expert systems with applications ( ) – www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa http://www.theeca.org/ f. garcı́a-sánchez et al. / expert systems with applications ( ) – the structure of this paper is as follows. section offers a brief description of the methodology applied. in section , the e-commerce framework used in this work is introduced. section describes the system prototype developed making use of the framework presented. the validation of the system prototype is shown in section . finally, some conclusions are put forward in section . . methodology in this work, we have applied a methodology that integrates different advanced information technologies with the aim of building a framework for developing e-commerce applications. these are: intelligent agents, which have been used to represent the different entities we find in a transaction process (buyers/sellers); decision support sys- tems (dsss) theory, applied to help buyers in some decision processes; ontologies that establish a reusable, shareable semantics for the terminology, so that different types of agents are able to communicate among themselves; and nlp, which permits to better determine the users desires (since the users can use their own terminology). in the following lines, a brief description of each of them is realised. . . multiagent systems a multiagent system is seen as a system that consists of a group of agents that can potentially interact with each other (vlassis, ). the term agent is defined as follows (wooldridge, ): “an agent is a computer system that is situated in some environment, and that is capable of autonomous action in this environment in order to meet its design objectives”. for this author, an agent is intelligent if it fulfils the following properties: reactivity (it responds in a timely fashion when needed), pro-activeness (it satisfies internal goals; it takes actions when it seems it will be useful) and social ability (it interacts with other agents in order to satisfy goals). in order to design and construct agents it will be necessary to take the following points into account: (i) agents theory (formal specification that describes what properties the agents must fulfil), (ii) agent languages (tools for the design and construction of agent-based systems, e.g.: agent-oriented programming—agent —, concurrent metatem) and, (iii) agent architectures (agents’ internal structure which can be logic-based, reactive, belief-desire-intention-based or layered). agents can be useful as stand-alone entities that are delegated particular tasks on behalf of a user. however, in the majority of cases agents exist in environments that contain other agents. in these multiagent systems (mass), the global behaviour derives from the interaction among the constituent agents. there are two main classes of mass (wooldridge, ): distributed problem solving systems, in which the component agents are explicitly designed to cooperatively achieve a given goal; and open systems, whose agents are not co-designed to share a common goal, and have been possibly developed by different people to achieve possibly different objectives. moreover, the com- position of the system can dynamically vary as agents enter and leave the system. when a group of individual agents form a mas, the presence of a mechanism to coordinate such a group as well as a communication language becomes necessary. among the coordination mechanisms the cases of agents having common objectives (and, therefore, they cooperate) can be distinguished from those where agents are self-interested and have conflictive objectives with other agents. for the latter case we will need negotiation. the most usual cooperation mechanisms are: organizational structures, multiagent planning—centralized and distributed—, con- tract nets and functionally exact cooperation. among the ones belonging to negotiation we may highlight coalitions formation, market mechanism, the bargaining theory, vote, auctions and task assignment between two agents. agent communication languages allow agents to interact each other while they hide their internal work details by exchanging information and knowledge. examples of those ones are kif (knowledge interchange format), kqml (knowledge query and manipulation language) and fipa- acl (foundations for intelligent physical agents-agents communication language). . . . multiagent platforms for this project, we have employed available implemen- tations of agent platforms which conform to fipa (the foundations for intelligent physical agents) specifications, namely, a group of normative rules that permit an agent society to operate among themselves. this model identifies some necessary agent’s roles for the platform management: the ams (agent management system), the df (directory facilitator), the acc (agent communication channel), the ipmt (internal platform message transport) and the iiop (internet interface object protocol) (fipa-os developers guide, ). the communication language among agents that fipa proposes is fipa-acl, which specifies a standard message passing language, establishing codification, semantics and message pragmatics. finally, fipa standard includes the non-agent software integration specification, the agents mobility and security, the ontology service, and the human-agent interaction. amongst the frameworks that satisfy the fipa standard, the more popular are zeus, jade and fipa-os. we have selected this last one in this work. fipa-os is an agent framework developed with the aim of building hetero- geneous agent platforms, agents and services that adjust to fipa standard. in addition to the advantage to be an open system in terms of license of use and extensibility, like f. garcı́a-sánchez et al. / expert systems with applications ( ) – zeus and jade, fipa-os presents some additional advantages: , interoperability with other fipa agents platforms not developed with fipa-os. , it allows for agents heterogeneity, that is, agents implemented with different programming languages. , it supports fipa specification for agents management. , it provides abstractions and application program inter- faces (apis) that permit extending and integrating an agents’s platform with other already existent software platforms. . . . agents for electronic commerce according to wooldridge ( ) agents make it possible the second-generation e-commerce systems, in which many aspects of a consumer’s buying behaviour is automated. one phased model that attempts to describe consumer- buying behaviour is the following (guttman, moukas, & maes, ): . need identification. this stage characterizes the con- sumer becoming aware of some need that is not satisfied. . product brokering. a would-be consumer obtains information related to available products in order to determine what product to buy. . merchant brokering. the consumer selects the supplier. this stage will typically involve examining offers from a range of different merchants. . negotiation. the terms of the transaction are agreed between the would-be consumer and the would-be mer- chant. in some markets, the negotiation stage is empty— the terms of agreement are fixed and not negotiable. . purchase and delivery. the transaction is actually carried out, and the good delivered. . product service and evaluation. the post-purchase stage involves product service, customer service, etc. agents have been widely promoted as being able to (partially) automate some of these stages. . . decision support systems (dss) “in the rush to open their website, e-commerce sites too often fail to support buyer decision making and search, resulting in a loss of sale and the customer’s repeat business” (silverman et al., ). dsss cover a wide variety of systems, tools and technologies that support decision-making in semi-structured and unstructured situ- ations where no one knows exactly how the decision should be made. it provides information, models, and data manipulation tools to solve the structured parts of the problem and help isolating places where both judgment and experience are required. as silverman remarks, in online shopping sites buyers have decisions to work out and tasks to perform that can be facilitated by means of dss software. we help clients with part of this process using a natural language system (based on a morphological analysis) that allows the system to find out the user’s desires. besides, a sorting process helps buyers to determine the most convenient product for them. . . ontologies defining a common vocabulary (apart from the com- munication language selected) is necessary for the agents to communicate with one another. thus, the communication is facilitated through one ontology that defines the concepts and the relations between the concepts of a particular domain. with all, an ontology is viewed in this work as a specification of a domain knowledge conceptualisation (van heijst, schreiber, & wielinga, ), and is represented through a set of concepts, while concepts are defined through sets of both attributes and interconceptual relations. so, the main ontological entity in the system developed is the concept but the use of the other ontological entities such as attributes is also possible in the model in order to provide the system with powerful representational capabilities. . . natural language processing (nlp) in the approach presented in this work, the methodology presented in valencia-garcı́a, ruiz-sánchez, vivancos-vi- cente, fernández-breis, and martı́nez-béjar ( ) has been used to get knowledge from text. this methodology, which uses ontologies and one incremental knowledge acquisition technique termed mcrdr (kang, ), is based on the idea that relationships between concepts are usually associated to verbs in natural language. this methodology uses the mentioned technique and the grammar category of the words in the current sentence to infer other knowledge entities (e.g. concept, attributes and values) in order to create an ontology from a text fragment. . a framework for developing e-commerce applications . . overview of the system the system architecture, which is shown in fig. , has four principal interconnected components. the most important component is the multiagent platform, where two agent types have been implemented, namely a buyer and a seller. these agents communicate with one another by means of acl (agent communication language) messages. these agents must register in the agent platform (ap), which specifies several types of agents that can facilitate the multiagent communication. in particular, the two agent platform types that will be necessary to get registered in are: the directory facil- itator (df), which provides a yellow pages service to fig. . framework general structure. f. garcı́a-sánchez et al. / expert systems with applications ( ) – the other agents, and the agent management system (ams), which provides platform-typical management functions (such as life cycle monitoring, checking of all the entities’ correct behaviour and ‘white pages’). the buyer agent functions include to get external petitions, process them and return the obtained results. the petition processing may include nlp, product ontology matching, planning, plan execution and answer composition. regard- ing the seller agent, this has been provided only with communication services with the buyer agents, so that agent is able to answer for two petitions types: questions about a particular type of product and orders. it can also communicate with a product database, which contains the product catalog of a manufacturer. the second system component is the database (one for each seller agent), which can be shared by the agents. the database can be acceded by the seller agents using the data access object (dao) pattern. the database is composed by the products a manufacturer wants to sell, so that the database contents are ontology-dependent. another important component of the system is the product ontology. the ontology is saved in an xml file and represents the specific relevant domain knowledge we are dealing with, establishing a common set of words that allows different types of agents to communicate one another. the last main component is the web application, which employs the fipa-os platform together with the above- referred agents. for the application design, we have made use of the struts framework and java technology. in the following paragraphs, all these components are detailed more precisely. . . fipa-os and intelligent agents according to fipa-os developers guide ( ), fipa- os is a component-orientated toolkit for constructing fipa compliant agents using mandatory components (i.e. components required by all fipa-os agents to execute), components with switchable implementations, and optional components (i.e. components that a fipa-os agent can optionally use) (see fig. ). agent shell (fipaosagent). it provides a shell for agent implementation to use by simply extending the fipaosa- gent class. the fipaosagent shell is responsible for loading an agent’s profile, and initialising the other components which the agent is composed by. task manager. it provides the ability to split the functionality of an agent into smaller, disjoint units of work known as tasks. these can be defined as self-contained pieces of code that carry out some task and (optionally) return a result, have the ability to send and receive messages, and have little or preferably no dependence on the agent which they are executed within. conversation manager. it provides the ability to keep track of the conversation state at the performative level, as well as a mechanism for grouping messages of the same conversation together. message transport service. it provides the ability to send/receive messages to/from an agent implementation. . . . the buyer agent the buyer agent’s target is to look for the products that a customer requests for (starting from a sentence) as well as asking for a concrete product. fig. . components within fipa-os. f. garcı́a-sánchez et al. / expert systems with applications ( ) – agent initialisation. it involves four main tasks. first, it is needed to initialise the nlp tool. second, the agent must be registered in the fipa agent platform (ams and df agents). third, messages buffers must be activated before conversa- tions take place. finally, the agent has to get all products categories from the ontology (main concepts, synonymous, attributes, relationships,.); the agent will use this infor- mation when that one communicates with other agents. products search. this process begins when a customer (i.e. the application user) writes a sentence denoting the product that (s)he wants to get, and finishes when the system returns the products matching what the user wants. in order to do this, it is necessary a planner that determines the steps that the agent must follow. for it, we have used an strips- type planner and some operators (see section . . for details). when the system has the plan, it is necessary to execute it, so that the possible steps then are the following: , sentence analysis. through this step, the system obtains the concepts underlying the sentence introduced by the customer. for it, the sentences analyser referred to before in this paper is used. , category obtaining. the system determines the product categories that the customer is looking for. for this, the system compares the concepts found in the sentence with the categories extracted from the product ontology. if the system does not find any category that matches any of the extracted concepts, then the plan execution is stopped. , manufacturer finding. this is executed thanks to a task included in the agent’s implementation. the task target is to look for seller agents in the ‘yellow pages’. for it, the system incorporates the directory facilitator agent—df—, which returns a collection with all fig. . message format for info the identifiers (id) of the agents found. in order to match this id with the agent transport address, the system uses the agent management system (ams), i.e. white pages. if the df does not find any agent, then the plan execution stops. , information request. for each category and seller agent found, a task is created that starts a conversation with the found agent and returns all products found grouped by seller agent. the information returned has the following format (fig. ).where ‘cat-i’ represents the ith category, ‘man-j’ the jth manufacturer and ‘prod-ij’ is a product of the ith category offered by the jth manufacturer.for each category, the system will return all the products (and its respective manufacturers) found with respect to that category.it is important to emphasize that each seller agent works in parallel to find the products of each category. the buyer agent saves this information then and asks for the next category. , offer selection. once a buyer agent has the information described before, the only remaining thing it has to do is to sort the products of each category.the result of this process is a collection with the following format (fig. ).where ‘bestpr’ represents the best product for the category under question, ‘manbp’ represents the manufacturer for such a product, ‘ ndpr’ represent the second best product for the category, ‘man p’ is the manufacturer for the mentioned product, and so on.this will be returned to the customer. asking for a product. in order to execute a product petition, a search must have taken place previously, because it is necessary to indicate the seller agent id, in addition to other product-related features such as the amount of products, the product type and the product code. rmation request. table an operator for the blocks world problem pick-up(block) preconditions clear(block) on-table(block) arm-empty add list holding(block) delete list on-table(block) clear(block) arm-empty fig. . message format for offer selection. f. garcı́a-sánchez et al. / expert systems with applications ( ) – the task to attend the request hence needs the product code (this must agree with the one that the seller agent has registered), the product type or category, the agentid of the seller agent that offers the product and the number of units the customer requests. all this information is then sent to the seller through an acl message. . . . the seller agent the seller agent is in charge of performing three tasks: one consists of attending petitions and determining which task to execute; the second is to keep the communication with the buyer agent which requests some products, and the last carries out the petition record of a given product. therefore, this agent’s main objective is to receive petitions and questions from different buyer agents. agent initialisation. this process is similar to the agent initialisation process in the buyer agent. this agent does not use nlp neither planning. search task. this task is executed when a ‘query’-type message is received. the aim of this task is to return all the products found in the database with the category indicated in the query. petition task. this task is called when a message ‘petition’ is received. the function of this task is to insert a record of the product sale into the database. this task can check whether the product is available in stock or it is necessary to make a petition to the providers. . . . the planner strips (fikes & nilsson, ) is a linear planner that attempts to find a sequence of operators in a space of world models to transform a given initial world model into a model in which a given goal formula can be proven true. it represents a world model as an arbitrary collection of first-order predicate calculus formulae and works with models consisting of a large number of formulae. it uses a resolution theorem prover to answer questions of particular models and mean-ends analysis as a guide towards the desired goal-satisfying model. for every world model, we have a set of applicable operators, which transform the world model into another world model. the problem solver finds some composition of operators that transforms a given initial world model into one that satisfies a stated goal condition. in strips, a world model is represented by a set of well formed first-order predicate calculus formulae. each operator in a solution corresponds to an action routine whose execution causes a robot to take certain actions. each operator in strips is composed by: – a set of preconditions. to execute the action related to the operator it is necessary that preconditions are true before the operator can be applied. – delete list, which is a set of formulae that will not be true after the operator has been applied (so that the planner has to delete them from the current world model). – add list, which is a set of formulae that will be true after the operator has been applied (so that the planner has to add them to the current world model). for example, an operator for the blocks world problem is shown in table . in table , clear(block), on-table (block), arm-empty and holding(block) are well formed formulae. strips, like most of planners, has been applied to the blocks world problem as an effective benchmark (slaney & thiébaux, ). the blocks world problem consists of a finite number of blocks stacked into towers on a table large enough to hold them all. the towers’ positioning on the table is irrelevant. the blocks world planning problem is to turn a blocks initial state into a goal state, by moving one block at a time from the top of a tower onto another tower or to the table. the optimal block world planning problem consist in doing so in a minimal number of moves. most of planners adopt the strips representation and search forward the state space. the grt planner (refanidis & vlahavas, ) is a domain-independent heuristic planner, which solves planning problems calculating in a first, phase the distances between the facts and the goals of the problem. the second phase consists of a simple best first search strategy using the distances calculated in the previous phase. this planner has been validated in several domains like blocks-world domain, in which grt can easily solve problems with more than blocks. another planner, which adopts the strips represen- tation is the fast-forward planning system (hoffmann, ), which attacks the planning problems by forward table the set of well formed formulae formula description preference(s) the customer’s desires contained in the sentence s concept(co, s) the concept co in s category(c, co, s) the category c is one of the customer’s preferences related to co in s manufacturer(ma) the manufacturer ma is registered in the system, therefore, it is available for possible petitions info(of, ma, c) the manufacturer ma presents the offer of for c sort(of) the offer’s group is sorted out attending to the relation quality-price f. garcı́a-sánchez et al. / expert systems with applications ( ) – searching in the state space, guided by a heuristic function. this function is extracted from the domain description relaxing the planning problem by ignoring parts of its specification, concretely the delete lists of operators. in this work, we have used a strips planner for allowing agents to determine dynamically the actions to be performed. more precisely, this planner has been implemented into the buyer agent to accomplish the product search task. the well formed formulae of first-order predicate calculus used in this planner are showed in table . the initial world model is formed by the formulae which represent the user’s preferences. the formulae of the final world model represent an ordered set containing the products that the customer may be interested in (according to his/her likes and dislikes). the operators which have been designed can be viewed in table . . . . natural language processing the nlp application used in this approach is based on the work presented in valencia-garcı́a et al. ( ), where table the strips operators operator findmanufac- turer( ) preconditions delete list add list manufactu analyze(f) preconditions preference manufactu delete list manufactu add list concept(co getcategory(co) preconditions concept(co delete list concept(co add list category(c requestinfo(c) preconditions manufactu category(c delete list category(c add list info(of, ma selectoffer(of) preconditions info(of, ma delete list info(of, ma add list sort(of) the authors described a software tool for ontology building from natural language texts. the principal idea sustaining our approach is that in natural language relationships between knowledge entities are usually associated to verbs. the knowledge acquisition process is divided into three sequential phases, which have been implemented into three separate modules: pos-tagging, concept search and inference. pos-tagging. the main objective of this process is to obtain the grammatical category of each word in the current sentence. for this purpose, the pos-tagger described in ruiz-sánchez, valencia-garcı́a, fernández-breis, martı́- nez-béjar, and compton ( ) is used. this is the first sub- process to be carried out. concept search. through this process, linguistic expressions representing concepts are identified. the associations between linguistic expressions and concepts have been stored in the conceptual knowledge base in a previous training of the system. this process is quite simple and as a result we obtain all the expressions of the fragment which are already in the conceptual knowledge base. inference. in natural language, relationships between concepts are usually associated to verbs. although the sub- process in which mcrdr acts is mainly concerned with obtaining relationships between concepts, it can also be used to get other knowledge categories like concepts, attributes, or values. this mcrdr component is formed by a knowledge-base that contains linguistic expressions representing generic conceptual relationships, and by an mcrdr subsystem that infers the participants in these relationships. we describe next the modus operandi of this process. firstly, the verb in the current sentence is identified. then, the user (of the system if the knowledge base is not empty) searches for the type of semantic relationship associated to that verb. once the type of relation associated to the main verb in the current sentence has been found, the mcrdr sub-system is applied to extract knowledge by action return all the manufacturer registered rer(ma) (s) obtain the relevant concepts within a sentence rer(ma) rer(ma) , s) ,s) try to match a concept with the set of product categories ,s) , co, s) rer(ma) make a request to a manufacturer about a category , co, s) , co, s) , c) , c) sort a set of offers related to the same category , c) f. garcı́a-sánchez et al. / expert systems with applications ( ) – means of the grammatical category of the words, their position in the current sentence, and the type of relation associated to the verb, if any. the inputs of this tool are the product queries, in natural language, introduced by the customers. an example of these queries could be: “i need a gb hard disk”. then, the system would obtain, for instance, an ontology formed by the concept ‘hard disk’, so that the attribute ‘capacity’ of this concept is ‘ gb’. as it has been indicated before, the nlp system needs a previous training in the domain. this training is based on the study of a set of texts in natural language related to the domain. in our case, this domain has been a hardware domain. . . product ontology in order to maintain a conversation, agents must have a common language. this is established by means of an ontology, which contains the main concepts owning to the domain we are dealing with. in addition to this information, the ontology also includes attributes, values, relations between concepts and axioms so that consistency checking and inferences are done. these concepts are the ones the buyer agent utilizes to indicate the customer’s preferences to the seller agent and the ones the seller agent utilizes to access the database. an (incomplete) example of a product ontology related to the computer products domain is shown in fig. . in this example, ontology shows how a computer is composed by several elements: monitor, keyboard, mouse, motherboard, processor, etc. . . . ontology editor to specify the product ontology, a graphic tool prototype has been implemented that allows to introduce concepts, attributes and semantic relations in a intuitive way. it also generates an xml document with the designed ontology’s description. fig. . computer pro the xml file has the following characteristics: , ducts concept: !concept commentz“” namez“computer”o !alternative-nameso !nameoworkstation!/nameo !nameopc!/nameo !nameolaptop!nameo !/alternative-nameso !specific-attributeso !attribute commentz“” namez“trade- mark” typez“string”o!/attributeo !attribute commentz“” namez“model” typez“string”o!/attributeo !attribute commentz“” namez“descrip- tion” typez“string”o!/attributeo !attribute commentz“” namez“guar- anty” typez“integer”o!/attributeo !attribute commentz“” namez“price” typez“integer”o!/attributeo !/specific-attributeso !/concepto ont it can be noticed that every concept, besides the ‘main name’, has also alternative names (that we will use like synonyms), in addition to its attributes. , relation: !relation namez“component_object”o !concept_nameoprocessor! /concept_nameo !concept_nameocomputer! /concept_nameo !relation_typeo !propertyononsymmetry!/propertyo !/relation_typeo !/relationo ology. f. garcı́a-sánchez et al. / expert systems with applications ( ) – in this case, in addition to the concepts taking part in the semantic relation under question, the relation will have a name with the relation type and eventually some other properties associated to that relation. . . . the analyser there are two possible parsers. the api sax accom- plishes a sequential access to the document and provides an event based programming model (callbacks). on the other hand, the api dom makes a tree structure from the document. the main advantages of api sax are its simplicity and that it consumes less time of processing. so, we have chosen this option because we do not need to update the document and the api sax is a more simple solution to do the queries in the ontology. thus, every agent only needs to do one document analysis. the buyer agent, once the concepts of the customer sentence have been found, should compare these concepts with that found in the ontology in such a way that, if they match, we understand that it is a product the customer wants to buy. . . product database the product each manufacturer possesses has been stored in a database. for it, mysql database server has been chosen because, apart from being open source (uses the gpl—gnu general public license), it is very fast, reliable, and easy to use mysql reference manual ( ). besides, the mysql database software is a client/server system that consists of a multi-threaded sql server that supports different backends, several different client programs and libraries, administrative tools and a wide range of application programming interfaces (apis). a jdbc–odbc bridge driver has been utilized to access the database from java due to its efficiently using native fig. . connection pool database client code. besides, the drive may be changed easily in the web application description file. in order to add flexibility to the database management the dao pattern has been used, so that persistent information from different sources (relational databases, ldap, xml, etc.) can be stored and recovered. the benefits from utilizing this solution are the following: transparency-oriented, easier component migration, complexity reduction and centraliza- tion of the access to data through a single layer. we also use a connection pool to improve the performance (see fig. ). the general structure of the dao pattern is shown in fig. . . . web application with the purpose of applying the designed agents in a real environment, the fipa-os framework has been integrated into a web application. . . . agents integration we have to distinguish three subjects: the agentloader creation, the buyer agent creation and the seller agent creation. agentloader creation. the environment initialisation is needed in order to start working with fipa-os. this is done by ‘loader’, which reads configuration files and creates the ams and df agents (those that form the platform kernel). the agentloader must be created by the frontcontroller servlet (see section . . ) at the application initialisation time. the buyer agent creation. each customer has a buyer agent associated to him/her. when a customer starts to use the application, a buyer agent is created. the problem observed then was that the agent creation process is very slowly. implementation. fig. . dao pattern implementation. f. garcı́a-sánchez et al. / expert systems with applications ( ) – the solution provided for that problem was to use the connection pool’s idea commented before for the database, so that when there are not enough agents to associate to customers new agents are created. the seller agent creation. for each manufacturer, we will have previously set up a database containing the stock together with a seller agent associated to it and registered in the platform. in this way, the customer through his buyer agent will have access to the stock of a large number of manufacturers being able to choose a better product. . . . application design the design is based on the apache struts web application framework. in this way, we only have one servlet (frontcontroller) where all petitions go to. this servlet studies every petition and executes the action that corresponds to the petition. . the system prototype a software tool based on the approach described in this work has been designed and implemented for facilitating e-commerce. this is a web application implemented with java technology and designed following struts recommen- dations. the application domain on which it has been applied is computer hardware. the user has been provided with two ways to introduce a query. on the one hand, it is possible to write free text with the user’s desires and, on the other hand, the user can select specific product categories from those indicated in fig. . we have trained the system with some sentences a user can write to ask for a hardware component. in this process, when we introduce a sentence to the tool, we have to establish the concepts, attributes, values and relations this sentence contains. for example, if we say i want a gb hard disk, we will have to indicate that there is a concept, ‘hard disk’, with an attribute, capacity, which value in this example is ‘ gb’. once the tool have learnt this, if we introduce a new sentence like i want a mb memory, the tool will automatically determine that there is a concept ‘memory’, with the attribute capacity which value is ‘ mb’. once the user has entered information into the system about the wanted products, the system creates a set with the product categories that match the user wishes. once the communication with the seller agents is established, fig. . screenshot showing product categories. f. garcı́a-sánchez et al. / expert systems with applications ( ) – the application shows a page with the best products of each category. then, the user is shown a list of all the products related to a found category sorted by an internal relation quality-prize. finally, for each product the user is interested in (according to the information supplied by her/him to the system), the user will be able to ask for the amount (s)he wants (see fig. ). internally, this request will be addressed fig. . list of products so to the seller agent that represents the manufacturer which has the product under question in its catalog. . validation of the system the system described in previous sections has been employed by eight users, who were given two hardware rted by categories. table system scoring question score (by user) u u u u u u u u previous experience in suchlike application interface performance – application utility – easy use agree with solution utility (do they think new characteristics are important (dss, nlp,.?)) – f. garcı́a-sánchez et al. / expert systems with applications ( ) – product catalogs. each catalog had products belonging to seven categories: hard disk, mouse, graphic card, sound card, monitor, motherboard and cd player. moreover, each catalog possessed three items. the mode the system was deployed is the following. in one computer, we installed the web application and one of the two seller agents (the one which accesses to its catalog). on the other hand, in a distinct computer we set up then the seller agent to make queries on the another available catalog. the users were asked to use the system during half an hour and to make at least requests to the system. utilities like search by all product categories to browse between products in an easy way were assessed as important for them. the main problem found by them was the performance (caused by the time consumed by the nlp module). they all think that the interface is intuitive but too simple and agree in the fact that the functionalities added to the e-commerce application are positive. finally, the goodness of solution was well assed for most of them: they think that it is necessary to take additional attributes into account (price, brand, product characteristics, etc.) in order to minimize the space of search. these results are summarized in table . . discussion and conclusion the results presented in this paper contribute to the development of intelligent electronic-commerce appli- cations that are capable of simulating real world transactions. concerning related work, a similar solution to ours can be found in silverman et al. ( ). however, while in our solution we split the functionality into two different agent roles (buyer and seller), the quoted author makes use of an unique search agent that integrates all the functionality. using intelligent agent technologies, nlp, decision support systems, ontologies and web technologies we have made a system that facilitates b c (business-to-consumer) and b b (business-to-business) transactions. the agents have been designed and implemented within a multiagent framework, fipa-os, that follow a broad standard such as fipa (foundation for intelligent physical agents). it allows to develop a distributed application in a very simple way. all this makes the development process more realistic, simulating the interaction between a store customer and the seller, which attempts to help the buyer to find the product that (s)he is looking for. at the same time, the solution presented in this paper also gives the user an object pre-classification that supports the user decision. the system developed has made use of advanced technologies to build an intelligent application but there are some aspects that we have not taken into account such as security, auctions, efficiency, etc. in particular, a few interesting issues should be explored and improved in future system improvements. first, a more complete ontology (i.e. with more knowledge) must be created. the application must be able to take into account and to use all the knowledge contained in the ontology (axioms, attri- butes, etc.). second, it is important to better train the nlp tool allowing more free text, so that the application be able to understand the customer’s wishes and to do a better search. third, in the current approach the agents must follow a rigid plan created by an strips-like planner. we can use a more flexible planner, which uses more knowledge about the environments and creates plans that take into account the possible contingencies that could take place. fourth, we can make the application totally product- independent. this implies the intensive use of the ontology in all application fields. fifth, both the buyer user and the seller user might be given the possibility of configuring their respective agents. so, a buyer agent could be configured with the buyer user’s preferences (e.g. the cheapest product, or the one with more quality,.) and something similar with the seller agent. finally, it could be interesting to use xml messages for the agents to communicate with one another, as a number of manufacturers currently do. acknowledgements we thank the spanish ministry for science and technology for its support for the development of f. garcı́a-sánchez et al. / expert systems with applications ( ) – the system through projects tic - , fit- - - and fit- - - ; the regional govern- ment of murcia (spain) through project i siu ; and seneca foundation through project pi- / /fs/ . we also thank the european commission for its support under project alfa ii fa. references fikes, r. e., & nilsson, n. j. ( ). strips: a new approach to the application of theorem proving to problem solving. artificial intelli- gence, , – . fipa-os developers guide ( ). http://fipa-os.sourceforge.net/, february. guttman, r. h., moukas, a. g., & maes, p. ( ). agent-mediated electronic commerce: a survey. the knowledge engineering review, ( ), – . hoffmann, j. ( ). the fast-forward planning system. ai magazine, – . kang, b. ( ). multiple classification ripple down rules. phd thesis, university of new south wales. mysql reference manual ( ). http://www.mysql.com/. refanidis, i., & vlahavas, i. ( ). the grt planner. ai magazine, – . ruiz-sánchez, j. m., valencia-garcı́a, r., fernández-breis, j. t., martı́nez béjar, r., & compton, p. ( ). an approach for incremental knowledge acquisition from text. expert systems with applications, , – . silverman, b. g., bachann, m., & al-akharas, k. ( ). implications of buyer decisión theory for design of e-commerce websites. international journal of human computer studies, , – . slaney, j., & thiébaux, s. ( ). blocks world revisited. artificial intelligence, , – . valencia-garcı́a, r., ruiz-sánchez, j. m., vivancos-vicente, p. j., fernández-breis, j. t., & martı́nez-béjar, r. ( ). an incremental approach for discovering medical knowledge from texts. expert systems with applications, ( ), – . van heijst, g., schreiber, a. t., & wielinga, b. j. ( ). using explicit ontologies in kbs development. international journal of human computer studies, , – . vlassis, n. ( ). a concise introduction to multiagent systems and distributed ai. http://carol.science.uva.nl/~vlassis/cimasdai/. wooldridge, m. ( ). an introduction to multiagent systems. new york: wiley. http://fipa-os.sourceforge.net/ http://www.mysql.com/ http://carol.science.uva.nl/~vlassis/cimasdai/ an integrated approach for developing e-commerce applications introduction methodology multiagent systems decision support systems (dss) ontologies natural language processing (nlp) a framework for developing e-commerce applications overview of the system fipa-os and intelligent agents product ontology product database web application the system prototype validation of the system discussion and conclusion acknowledgements references perception granular computing in visual haze-free task expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a perception granular computing in visual haze-free task - /$ - see front matter crown copyright � published by elsevier ltd. all rights reserved. http://dx.doi.org/ . /j.eswa. . . ⇑ corresponding author. tel.: + . e-mail addresses: huhong@ict.ac.cn (h. hu), pangl@ics.ict.ac.cn (l. pang), tiandp@ics.ict.ac.cn (d. tian), shizz@ics.ict.ac.cn (z. shi). hong hu, liang pang ⇑, dongping tian, zhongzhi shi key laboratory of intelligent information processing, institute of computing technology, chinese academy of science, beijing , china a r t i c l e i n f o a b s t r a c t keywords: granular computing leveled granular system fuzzy logic machine learning haze free brain-like computer in the past decade, granular computing (grc) has been an active topic of research in machine learning and computer vision. however, the granularity division is itself an open and complex problem. deep learning, at the same time, has been proposed by geoffrey hinton, which simulates the hierarchical structure of human brain, processes data from lower level to higher level and gradually composes more and more semantic concepts. the information similarity, proximity and functionality constitute the key points in the original insight of granular computing proposed by zadeh. many grc researches are based on the equivalence relation or the more general tolerance relation, either of which can be described by some dis- tance functions. the information similarity and proximity depended on the samples distribution can be easily described by the fuzzy logic. from this point of view, grc can be considered as a set of fuzzy logical formulas, which is geometrically defined as a layered framework in a multi-scale granular system. the necessity of such kind multi-scale layered granular system can be supported by the columnar organiza- tion of the neocortex. so the granular system proposed in this paper can be viewed as a new explanation of deep learning that simulates the hierarchical structure of human brain. in view of this, a novel learning approach, which combines fuzzy logical designing with machine learning, is proposed in this paper to construct a grc system to explore a novel direction for deep learning. unlike those previous works on the theoretical framework of grc, our granular system is abstracted from brain science and information science, so it can be used to guide the research of image processing and pattern recognition. finally, we take the task of haze-free as an example to demonstrate that our multi-scale grc has high ability to increase the texture information entropy and improve the effect of haze-removing. crown copyright � published by elsevier ltd. all rights reserved. . introduction lin ( ) pointed out that granulation seems to be a natural methodology deeply rooted in human thinking. many daily things are routinely granulated into sub-things (lin, ). in the ieee- grc conference of information about the granular computing (grc), the outline of grc is defined as a general computation theory for effectively using granules such as classes, clusters, subsets, groups and intervals to build an efficient computational model for complex applications with huge amounts of data, information and knowledge (zadeh, ). just as the scholars summarized in the ieee-grc conference though the label is relatively recent, the basic notions and principles of grc, though under different names, have appeared in many related fields, such as information hiding in programming, granularity in artificial intelligence, divide and conquer in theoretical computer science, interval computing, cluster analysis, fuzzy and rough set theories, neutrosophic com- puting, quotient space theory, belief functions, machine learning, databases and many others (bargiela & pedrycz, ). the above definition of grc is too augmental and the subjects about classes, clusters, subsets, groups and intervals have already studied by arti- ficial intelligence and mathematics for a long time. what is really new point for grc? we think that the new or main point of the grc lies in the original insight of grc proposed by zadeh, in which there are three basic concepts that underlie human cognition: granulation, organization and causation. informally, granulation involves decomposition of whole into parts; organization involves integration of parts into whole; and causation involves association of causes with effects. granulation of an object a leads to a collection of granules of a, with a granule being a clump of points (objects) drawn together by indistinguishability, similarity, proximity or functionality (zadeh, ). in this original insight of grc, zadel pointed out three important aspects about grc: ( ) the grc is a main character of human cognition, ( ) so called grc is based on indistinguishability, similarity, proximity or functional- ity, ( ) there is a close relationship among granulation, organiza- tion and causation. based on these points, we think that it is necessary to find some key points of grc in human cognition. there are two kinds of grc research: perception-level and knowledge-level. a perception-level grc does a series feature transformation and tries to find meta-knowledge implied in sam- ples; a knowledge grc tries to process knowledge or structure http://crossmark.crossref.org/dialog/?doi= . /j.eswa. . . &domain=pdf http://dx.doi.org/ . /j.eswa. . . mailto:huhong@ict.ac.cn mailto:pangl@ics.ict.ac.cn mailto:tiandp@ics.ict.ac.cn mailto:shizz@ics.ict.ac.cn http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa h. hu et al. / expert systems with applications ( ) – information based on metal-knowledge. in this paper we focus on the perception-level. indistinguishability, similarity and proximity can be described by equivalence relation or tolerance relation, and these relations can be described by some kind distance functions. from the rele- vant literature, it is easy to see that many grc researches focus on classification and clustering (yao, , ; zhang & zhang, , a). zhang and zhang ( , a, b, ) use the quotient space theory to try to study indistinguishability and similarity. yao ( ) extends the equivalent class to rough approximation set. the quotient space structure described by equivalence relation is used to probe the structure of granules such as classes, clusters, subsets, groups etc. in a more general way, lin ( , ) and yao ( , ) use binary relations and neigh- borhood systems to study indistinguishability and similarity respectively, the geometric concepts: partitions, covering and topology, and neighborhood can be described by binary relations in the algebra. pedrycz, hirota, pedrycz, and dong ( ) define granular on fuzzy sets and discuss several operations and their granular consistency (pedrycz et al., ). lin ( ) gives out a summary about the history of granular computing, he discusses all formal description of granular computing and some further directions e.g. grc, databases and data mining, grc and clouding computing etc. in fact, grc should be discussed in the framework of human cognition from perception to pattern recognition and knowledge processing. although the concept of granular comput- ing has been proposed more than ten years, only few people pay attention to this subject. in fact granular computing pays much more attention to the leveled computing of intelligence. just as yao ( ) pointed out: ‘‘granules in the family are called focal elements of discussion at the level. each level is represented by a plane. while granules at the same level are of similar nature, gran- ules at different levels may be very different. consequently, we may use different vocabularies and languages for descriptions at different levels.’’ the leveled computation revealed by granular computing is very important for machine learning, e.g. the famous approach deep learning. the term deep learning gained much attraction in the mid- s after a publication by bargiela and pedrycz ( ), castro ( ). nowadays it becomes a huge wave of technology trend for big data and artificial intelligence. deep learning simulates the hierarchical structure of human brain, pro- cesses data from lower level to higher level, and gradually com- poses more and more semantic concepts. these facts mean that deep learning has a close relationship with the granular computing. all the granular computing researches aforementioned, in gen- eral, neglect information transformation and feature abstraction, which are very important for deep learning. in this paper, we pro- pose a novel framework of granular system, which has ability to process information transformation and object-background sepa- ration. we take the haze-free task as an example to validate the ability of our granular system. these facts mean that the deep learning has a close relationship with the granular computing proposed by us. the main contributions of this paper include: ( ) the basic notions and principles of grc termed with differ- ent names, but they have appeared in many related fields, such as information hiding in programming, granularity in artificial intelligence, divide and conquer in theoretical com- puter science, interval computing, cluster analysis, fuzzy and rough set theories, neurotrophic computing, quotient space theory, belief functions, machine learning, databases, and many others, so old version of granular computing is just an abstraction of old methods, in this paper we give a novel concrete model for granular computing which has a multi scale layered structure from feature abstraction to classification. ( ) our novel granular system has a close relation with deep learning, so it develops a new focus for deep learning. it is the first time that fuzzy logic is introduced for leveled fea- ture abstraction in deep learning. ( ) although fussy logic is often mentioned in granular comput- ing, for example,fuzzy logic and rough set technique are used by lin ( ) for word computing, and liu, xiong, and wu ( ) use fuzzy lattices in the classification based on hyperspherical granular computing (liu et al., ). in this paper, fuzzy logic is not only used for describing granu- lar similar to hyperspherical granular in liu et al. ( ), but also for feature abstraction and classification. for this pur- pose, we propose a novel and effective approach which is combined fuzzy logical designing, psvm and back propagation. ( ) the granular computing proposed by us gives a novel approach for the task of haze-free, the experiments’ result show that this approach is sound. the rest of this paper is organized as follows: in section ,we try to give out a formal definition of granular system and granular computing based on the a tolerance relation, which is described by fuzzy logical formula; in section , we discuss the algorithm to design a granular system; in section , we give out a concrete example of designing a granular system for haze-free task; at last section is the discussion and looking forward to the future. . granular system based on tolerance relation the difference between a granular system based on equivalence relation and a granular system based on tolerance relation is that an equivalence relation will divide a space into nonoverlapping covering while a tolerance relation will create overlapping cover- ing of this space. yiyu yao proposed granular computing paradigm for concept learning in which two learning strategies are investigated. a global attribute-oriented strategy searches for a good partition of a uni- verse of objects and a local attribute–value-oriented strategy searches for a good covering (yao & deng, ). in this paper, granular computing is started from feature vectors e.g. images, not attributes. in order to simulation perception procession of our cognition, we define a set of multi-scale nested convex regions with a corresponding computing based on this set. there are two main purposes to build such a granular system based on tolerance relation: ( ) granular systems are designed to describe similarity and proximity of information, which can be described by toler- ance relation. granular systems based on tolerance relation can be viewed as a topological structure built by topological bases on the topological space ðx; sÞ induced from a metric space ðx; disÞ by the metric dis. granular systems based on tolerance relation can be used to describe domain stricture, which represents indistinguishability, similarity and prox- imity of examples.classification is determined by the indis- tinguishability, similarity and proximity of information. there are two kind similarity among examples-static simi- larity and dynamical similarity. (a) if elements of classes are distributed in standard convex regions, we can use some kind distance function to describe classes distribution domains. in this case, simi- larity between two objects can be intuitively described by distance functions. if disðx; yÞ is a distance function h. hu et al. / expert systems with applications ( ) – in the n-dimensional space rn and c is a point in rn , the formula disðc; yÞ < r described a convex region d in rn which takes c as its center. every point y in this region is equal to y ¼ c þ e, here e is some kind noise, if d is just a ball, e can be viewed as white noise which has an amplitude less than r.we denote such kind similarity as static similarity. (b) the dynamical similarity is different from static similar- ity, if one object o continuously changes to another object o , e.g. a tadpole continuously grow up to a frog, then o and o are dynamical similar, i.e. if all elements in a class a are dynamical similar, the distribution domain of a is a connected domain. dynamical similarity will cause distribution domain become very complicate and have a nonlinear borderline. although,a dynamical similarity may cause the inner class difference larger than the among classes difference, it can also be described by equivalence relation or tolerance relation. the difference of a granular system based equivalence relation and granular system based on tolerance relation is that an equivalence relation will divide a space into non overlapping covering and a tolerance relation will create overlapping covering of this space. the relation described by the formula disðx; yÞ < r is the special case of a tolerance relation. ( ) the main purpose of information transformation in pat- tern recognition is to recognize or classify different objects from their mixture, so information transformation used in pattern recognition should be taken place in a granular system which describes static similarity and dynamical similarity. this is the main point we will dis- cuss in this paper. now we try to use fuzzy logical formula based on distance func- tion to define granular systems. there are three distance axioms: ð Þdisða;aÞ¼ ;ð Þdisða;bÞ¼disðb;aÞ;ð Þdisða;bÞþdisðb;cÞp disða;cÞ. disðx;yÞ r (strong r-cut set) or sprða; cjdis; d; xÞ¼ spða; cjdis; d; xÞ p r (r-cut set). sprðx; cjdis; d; xÞ defines an open convex region and denoted as a granule, and sprðx; cjdis; d; xÞ defines a closed convex region. definition (leveled perception granular system based on tolerance relation gsys). a granular system based on tolerance relations of distance function (granular system for short) is a set of granules, every granule is a -tuple fg; sfg, here g is a convex region which is described by tolerance relations of distance function, and sf is a set of fuzzy logical functions (denoted as ‘‘adjoint functions’’) which are computed from the convex region g, the outputs of all fuzzy logical functions in sf are denoted as ‘‘an adjoint vector’’ of this granule. the granules of gsys have the following attributions. ( ) multi-scale leveled structure: the metric space x (e.g. a finite connected volume region in the n-dimensional real space rn) is the only level granule, the level granule is denoted as gðcoeg Þ¼ fx; sfg, where coeg is a coefficient set, coeg is usually empty, and the function sf is a set of fuzzy logical functions (denoted as adjoint functions). the convex region of a level granule gðcoeg Þ¼ fg ; sf g is defined by the conjunction of finite number r-cut sets spr a; c jdis ; d ; x � � ; spr a; c jdis ; d ; x � � ; . . . ; spr a; c k jdis ; � d k ; xÞ, where coeg is its coefficient set to define the convex region g , so g can also be written as g ðcoeg Þ and coeg ¼ c ; c ; . . . ; c k x; d ; d ; . . . ; d k ; dis n o . the first level of our granular system is denoted as c ðxÞ¼ fgðcoeg Þg. if the l level granules gðcoeglÞ¼ fgl; sflg have been defined, the l þ level granules can be defined as gðcoeglþ Þ, which can be defined as the convex region created by the intersection of gl and finite number strong r-cut sets: spr a; clþ jdis lþ ; dlþ ; x � � ; spr a; clþ jdis lþ ; dlþ ; x � � ; � � � ; spr a; clþ klþ � jdislþ ; dlþ klþ ; xÞ. a level l þ granule’s coefficient is coeglþ ¼ fclþ ; c lþ ; . . . ; c lþ klþ x; dlþ ; d lþ ; . . . d lþ klþ ; dislþ g, where klþ is denoted as the number of simple logical formulas in same level. the centerofgðcoeglþ Þ is clþ i gðcoeg lÞ, and its radius is dlþ i d l i; lim l ! d l i ¼ . the gðcoeg Þ is called as the ‘‘the father granule of gðcoeglþ Þ ’’. ( ) granular computing (grc): a granular computing is described by a set of fuzzy logical formula upon above multi scale leveled structure of convex regions. the purpose of a granular computing (grc) is to transfer fea- ture information and classy points in an input space x, so at least one level of a granular computing (grc) outputs a fuzzy label for points in an input space x. if there are totally m classes, a fuzzy label l is a m dimensional fuzzy vector l ¼fl ; l ; . . . ; lmg; and p i¼ ;...;mli ¼ . castro ( ) proved that fuzzy logic controllers using fuzzy rules are universal approximations, later, li and philip chen ( ) shows a proof of the equality between a forward neu- ral circuit (or circuit) and a fuzzy logical inference. so it is not difficulty to prove that any continuous functions f : rn�!½ ; �n can be simulated by such kind nested layered granular computing with arbitrary small errors. a level l þ adjoint function f lþ receives its input from the outputs of level l þ p; p > adjoint functions, i.e. if a leveled granular system gsys has k levels, grc is taken from level k to , so the level of a grc is upside down with the level of the gsys. the st level grc takes place in the smallest kth level granules’ convex regions of the gsys. two kinds layered computing can be taken place over a granular system. in the first kind layered computing,the adjoint feature vectors of larger scale level n granules are computed based on the h. hu et al. / expert systems with applications ( ) – adjoint feature vectors of smaller scale n þ level gran- ules,such kind layered computing has strictly nested struc- ture, (fig. (a)), and is denoted as ‘‘nested layered computing’’. in the second kind layered computing, adjoint feature vectors of level n granules can be computed based on all adjoint feature vectors of smaller granules, which have level greater than n (fig. (b)), such kind layered computing is denoted as ‘‘unnested layered computing’’. nested layered computing is a special case of unnested layered computing. ( ) radiuses of convex regions: a granular system can have countable infinite or finite levels. the radiuses of gran- ules’convex regions decrease and tend to zero when the level goes to infinite. ( ) centers’ grid: the centers of granules will distribute in a so called center grid, we call the set of all centers of granules of level l þ on the granule gðcoeglÞ as the center grid of level l granule gðcoeglÞ denoted as gclþ ðgðcoeglÞÞ . we denote the set of all centers of level l þ granules over x as gclþ ðxÞ and all centers of level l þ granules over a level k < l granule gðcoegkÞ as gclþ ðgðcoegkÞÞ. the center grid is usually discrete, but it can also be a contin- uous set e.g. the whole metric space x. ( ) shape of granules: the shapes of granules is defined by their distance functions disðÞ . if disðÞ can be an abstract distance function, then a granule’s convex region can be an arbitrary convex region. every level uses same distance function, so the granules in the same level have the same shape, but for different levels, granules’region may have different shapes. ( ) the cover over a granule: in order to create a cover over gðcoeglÞ, the elements in the centers’ grid gclþ ðgðcoeglÞÞ should be tight enough. such kind cover can be formerly defined as: clþ ðgðcoeglÞÞ¼ ^ i klþ sprðx; clþ i jdis lþ ; dlþ i ; xÞ^ gðcoeg lÞjclþ i gc lþ ðgðcoeglÞÞ; sprðx; clþ i jdislþ ; d lþ i ; xÞ^ gðcoeg lÞ – / � � all level l þ granules create a cover of the whole space x, and denoted as the level l þ cover or the level l þ layer of x, clþ ðxÞ¼ [ gðcoeglÞ clðxÞ fclþ ðgðcoeglÞÞg . radial basis neural network (haykin, ), which can be used to simulate continuous functions, is an example of two layers gran- ule system. definition (hyper-granules and mini-granules). all level n þ granules gðcoegnþ Þ, which are contained in a level n granule gðcoegnÞ, denoted as ‘‘mini-granules’’, and the level n granule gðcoegnÞ is denoted as a ‘‘hyper-granule’’. fig. . two kind layered compu after the theory of fuzzy logic was conceived by zadeh ( ), many fuzzy logical systems have been presented, for example, the zadeh system, the probability system, the algebraic system, and bounded operator system, etc. according to universal approx- imation theorem (haykin, ), in this paper, the extended bounded operator is selected, which is denoted as ‘‘q-value weighted bounded operator’’. it is not difficult to prove that q-value weighted fuzzy logical formulas can precisely simulate any continuous functions f : rn�!½ ; �n with arbitrary small error, or vice versa, i.e. every grc can be completed by a set of fuzzy logical functions of q-value weighted bounded operator with arbitrary small error. definition (bounded operator fð�f ;�fÞ). bounded product: x�f y ¼ maxð ; x þ y � Þ, and bounded sum: x�f y ¼ minð ; x þ yÞ, where x; y . in order to simulate grc, it is necessary to extend the bounded operator to weighted bounded operator. the fuzzy formulas de- fined by q-value weighted bounded operators is denoted as q-value weighted fuzzy logical functions. definition (q-value weighted bounded operator fð�f ;�fÞ). q-value weighted bounded product: p �f p ¼ f�f ðp ; p ; w ; w Þ ¼ maxð ; w p þ w p �ðw þ w � ÞqÞ ð Þ q-value weighted bounded sum: p �f p ¼ f�f ðp ; q ; w ; w Þ¼ minðq; w p þ w p Þ ð Þ where p ; p q. for association and distribution rules, we define: ðp df p Þhf p ¼fhf ðfdf ðp ;p ;w ;w Þ;p ; ;w Þ and p dfðp hf p Þ¼ fdf ðp ; fhf ðp ; p ; w ; w Þ; w ; Þ, here df ; hf ¼�f or �f . we can prove that �f and �f follow the associative condition (see appendix c) and x �f x �f x . . .�f xn ¼ min q; x i n wixi ! ð Þ x �f x �f x . . .�f xn ¼ max ; x i n wi xi � x i n wi � ! q ! ð Þ for more above q -value weighted bounded operator fð�f ;�fÞ fol- lows the demorgan law, i.e. ting over granular system. h. hu et al. / expert systems with applications ( ) – nðx �f x �f x . . .�f xnÞ¼ q � min q; x i n wi xi ! ¼ max ; q � x i n wi xi ! ¼ max ; x i n wiðq � xiÞ�ð x i n wi � Þq ! ¼ nðx Þ�f nðx Þ�f nðx Þ � � ��f nðxnÞ: ð Þ but for the q-value weighted bounded operator fð�f ;�fÞ, the distribution condition is usually not hold, and the boundary condi- tion is hold only all weights equal to , for p �f q ¼ f�f ðp ; q; w ;w Þ¼maxð ;w p þð �w ÞqÞ and p �f q ¼ f�f ðp ; q; w ; w Þ¼ minðq; w p þ w qÞ. in this paper, we show that the task of haze-free can be completed by a common grc based on fuzzy logical formulas of bounded fuzzy operator. . hybrid designing of leveled perception granular system based on fuzzy logic and psvm owing to the limitation of the scope, in this paper only nested layered grc is discussed. a nested layered grc is defined by the in- put and output relation of a granular computing on a granular sys- tem. there are three kinds relations between nearby layers (layers k and k þ ) of a nested grc: ( ) binary logic; ( ) fuzzy logic; ( ) alogical relation. because fuzzy logic and binary logic are all created by the sig- moid function, so back propagation method can be used to mod- ify weights of all layers. in order to speed up the learning process, for a layered grc, we combine logical designing with psvm (fung & mangasarian, ), such kind novel approach is called as ‘‘logical support vector machine (lpsvm)’’. for nested layered grc, parameters in the binary logical layers can be di- rectly designated according to the binary relation; for the fuzzy logical layers, parameters can also be set according to these lay- ers’functions, but a suitable small adjustment by back propaga- tion is necessary, this is similar to the deep learning proposed by geoffrey hinton such that a many-layered neural network could be effectively pre-trained one layer at a time, treating each layer in turn as an unsupervised restricted boltzmann machi- ne,then using supervised backpropagation for fine-tuning. for the non logical (alogical) layer, parameters should be learned based on samples according to the input and output relation function fiðx ; x ; x ; . . . ; xnÞ, we can use back propagation method or psvm, to learn weights for fiðx ; x ; x ; . . . ; xnÞ . the designing strategy of lpsvm: � step : except for the alogical layer’s weights, designing the lay- ers’ weights according to the logical (binary or fuzzy) relations, for fuzzy logical relations, a suitable modification of weights maybe be necessary according to the task of this layer; � step : alogical layers’ weights are computed for the input layer to the last output layer. for an alogical layer i, if x is the input train set, computing the inner layers’ output from the st layer to the ði � Þth layer based on x; � step : using psvm to compute the ith layer’s weights w i according to ( ); � step : repeat the step to step , until the output error is small enough. � step : using back propagating approach to modify all layers weights. the weight vector w l of w i ¼ x du ð Þ where the weight vector w i of the node, u is computed by ( ) and x and d are the problem data, i.e. x ¼ ½x ; . . . ; xn�, and diagonal ma- trix d ¼ y .. . . . . .. . . . . yn ;ðxi; yiÞ is a training sample with xi feature vector and target yi . u ¼ i m þ dðxx þ ee Þd � �� e ð Þ where m is a positive parameter selected for guarantee of a small magnitude kw ik; i is the identity matrix, and e is a vector with all elements are . . granular system for visual task the columnar organization of our brain’s primary visual cortex strongly supports the granular system defined aforementioned. many functions of the primary visual cortex are still unknown, but the columnar organization is well understood. the lateral geniculate nucleus (lgn) transfers information from eyes to brain stem and primary visual cortex (v ) (mountcastle, ). colum- nar organization of v plays an important role in the processing of visual information. local similarity of information processing gives rise to colum- nar organization has a granular structure. v is composed of a grid of ð mm Þ neural area of hypercolumns (hc) in our brain’s pri- mary visual cortex. every hypercolumn contains a set of minicol- umns (mc), which have same focus. each hypercolumn analyzes information from one small region (described by a distance func- tion) of the retina. adjacent hypercolumns analyze information from adjacent areas of the retina, so the structure of a columnar organization can be described by a set of fuzzy logical formulas similar to a granular system. hypercolumns (or supercolumns), minicolumns (mc) can be viewed as granules. similar to the pri- mary visual cortex,in our granular system, there are two kind gran- ules:hyper-granule and mini-granules in some levels of our granular system. a hyper-granule contain a bundle of mini- granules. definition (perception granular system of columnar organization (cogsys)). a perception columnar organization is a special per- ception granular system, in which, there is at least one hyper- granule gðcoegnþ Þ such that all mini-granules included in it have same convex region, but different adjoint functions. in this paper, in order to simulating visual cortex, a granular system of columnar organization (cogsys) is designed for the haze-free task.in our hybrid designing approach (lpsvm), we firstly design leveled granular systems with the help of fuzzy lo- gic, and then we use psvm to accomplish the learning for some concrete visual tasks. . . the theory of image matting according to levin, lischinski, and weiss ( ), image matting refers to the problem of softly extracting the foreground object from an input image and a trimap image. ‘‘tripmap’’ means three kinds of regions, white denotes definite foreground region, black denotes definite background region and gray denotes undefined region. formally, image matting methods take i as an input, which is assumed to be a composite of a foreground image f o and a back- ground b in a linear form and can be written as i ¼ af o þð � aÞb . for the haze-free task, the fuzzy label of haze or non-haze is described by the parameter a . and the task of image matting tries to find a function fo ¼ ffoðiÞ . closed form solution assumes that a is a linear function of the input image i in a small window h. hu et al. / expert systems with applications ( ) – w : ai aii þ b; i w . then to solve a spare linear system to get the alpha matte. our grc approach gets rid of the linear assump- tion between a and i. instead, we try to introduce nonlinear rela- tion between a and i: aw ¼ fðw iwÞ ð Þ here w iw is the image block included in the small window w, and aw is its center pixel’s fuzzy label. we take color or texture in local win- dow as our input feature, and the trimap image as the target. after training, the neural fuzzy logical network will generate the result of alpha matte. in the application of alpha matting, our method can re- move the haze using dark channel prior as the trimap. . . leveled perception granular system for haze-free task in this section, we try to design a perception granular system of columnar organization (cogsys) for the haze-free task, here only nested layered grc is needed. the recognition of our leveled fig. . a layers’ structure of a granular sy granular system (see fig. ) is started with the recognition orien- tation or simple structure of local patterns, then the trimap image is computed based on these local patterns. eq. ( ), which has a high ability to simulate fuzzy logic operator (see the detail in the appendix) is used to design grc. the weight wi in eq. ( ) can be viewed as connections among granules. a nested layered grc is de- fined by the input and output relation of a granular computing on a granular system. just as above mentioned, there are three ways to design weights of a layered grc:according to the binary or fuzzy logical relation about this layered grc and according to the input and output relation function fiðx ; x ; x ; . . . ; xnÞ from training samples. ulþ ;i ¼ x k wlþ ;i;k � ilþ ;k;i olþ ;i ¼ sigmðulþ ;i; t lþ ;i; kÞ ð Þ where sigmðÞ is a sigmoid function eq. ( ), and olþ ;i is the output of a level l þ granule. stem for haze-background separation. h. hu et al. / expert systems with applications ( ) – sigmðx; t; kÞ¼ =f þ expf�k � ðx � tÞgg ð Þ the theorem discussed in the appendix a guarantee that above defined granule computing can simulate a boolean function with arbitrary small error. as the designing of gsys contains two parts: ( ) convex regions, ( ) adjoint fuzzy logical functions for grc. the following gsys for haze-free task is a very simple convex regions can be described by distance function. the input space is just an image, which is a -dimensional space x ¼fðx; y; r; g; bÞg, here every example ðx; y; r; g; bÞ represents a pix- el of this image, ðx; yÞ is the pixel’s location and ðr; g; bÞ is pixel’s color value. the nested granular system is build on the image. a granular system is built upon images, with fuzzy logical formula spða; cjdis; d; xÞ here x ¼ð ; ; ; ; Þ, and d ¼ for level grc, and d > for higher level grcs. all levels’ centers are located on the whole image plane, so every centers grid is just the image plane and granules are overlapped. in the following pages, we focus on the designing of adjoint fuz- zy logical functions for grc. if there are k levels in our gsys, the kth level receives the input image i, and the first level granule outputs the result haze free im- age. the relation between input and output of a level-l granule is described by eq. ( ). the weights among granules can be designed by lpsvm, the weights of st and nd layers are designed by fuzzy logic, and the weights of the rd layer are designed by psvm to learn the trimap image. for the sake of simplicity, in the following gsys, we use the or- der of grc level which is upside down with the granular system le- vel, and one layer may contain two grc levels. the grc of cogsys is formally defined as bellow: ( ) the st layer – fuzzy logical layer every hyper-granule (fig. ) in the st layer tries to change a pixels’ image block ib into a binary pixels’texture pattern. the input image is normalized. a hyper-granule hg ¼ðg; sfÞ in the st layer contains mini-granules to focus a small window, every mini-granule focuses only one pixel, so the convex region of a hyper-granule is described by disðx; cÞ� . a hyper-granule completes the task of image processing. there are three kinds fuzzy logical functions in a hyper-granule’s sf : fig. . pattern just a every g ( ) in a local image pattern recognition way (lipw): the st processing directly transforms every pixel’s value to a fuzzy logical one by a sigmoid function. f ðfibg Þ¼ fsigmðpi;jÞg ð Þ every st layer granule tries to change a local image into a binary texture . for a hyper granule is defined by a distance function disðx; cÞ < , which is small window, a hyper-granule in the st layer contains granules, and ranule focuses only one pixel. here pi;j; i; j ¼ ; ; is the rgb pixel value in a small window; ( ) in a local binary pattern operator simulating way (lbpw). the nd processing is also completed by a sig- moid function; the difference is that every boundary pix- el’s value is fuzzy exclusive or � with the center pixel’s value before sending it to a sigmoid function, f ðfibg Þ¼ ffðpi;jÞg ð Þ here fðpi;jÞ¼ sigmðpi;j � p ; Þ when i; j – , and fðp ; Þ¼ . f is similar to a local binary pattern operator (lbp) mentioned in ojala, pietikäinen, and harwood ( ) as a mean of sum- marizing local gray-level structure. the operator takes a local neighborhood around each pixel, thresholds the pixels of the neighborhood at the value of the central pixel and uses the resulting binary-valued image patch as a local image descrip- tor. it was originally defined for neighborhoods, giving bit codes based on the pixels around the central one. such processing emphasizes the contrast of texture, and our exper- iments support this fact. ( ) hybrid lipw and lbpw (lbipw). the adjoint function f ð�Þ in lbipw is same as f ð�Þ in lbpw, except that fðp ; Þ¼ p ; in f ð�Þ, while fðp ; Þ¼ in f ð�Þ . every granule in a st layer’s granule has only one input weight wij in fig. , which equals ; when k !þ , the coef- ficient k in eq. ( ) changes the outputs from fuzzy values to binary numbers. ( ) the nd layer–binary logical layer every nd layer mini- granules try to recognize a definite shape (see fig. ), so they share the same convex region with a st layer hyper-gran- ule, which focuses on the same small window in an image, and can be described by disðx; cÞ < . if there are total q local small patterns, a hyper-granule in the nd layer con- tains q (in our system q ¼ or ) mini-granules of the nd layer, which have same receptive field, but with a differ- ent adjoint fuzzy logical function, which tries to recognize a definite shape from the output of a st layer hyper-granule. for example, the ‘‘\’’ shape in fig. can be described by a adjoint fuzzy logical formula (eq. ( )). the ‘‘and’’ operator for inputs in eq. ( )can be created by a granule mc (see. fig. ). in eq. ( ), every pixel pij has two states mij and mij . suppose the unified gray value (or rgb value) of pij is gij, and an image module needs a high value gij at the place of mij and a low value at mij . so the input for the gran- ule mc at mij is iij ¼ gij, and at mij is iij ¼�ð : � gijÞ. a not gate mc is needed for iij ¼�ð : � gijÞ, here gij; i; j ¼ ; ; is the output of a st layer hyper-granule. p¼m ^m ^m ^m ^m ^m ^m ^m ^m ð Þ wij ¼ ; if the jth bit of a binary pattern ¼ � ; if the jth bit of a binary pattern ¼ � ð Þ where for lipw and lbipw, j ¼ ; ; ; . . . ; ; for lbpw, the cen- ter st-layer granule is useless, so j ¼ ; ; ; . . . . there are three kinds hyper-granules in the nd-layer, which receive three differ- ent outputs of a st-layer’s hyper-granule, so a hyper-granule in the nd-layer may work in one of following three ways: . in the local image pattern recognition way (lipw): every nd layer hyper-granule contains nd-layer’s mini-granules, and inputs of these nd-layer’s mini-granules come from a st-layer’s hyper-granule which works in lipw way. every nd-layer’s hyper-granule tries to classify the image block in this window into binary texture patterns (btp), e.g. eight important btps are shown in fig. . the pixel value is ‘‘ ’’ for fig. . every the nd layer’s granule contains or granule which corresponds to or modules in above picture. fig. . a hyper-granule in the nd layer contains q granules which have same receptive field and try to recognize q definite small shapes. a ‘and’ granule is needed for every nd layer granule. h. hu et al. / expert systems with applications ( ) – white and ‘‘ ’’ for black. in this mode, granules of the st layer output a vector, i.e., a fuzzy logical pattern of a btp, which is computed by a sigmoid function. . in the local binary pattern operator simulating way (lbpw), a nd-layer’s hyper-granule contains nd-layer’s mini- granules which receive input from the output of a st-layer’s hyper-granule, which works in the way of lbpw. . in the hybrid lipw and lbpw (lbipw) way, a nd-layer’s hyper- granule contains nd-layer’s mini-granules which receive input from the output of a st-layer’s hyper-granule, which has -dimensions. in our system, a gsys is built for every color channel r,g or b, so a hyper-granule in the nd layer has a dimensions output or dimensions output. as a binary logical layer, in order to recognize a binary pattern, an ‘and’ granule with index i is needed (see fig. ) for every nd- layer granule, and the weights of this ‘and’ granule to the st-layer granules are set as eq. ( ), the corresponding parameters in eq. ( ) are set as the threshold t i ¼ : , and k ¼ : . . . . the rd layer – alogical layer the convex region of this layer can also be described by disðx; cÞ < . the output of a hyper-granule in the nd layer, which has or dimensions, is transformed to the rd-layer granules to compute the similarity parameter or fuzzy value ai in eq. ( ), the weights of this layer is computed by psvm, the target is provided by so called dark channel prior which is computed by the approach mentioned in he, sun, and tang ( ). as all ai are optimised on the whole image, in this layer,the whole image is the only convex region. as the small windows focused by hy- per-granules in the nd-layer are overlapped, the focuses of rd- layer’s granules are also overlapped. . . . the th layer – fuzzy logical layer in this layer, a granule tries to remove the haze from original image. a granule in the th layer computes a pixel of a haze free image according to fuzzy logical equation eq. ( ) iiðxÞ¼ minfq; aiðxÞ � jiðxÞþð � aiðxÞÞ � aig¼ jiðxÞ�f ai ð Þ where ji is the haze free image, ii is the original image, ai is the global atmospheric light which can be estimated from dark channel prior, ai is the alpha matte generated by rd layer, and �f is the q-value weighted bounded sum with weights w ¼ aiðxÞ; w ¼ � aiðxÞ, here q is max gray or rgb value of a pixel, and aiðxÞ and ð � aiðxÞÞ are weights. although we can use back propagation approach to compute pixels’ value jiðxÞ given the haze image pixel value iiðxÞ based on eq. ( ), for the sake of simplicity, we directly use the eq. ( ) mentioned by he et al. ( ) to com- pute the haze free image. as every aiðxÞ is computed upon the whole image, the pixel of haze-free image is also computed upon whole image, so the whole image is also the convex region of this layer. jiðxÞ¼ iiðxÞ� ai maxðaiðxÞ; a Þ þ ai ð Þ where ji is the haze free image, ii is the original image, ai is the glo- bal atmospheric light which can be estimated from dark channel prior, ai is the alpha matte generated by rd layer, and a is a threshold, a typical value is . . . experiments result the haze-free experiment result ( ) the haze-free and texture information entropy texture information can give out a rough measure about the effect of haze-freeing, we use the entropy of the texture his- togram to measure the effect of deleting haze from images. the entropy of the histogram is described in eq. ( ). haze makes the texture of an image unclear, so theoretically speaking, haze removing will increase the entropy of the texture histogram. entropy : h ¼� xg� i¼ pðiÞlog ½pðiÞ� ð Þ fig. . the processing result of granular system for visual haze-free task. h. hu et al. / expert systems with applications ( ) – the pðiÞ denote the rate of each pattern in histogram. in general pðiÞ define in eq. ( ). here patterns we use are the lbps in eq. ( ), where sigmðin � icÞ¼ , if in > ic þ else sigmðin � icÞ¼ . table the tex area : area area area fig. . granule pðiÞ¼ hðiÞ=ðnmÞ; i ¼ ; ; . . . ; g � ð Þ in fig. (a), (b), (c), (d) and (e) are the results of lbpw, lbipw, lipw, the linear mode (lmkh) by he et al. ( ), and the original image respectively. from fig. , we can see that the texture structure in the waist of a mountain becomes vaguer from lbpw, lbipw, lipw to lmkh. for the sake of the nd kind processing in the st-layer’s granules pays much more attention to the contrast, lbpw has the highest ability to remove the haze, lbpw and lipw are complemen- tary approaches, lbipw, which is the cooperation of them, has a similar ability as the linear approach proposed by he et al. ( ). according to the results showed in the table , which are about tex- ture information entropy of the image, we can see that the texture information entropy is increased after haze-free processing, so our approaches have higher ability to increase the texture information entropy than the linear approach proposed by he et al. ( ). ture information entropy of the image blocks (area : the waist of a mountain; right bottom corner) in the fig. . lbpw lbipw lipw lmkh original . . . . . . . . . . the relation among the precision (rmse) of psvm learning and k parameters in th . theoretically speaking, lbpw is a pure texture processing, so lbpw has a highest value, lipw is much more weaker than lbpw, lbipw is the hybrid of lbpw and lipw, so it has a average ability. the tex- ture information entropy of the area correctly reflects this fact. but for the area , as it already has a clearest texture structure in the ori- ginal image, the deleting of haze may cause overdone. the texture information is over emphasized by lbpw in the aera , so it has a lowest texture information entropy and almost becomes a dark area. this fact means that overtreatment is more easier to appear in a non linear processing than a linear one in the haze-free task. ( ) the effect about the degree of fuzzyness just as the theorem mentioned above, the parameter k in eq. ( ) can control the fuzzyness of a granule, when the parameter k in eq. ( ) tends to infinite, a granule behaves from a fuzzy logical formula to a binary logical formula. this experiment is about the relation among the precision (rmse) of psvm learning and k parameters in the first and second layer. lbpw is a pure texture processing and pays much more attention to the contrast of an image’s nearby pixels, a set of large k is necessary for a low rmse, which corre- sponds to binary logic; but lbipw and lipw aphe pear to prefer fuzzy logic for a set of small k when rmse is small. a possible explanation for this fact is that lbp proposed by ojala et al. ( ) is binary, not fuzzy, and has a sound clas- sification ability for image understanding under binary pat- tern, but lbipw and lipw are not binary, they have fuzzy information at least for the center pixel of a small window (fig. ). e first and second layer, the parameter k in eq. ( ) can control the fuzzyness of a fig. . simulating fuzzy logical and-or by changing thresholds of eq. ( ). the x- axis is the threshold value divided by . , the y-axis is errg. the real line is errand between i �f i and v i , and the dot line is the error between i �f i and v i . fig. . more result of granular system for visual haze-free task. h. hu et al. / expert systems with applications ( ) – ( ) the comparison between our approach and lmkh to illustrate the effect of our approach in haze-free task, we apply it on the other images and compare with lmkh (fig. ). half of the result is better than the lmkh, the rest is as good as the lmkh by manual evaluation. . discussion in this paper, we give out a concrete example to show that the theory of grc can help us to design the brain-like computer. the experimental results show that lpsvm is a promising approach for designing of a granular system similar to a columnar organiza- tion for image haze-removing task. the concept of granular com- puting is proposed by bargiela and pedrycz ( ). just as he said: a granule is a clump of objects (points) drawn together by indistinguishability, similarity, proximity or functionality. the nec- essary of granular computing to study the information transforma- tion in the pattern recognition lies in indistinguishability, similarity, proximity or functionality of sensed information due to the local similarity in the information processing of pattern rec- ognition, multi-scale information processing is a common phe- nomenon in pattern recognition. in actual fact, the grc based on the leveled granular system aforementioned can simulate all mul- ti-scale information processing with arbitrary small error. this fact is very important for the hot approach-deep learning. in this paper, we use a novel designing approach (lpsvm) to design a granular system similar to the structure of columnar organization of visual cortex, we demonstrate that fuzzy logic and machine learning can be hybrid and cooperated easily to design a granular system. this approach not only give out a novel concrete realization of abstract models for granular computing mentioned lin ( ), but also gives a new focus for deep learning. for more,the corre- sponding of grc can simulate multi-scale information processing for the task of haze-free of images, and our experiments show that h. hu et al. / expert systems with applications ( ) – our approach has some approvement for the task of haze-free com- paring to the approach proposed by the linear approach proposed by he et al. ( ). for further directions, although our lpsvm gives out a concrete example for designing a granular system for haze free task, many details of lpsvm should be studied in the task of pattern recogni- tion under the framework of deep learning, especially for the lay- ered feature abstraction in the task of pattern recognition. for more, we will extend the investigation by looking at other nested layered computing for more complex tasks. however, since layered computing has no feedback, which is important for many visual tasks in dynamical situations, we also plan to extend our layered granular computing to a more general one which allows for both feedback and dynamical regulation for the task of computer vision. acknowledgments this work is partially supported by the national program on key basic research project ( program) (no. cb ), the national natural science foundation of china (nos. , , , ), the national high- tech r and d program of china ( program) (no. aa ), the national science and technology support program ( ba b ) and the china information technology security evaluation center (cnitsec-ky- - / ). appendix a. sigmoid function and binary logic theorem . suppose in eq. ( ), every wlik ¼ bk t i; bk > ; t i > ; k k, for more, c ¼fsiji ¼ ; . . . ; lg is a class of index sets, and every index set si is a subset of f ; ; ; . . . ; kg, then we have: ( ) if fðx ; x ; . . . ; xkÞ¼ _ l¼ ;...;l ð^xjiÞ jl sl is a disjunctive normal form (dnf) formula, and the class c ¼fsiji ¼ ; . . . ; lg is the class which has the following two characters: ( ). for every si; sj c; si \ sj – sk c for all k and i – j (this condition assures that fðx ; x ; . . . ; xkÞ has a simplest form); ( ). every si has the character p j sl bj > , where si c, and any index sets s r c have character p j s bj < , or if p j s bj > , there must be an index set si c such that s \ si ¼ si (this condition assures c is the largest), then the output described by eq. ( ) can simulate the dnf formula fðx ; x ; . . . ; xkÞ¼ _ l¼ ;...;l ð^xjiÞ i sl with arbitrary small error, where xi ¼ zi, if the corresponding input ii ¼ zi, or xi ¼ �zi if ii ¼ � zi . ( ) if a neural cell described by eq. ( ) can simulate the boolean formula fðx ; x ; . . . ; xkÞ with arbitrary small error, and ð^xiÞ i sl is an item in the disjunctive normal form of fðx ; x ; . . . ; xkÞ, i.e. fðx ; x ; . . . ; xkÞ¼ at xj ¼ for all j sl and xj ¼ for all j r sl, then p i sl bi > . ( ) if a couple of index sets sl and sl can be found in the formula fðx ;x ;. . .;xkÞ¼ _ l¼ ;...;k; ð ^ t sl xtÞ, such that ð ^ t sl xt Þ^ð ^ t sl xt Þ¼ zi ^ �zi ¼ false, then the output described by eq. ( ) can’t simulate the formula fðx ; x ; . . . ; xkÞ . proof. ( ) if it ¼ , for all t sl, and it ¼ , for all t r sl, becausep i sl bi > , then for the index set sl is a subset of f ; ; ; . . . ; kg, we have v i ¼ =½expð�kðui � t iÞþ � ¼ =½expð�kð x k k wikik � t iÞÞþ � ¼ =½expð�kð x i sl bi � Þt iÞÞþ �; so limk!þ v i ¼ ¼ fðx ; x ; . . . ; xkÞ. if it ¼ ; t s ; it ¼ ; t r s and s r c, then according to the condition of this theorem: if p i s bi < , limk!þ v i ¼ ¼ fðx ; x ; . . . ; xkÞ; if p i s bi > , then there is an in- dex set si c such that s \ si ¼ si , then limk!þ v i ¼ ¼ fðx ; x ; . . . ; xkÞ. so when k ! , the error between output described by eq. ( ) and fðx ; x ; . . . ; xkÞ trends to . ( ) if the output described by eq. ( ) can simulate the boolean formula fðx ; x ; . . . ; xkÞ which is not a constant with arbi- trary small error, and for a definite binary input x ; x ; . . . ; xk , then the arbitrary small error is achieved when k trends to infinite and ðui � t iÞ¼ p k sl wikik � t i – where sl is the set of the labels and ii ¼ , for all i sl , and ii ¼ , for all i r sl. the theorem’s condition supposes that every wik ¼ bkt i; bk > ; t i > ; k k, and x ; x ; . . . ; xk are bin- ary number or , so if fðx ; x ; . . . ; xkÞ is not a constant, when fðx ; x ; . . . ; xkÞ¼ , there must be limk!þ v i ¼ ; and when fðx ; x ; . . . ; xkÞ¼ , it is necessary for limk!þ v i ¼ . limk!þ v i ¼ needs that -kð p i sl bit i � t iÞ trends to minus infinite and limk!þ v i ¼ needs that -kð p i sl bit i � t iÞ trends to plus infinite. so if fðx ; x ; . . . ; xkÞ¼ at xj ¼ for all j sl and xj ¼ for all j r sl, in order to guarantee limk!þ errfðx ;x ;...;xkÞðwi; ; wi; ; . . . ; wi;k; t iÞ ¼ ; p i sl bi > must be hold, here errfðx ;x ;...;xkÞðwi; ; wi; ; . . . ; wi;k; t iÞ is the error between output described by eq. ( ) and fðx ; x ; . . . ; xkÞ. ( ) the third part of the theorem is based on the simple fact that for a single neuron v i is monotone on every input ii which can be zi or � zi. h appendix b. sigmoid function and fuzzy logic for more above granular computing can approximately simu- late bounded operator. bounded operator fð�f ;�fÞ bounded prod- uct p�f q ¼ maxð ; p þ q � Þ, bounded sum p�f q ¼ minð ; p þ qÞ. based on eq. ( ), the membrane potential’s fixed point under in- put ik is ui ¼ p k wikik and the output at the fixed point is v i ¼ =ðexpð�ui þ t iÞþ Þ. if there are only two inputs i ; i ði ; i ½ ; �Þ in eq. ( ), we set w ¼ : and w ¼ : , then ui ¼ i þ i . now we try to prove that the bounded operator fð�f ;�fÞ is the best fuzzy operator to simulate neural cells described by ( ) and the threshold ti can change the neural cell from the bounded oper- ator �f to �f by analyzing the output at the fixed point v i ¼ =ðexpð�ui þ t iÞþ Þ. if c > is a constant and ui ¼ i þ i p c, then =ðexpð�c þ t iÞþ Þ v i < . when ui ¼ i þ i !þ v i ! , so in this case, if c is large enough, v i . if -c ui ¼ i þ i c, then =ðexpðc þ t iÞþ Þ v i =ðexpð�c þ t iÞþ Þ, according to equation (a). we can select a t i, that makes jt i þ p j¼ ð�ui þ t iÞ j =j! � p k¼ ð� Þ k expð�kðui t iÞÞj small enough, then v i i þ i . h. hu et al. / expert systems with applications ( ) – v i ¼ =ðexpð�ui þ t iÞþ Þ ¼ � expð�ui þ t iÞþ x k¼ ð� Þk expð�kðui � t iÞÞ ¼ ui � t i � x j¼ ð�ui þ t iÞ j =j! þ x k¼ ð� Þk expð�kðui t iÞÞ ¼ ui � t i � x j¼ ð�ui þ t iÞ j =j! þ x k¼ ð� Þk expð�kðui t iÞÞ:ðaÞ so in this case, v i i �f i ¼ minð ; i þ i Þ. similarly, if ui ¼ i þ i !� v i ! . so when c is large enough and ui ¼ i þ i �c < , then v i . when -c ui ¼ i þ i c, if we select a suitable t i which makes t i þ p j¼ ð�ui þ t iÞ j =j! � p k¼ ð� Þ k expð�kðui � t iÞÞ , then v i i �f i ¼ maxð ; i þ i � Þ. based on above analysis, the bounded operator fuzzy system is suitable for grc described by eq. ( ) when ai ¼ : ; w ¼ : and w ¼ : . for arbitrary positive ai; w and w , we can use corre- sponding q-value weighted universal fuzzy logical function based on bounded operator to simulate such kind neural cells. if a weight w is negative, a n-norm operator nðxÞ¼ � x should be used. experiments done by scanning the whole region of ði ; i Þ in ½ ; � to find the suitable coefficients for �f and �f show that above analysis is sound. we denote the input in eq. ( ) as ~x ¼ði ; i Þ. the ‘‘error’’ for �f and ‘‘errand’’ for �f are shown in fig. as the solid line and the dotted line respectively. in fig. , the threshold t i is scanned from to . with step size . . the best t i in eq. ( ) for �f is . and the best t i in eq. ( ) for �f is , when a ¼ : ; w ¼ : and w ¼ : . in this case the ‘‘error’’ and ‘‘errand’’ is less than . . our experiments show that suitable t i can be found. so in most cases, the bounded operator fð�f ;�fÞ mentioned above is the suitable fuzzy logical framework for the neuron defined by eq. ( ). if the weight < w and < w , we should use a q-value weighted bounded operator fð�f ;�fÞ to rep- resent above neuron. appendix c. associative condition and demorgan law of q- weighted bounded operator it is easily to see �f follows the associative condition and x �f x �f x . . .�f xn ¼ minðq; p i nwixiÞ. for �f , we can prove the associative condition is hold also. the proof is listed as below: if w p þ w p �ðw þ w � Þq p , we have: ðp �f p Þ�f p ¼f�f ðf�f ðp ;p ;w ;w Þ;p ; ;w Þ ¼f�f ðw p þw p �ðw þw � Þq;p ; ;w Þ ¼maxð ;w p þw p �ðw þw � Þqþw p �ð þw � ÞqÞ ¼maxð ;w p þw p þw p �ðw þw þw � ÞqÞ; if w p þ w p �ðw þ w � Þq < , we have ðp �f p Þ�f p ¼f�f ðf�f ðp ;p ;w ;w Þ;p ; ;w Þ ¼f�f ð ;p ; ;w Þ¼maxð ; þw p �ð þw � ÞqÞ ¼maxð ;w p �w qÞ ¼ for p q ¼maxð ;w p þw p þw p �ðw þw þw � ÞqÞ; so ðp �f p Þ�f p ¼ p �fðp �f p Þ¼ maxð ; w p þ w p þ w p �ðw þ w þ w � ÞqÞ. by inductive approach, we can prove that �f also follows the associative condition and x �f x �f x . . .�f xn ¼ maxð ; p i n wixi �ð p i nwi � ÞqÞ. for more if we define nðpÞ¼ q � p (usually, a negative weight wi corresponds a n-norm), above weighted bounded operator fð�f ;�fÞ follows the demorgan law, i.e. nðx �f x �f x . . .�f xnÞ¼ q � min q; x i n wi xi ! ¼ max ; q � x i n wixi ! ¼ max ; x i n wiðq � xiÞ�ð x i n wi � Þq ! ¼ nðx Þ�f nðx Þ�f nðx Þ . . .�f nðxnÞ: references andrzej, b., & pedrycz, w. ( ). the roots of granular computing. in grc (pp. – ). castro, j. l. ( ). fuzzy logic controllers are universal approximators. ieee transactions on systems, man and cybernetics, ( ), – . fung, g., & mangasarian, o. l. ( ). proximal support vector machine classifiers. in proceedings of the seventh acm sigkdd international conference on knowledge discovery and data mining (pp. – ). acm. haykin, s. ( ). neural networks: a comprehensive foundation. prentice hall ptr. haykin, s. ( ). neural networks: a comprehensive foundation. englewood cliffs: prentive-hall. he, k., sun, j., & tang, x. ( ). single image haze removal using dark channel prior. ieee transactions on pattern analysis and machine intelligence, ( ), – . levin, a., lischinski, d., & weiss, y. ( ). a closed-form solution to natural image matting. ieee transactions on pattern analysis and machine intelligence, ( ), – . lin, t. y. ( ). granular computing on binary relations i: data mining and neighborhood systems. rough sets in knowledge discovery, , – . lin, t. y. ( ). granular computing: fuzzy logic and rough sets. computing with words in information/intelligent systems (vol. , pp. – ). springer. lin, t. y. ( ). neighborhood systems: a qualitative theory for fuzzy and rough sets. berkeley: university of california. . lin, t. y. ( ). granular computing: practices, theories, and future directions. in computational complexity (pp. – ). springer. li, h.-x., & philip chen, c. l. ( ). the equivalence between fuzzy logic systems and feedforward neural networks. ieee transactions on neural networks, ( ), – . liu, h., xiong, s., & wu, c.-a. ( ). hyperspherical granular computing classification algorithm based on fuzzy lattices. mathematical and computer modelling. mountcastle, v. b. ( ). the columnar organization of the neocortex. brain, ( ), – . ojala, timo, pietikäinen, matti, & harwood, david ( ). a comparative study of texture measures with classification based on featured distributions. pattern recognition, ( ), – . pedrycz, adam, hirota, kaoru, pedrycz, witold, & dong, fangyan ( ). granular representation and granular computing with fuzzy sets. fuzzy sets and systems, , – . yao, y. y. ( ). relational interpretations of neighborhood operators and rough set approximation operators. information sciences, ( ), – . yao, y. y. ( ). granular computing using neighborhood systems. in advances in soft computing (pp. – ). springer. yao, y. y. ( ). granular computing: basic issues and possible solutions. in proceedings of the th joint conference on information sciences (vol. , pp. – ) citeseer. yao, y. y. ( ). on modeling data mining with granular computing. in th annual international computer software and applications conference, . compsac (pp. – ). ieee. yao, y. y. ( ). information granulation and rough set approximation. international journal of intelligent systems, ( ), – . yao, y. ( ). granular computing for data mining. in defense and security symposium (pp. ). international society for optics and photonics. yao, y., & deng, x. ( ). a granular computing paradigm for concept learning. in emerging paradigms in machine learning (pp. – ). springer. zadeh, l. a. ( ). fuzzy sets. information and control, ( ), – . http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h h. hu et al. / expert systems with applications ( ) – zadeh, l. a. ( ). toward a theory of fuzzy information granulation and its centrality in human reasoning and fuzzy logic. fuzzy sets and systems, ( ), – . zhang, l., & zhang, b. ( ). theory of fuzzy quotient space (methods of fuzzy granular computing). journal of software, ( ), – . zhang, l., & zhang, b. ( a). the quotient space theory of problem solving. fundamenta informaticae, ( ), – . zhang, l., & zhang, b. ( b). the quotient space theory of problem solving. fundamenta informaticae, ( ), – . zhang, l., & zhang, b. ( ). quotient space model based hierarchical machine learning. international conference on neural networks and brain, . icnn&b’ (vol. ). ieee, pp. xiv–xiv. http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h perception granular computing in visual haze-free task introduction granular system based on tolerance relation hybrid designing of leveled perception granular system based on fuzzy logic and psvm granular system for visual task . the theory of image matting . leveled perception granular system for haze-free task . . the rd layer – alogical layer . . the th layer – fuzzy logical layer . experiments result discussion acknowledgments appendix a sigmoid function and binary logic appendix b sigmoid function and fuzzy logic appendix c associative condition and demorgan law of q-weighted bounded operator references this article appeared in a journal published by elsevier. the attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. in most cases authors are permitted to post their version of the article (e.g. in word or tex form) to their personal website or institutional repository. authors requiring further information regarding elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/authorsrights http://www.elsevier.com/authorsrights author's personal copy proximity measures for link prediction based on temporal events paulo r.s. soares, ricardo b.c. prudêncio ⇑ center of informatics (cin), federal university of pernambuco, recife, pe, brazil a r t i c l e i n f o keywords: link prediction temporal events neighborhood-based measures co-authorship networks a b s t r a c t link prediction is a well-known task from the social network analysis field that deals with the occur- rence of connections in a network. it consists of using the network structure up to a given time in order to predict the appearance of links in a close future. the majority of previous work in link prediction is focused on the application of proximity measures (e.g., path distance, common neighbors) to non-con- nected pairs of nodes at present time in order to predict new connections in the future. new links can be predicted for instance by ordering the pairs of nodes according to their proximity scores. a limitation usually observed in previous work is that only the current state of the network is used to compute the proximity scores, without taking any temporal information into account (i.e., a static graph representa- tion is adopted). in this work, we propose a new proximity measure for link prediction based on the con- cept of temporal events. in our work, we defined a temporal event related to a pair of nodes according to the creation, maintenance or interruption of the relationship between the nodes in consecutive periods of time. we proposed an event-based score which is updated along time by rewarding the temporal events observed between the pair of nodes under analysis and their neighborhood. the assigned rewards depend on the type of temporal event observed (e.g., if a link is conserved along time, a positive reward is assigned). hence, the dynamics of links as the network evolves is used to update representative scores to pairs of nodes, rewarding pairs which formed or preserved a link and penalizing the ones that are no longer connected. in the performed experiments, we evaluated the proposed event-based measure in different scenarios for link prediction using co-authorship networks. promising results were observed when the proposed measure was compared to both static proximity measures and a time series approach (a more competitive method) that also deploys temporal information for link prediction. � elsevier ltd. all rights reserved. . introduction as the amount of interaction among individuals increases in vir- tual environments, more useful social data become available, serv- ing as a basis for social network analysis. social networks are structures composed by individuals (entities) that can be con- nected by different forms of social relationship (such as friendship, in which two individuals are connected if they are friends) (ama- ral, scala, barthelemy, & stanley, ). in social networks, connec- tions and entities tend to appear and disappear along time, which turns them into highly dynamic and complex systems. social net- work analysis (sna) is a broad field of research that tries to deal with such complexity (wasserman & faust, ). several tasks can be associated to sna. in this paper, our specific aim is to inves- tigate the prediction of links in a social network, that is, we are fo- cused on predicting the most probable future connections based on previous states of the network. this is a well-known problem from the sna field called link prediction (hasan, chaoji, salem, & zaki, ). a lot of work has been devoted to cope with the link prediction problem (getoor & diehl, ; wang, satuluri, & parthasarathy, ; xiang, ). the majority of previous work is based on the application of proximity measures to non-connected pairs of nodes in the network at present time in order to predict new con- nections at future time. the proximity measures are used to asso- ciate scores to the pairs of nodes, which can be used either: ( ) in a unsupervised approach, in which a chosen score is simply ordered and links are predicted for the top ranked pairs; or ( ) in a super- vised approach, in which the link prediction is treated as a classi- fication task and different scores are used as predictor attributes by a learning algorithm. a limitation that can be pointed out in the previous work is that the proximity scores are usually calcu- lated without taking into account the evolution of the network. the proximity measures are computed using all network data up to the present moment (i.e., the current network state) without considering when links were created. hence, a potentially useful source of information for link prediction is not adequately leveraged. - /$ - see front matter � elsevier ltd. all rights reserved. http://dx.doi.org/ . /j.eswa. . . ⇑ corresponding author. tel.: + . e-mail addresses: prss@cin.ufpe.br (p.r.s. soares), rbcp@cin.ufpe.br (r.b.c. prudêncio). expert systems with applications ( ) – contents lists available at sciverse sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a author's personal copy in the current work, we propose a novel proximity measure in which temporal information is taken into account. in the proposed measure, we deployed the concept of temporal events, which are specific activities (e.g., creation or removal of a link) observed be- tween a pair of nodes in consecutive time intervals. for instance, an innovative event occurs when two nodes are not connected in a given time interval and a new link is created between them in the next interval. the proximity score for a pair of nodes is com- puted by monitoring along time the events observed around the nodes and their immediate neighbors. each category of temporal event defined in our work (innovative, conservative and regressive) is associated to a numerical reward and the proximity score in- creases or decreases along time depending on the temporal events observed in consecutive network frames. our proposal was sup- ported by the homans’ work that associates the strength of a con- nection between individuals to their frequency of interaction (which was modeled in our work by the temporal events) (homans, ) and the idea shared by many authors that a bigger common neighborhood between individuals is closely related to a higher probability of future connections (newman, ). in order to verify the viability of the proposed measure, we per- formed experiments on co-authorship networks extracted from four sections of the physics e-print arxiv. in these experiments, the proposed measure was evaluated in different scenarios by con- sidering for instance different reward values for the temporal events and by adopting a weighting function to give higher importance to more recent events. for a baseline comparison, experiments were performed with traditional proximity measures previously adopted in the literature. we also performed experiments with a time series approach (potgieter, april, cooke, & osunmakinde, ) which, similarly to our approach, also aims to consider temporal informa- tion to improve link prediction. in all experiments, the unsupervised approach was adopted to perform the prediction task once the scores were calculated and the area under the roc curve (auc) was used to evaluate the performance of prediction. the performance achieved by the event-based measure outperformed the results of the baseline methods in all the networks considered. section briefly presents the link prediction problem. section describes the event-based approach, showing the concepts used and the score calculation process. section presents the social net- work data adopted, the experiments and obtained results. finally, section concludes this work by presenting some final consider- ations and future work. . link prediction link prediction consists of predicting new connections or detecting hidden links in a network. it is a very important task applicable to a wide variety of areas, such as bibliographic domain, molecule biology, criminal investigations and recommending sys- tems (getoor & diehl, ; xiang, ). a traditional definition of the link prediction problem is expressed by: ‘‘given a snapshot of a social network at time t, we seek to accurately predict the edges that will be added to the network during the interval from time t to a given future time t ’’ (liben-nowell et al., ). among several approaches to treat the problem, the most widespread ones rely on the use of proximity measures between pairs of nodes (lu & zhou, ; xiang, ). as previously mentioned, the starting point of this approach is to extract the values/scores of different measures that indicate the similarity between pairs of nodes. these scores can be used either by unsupervised (liben-nowell et al., ; lu & zhou, ; murata & moriyasu, ) or supervised link prediction (hasan et al., ; lichtenwalter, lussier, & chawla, ; sá & prudêncio, ). in the former approach, a proximity measure is chosen and deployed to rank node pairs in the network. the top ranked ones are predicted to be linked. in the supervised link prediction, a set of proximity measures is cho- sen and adopted as predictor attributes by a classifier, learned based on historical data. each pair of non-connected nodes is de- scribed by the set of proximity scores. a classifier then uses these attributes to perform a binary classification to decide whether the link will be formed or not in the future. the proximity measures proposed and evaluated in the litera- ture can be broadly categorized into semantic or topological mea- sures (xiang, ). in the semantic measures (or node-wise measures), the nodes’ content is considered to measure proximity. for instance, in a co-authorship network, the similarity between keywords extracted from published papers can be used to predict future interaction among the authors (xiang ( )). different from the semantic measures, the topological strategy consists of deploying the network structure to compute the proximity scores (e.g. the number of common neighbors that two nodes share). topological measures are more commonly adopted in the litera- ture since they are more general and do not require the definition of rich features to describe content (in fact, rich content is not al- ways available depending on the social network considered). several topological measures were proposed in the literature mainly categorized into neighborhood-based or path-based mea- sures (hasan & zaki, ). the neighborhood-based measures take into account the immediate neighbors of the nodes. in general, these measures consider that two nodes are more likely to form a link if their sets of neighbors have a large overlap (xiang, ). among the neighborhood-based measures, we can mention common neighbors (newman, ), preferential attachment (barabasi et al., ; newman, ),adamic and adar ( ) and the jaccard’s coefficient (salton & mcgill, ). the path- based measures in turn define proximity between nodes by consid- ering the paths between them. the basic idea is that two nodes are more likely to form a link if there are short paths between them. the path-based measures range from the simple path-distance measure to more sophisticated definitions that consider ensembles of different paths, such as the katz measure (katz, ). in com- parative terms, the neighborhood-based methods are more wide- spread, due to both their computational eficiency and great performance observed in experiments (huan, ; liben-nowell et al., ; murata & moriyasu, ). the measure proposed in our work can be categorized as neighborhood-based, since it uses information about the connections around nodes to assign scores to them (see section ). most previous work performs link prediction by statically ana- lyzing the network data, that is, no temporal information is consid- ered to perform the prediction. however, temporal information (e.g., the moments when two nodes interacted in the past or the time when a connection was first observed) is an important aspect that should be considered during the link prediction (hasan & zaki, ). for instance, when computing a proximity score based on neighborhood, it would be interesting to consider not only how many but also when the links were formed between neighbors. recent activity between common neighbors may be more impor- tant than older activity. also, static approaches are appropriate to investigate whether a certain link will ever occur in a network but they are less useful, for instance, to applications for which the prediction of repeated link occurrences are of interest (huang & lin, ). in this section, we mention some previous work concerning the use of temporal information for link prediction. for instance, in tylenda, angelova, and bedathur ( ), the authors modeled a network as a weighted graph, in which the weight of a link was the age of the most recent activity between the corresponding http://www.arxiv.org p.r.s. soares, r.b.c. prudêncio / expert systems with applications ( ) – author's personal copy nodes. the link prediction task was then accomplished by deploy- ing extended versions of the proximity measures adequate for weighted networks (e.g., weighted adamic adar). bringmann, berlingerio, bonchi, and gionis ( ) proposed to mine network data augmented with temporal information in order to discover association rules (in terms of frequent subgraphs) that best ex- plained the network evolution. in juszczyszyn, musial, and budka ( ), the authors adopted a related approach in which the his- tory of the network (recorded during past time windows) is used to derive probabilities of transitions between triads of nodes. although promising results can be obtained in this approach, min- ing frequent subgraphs has shown to be a very expensive task (coenen, jiang, & zito, ). an alternative approach for deploying temporal information is to treat link prediction as a time series forecasting problem (huang & lin, ; potgieter et al., ; qiu, he, & yen, ; soares & prudêncio, ). huang and lin ( ) built a time series for each pair of nodes, in which each series observation is the frequency of occurrence of links between the nodes during a specific time peri- od. the time series forecasts produced by arima models (box & jenkins, ) were then used to estimate the probability of future link occurrence. in potgieter et al. ( ) and soares and prudêncio ( ), the authors adopted a similar idea, but in this case time series models were used to predict proximity scores. in this ap- proach, given a chosen proximity measure, a time series is built for each pair of nodes by computing the score in a sequence of time periods (i.e., using different snapshots of the network along time). a final proximity score is then obtained for each pair by forecasting its corresponding time series. different models were applied in the forecasting process, including arima models, smoothing methods and linear regression models. despite the good results obtained in experiments, a difficulty in the time series approach is to choose an adequate forecasting model, among the variety of models that can be applied. in fact, in soares and prudêncio ( ), the authors ob- served that a wrong decision concerning the forecasting model can have a negative impact in the link prediction performance. . event-based link prediction social networks are highly dynamic structures in which several connections and nodes tend to appear or disappear along time. the temporal evolution of these networks brings valuable information about how connections tend to be formed, and, for that, should be considered in the link prediction task. in the current work, we pro- pose a new proximity measure that takes into account the tempo- ral structure of a network and temporal events related to pairs of nodes in the network. the aim of our proposal is to combine ho- mans’ thought that the strength of a connection between two indi- viduals is directly associated with how often they interact with one another (modeled here by means of temporal events) (homans, ) and newman’s idea that the bigger the number of common neighbors between two nodes, the higher is their probability to be connected in the future (newman, ). the general idea of the proposed approach is to increase or de- crease the proximity score between two nodes depending on tem- poral events observed among the two nodes and their neighborhood. a temporal event, which will be explained better in the next section, is defined through the appearance or disap- pearance of links around nodes as the network evolves. in the pro- posed work, initially a temporal structure is created by extracting consecutive frames of the network at different time intervals. a proximity score is then computed for each pair of nodes by aggre- gating the rewards assigned for the temporal events observed dur- ing the transitions of frames. the proposed measure is explained in more detail in the next subsections. . . temporal structure in the proposed solution, we initially build a temporal structure n , by following a methodology described in soares and prudêncio ( ). first of all, the network is split into several time-sliced snapshots, which represent states of the network at different time intervals in the past. after that, frames are built by grouping con- secutive snapshots. the size of each frame is the same as the length of the prediction window, pre-defined in the link prediction task. each frame represents one time step in n . this methodology is detailed below. let gðv; eÞ be a graph representing a social network observed up to time t. each edge in e is represented by a triple ðu; v; tÞ, indicat- ing that the nodes u and v v had a social interaction at time t. hence, more than one edge related to a single pair of nodes can be observed in e if the nodes interacted in different moments in the past. let w be the length of the prediction window, i.e., our task is to predict new links concerning the future time interval from t þ to t þ w. let gt be the sub-graph of g containing only the edges observed at time t. let ½gt; gtþ ; . . . ; gtþm� be the frame formed by the disjoint union of the graphs from time t to t þ m. in our work, a set of n con- secutive frames n ¼ff ; . . . ; f k; . . . ; f ng of size w is extracted from the graph g. formally, n can be defined as n ¼f ½gt�nwþ ; gt�nwþ ; . . . ; gt�ðn� Þw�; . . . ; ½gt� wþ ; gt� wþ ; . . . ; gt�w�; ½gt�wþ ; gt�wþ ; . . . ; gt�g ð Þ the frame fk ¼ ½gt�ðn�kþ Þwþ ; . . . ; gt�ðn�kÞw� is hence the sub-graph of g containing all links observed in the k-th time interval of size w. as an example, consider a network observed up to the year t ¼ and a prediction window of length (i.e., the task is to predict new links from and ). if we extract n ¼ frames from the network, the following structure (n ¼ff ; f ; f g) is obtained n ¼f½g ; g �; ½g ; g �; ½g ; g �g ð Þ this structure can be used to analyze how the network evolved every years since . for this, we will introduce next section the concept of temporal events, that can be observed between sub- sequent frames. . . temporal events we define a temporal event as a specific activity between two nodes from a frame to its subsequent. a temporal event is the ac- tion that leads a dyad (a pair of nodes) from a state (connected or non-connected) to another. events can be categorized into one of three mutually exclusive types: conservative, innovative or regres- sive, defined below. � conservative a conservative event occurs when a relationship between two nodes is not dropped when the network evolves, that is, when two nodes share a link in a frame and this connection is pre- served in the subsequent one. for each pair of nodes ðu; vÞ, we define a reward cðu; v; kÞ related to frame fk , in order to take into account a conservative event during the transition from the ðk � Þth to the kth frame. formally: cðu; v; kÞ¼ c; if ðu; vÞ ek� \ ek ; otherwise � ð Þ in the above definition, ek� and ek are the sets of edges observed in frames fk� and fk respectively. the constant c indicates the reward for conservative events, which should be a non-negative value since the strength of a tie between two nodes is preserved. p.r.s. soares, r.b.c. prudêncio / expert systems with applications ( ) – author's personal copy � innovative innovative events represent the creation of a new link between two nodes on different frames. they happen when two nodes are not connected in a frame and a link is observed in the next frame. the innovative reward iðu; v; kÞ associated to a pair ðu; vÞ and a frame fk is: iðu; v; kÞ¼ i; if ðu; vÞ ek n ek� ; otherwise � ð Þ the constant i in the above equation indicates the reward for inno- vative events. its value should be positive since the tie between two nodes is strengthened. � regressive regressive events are opposite to innovative ones. this kind of event represents the removal of an existing link between two nodes from a frame to its subsequent. the regressive reward rðu; v; kÞ is defined as: rðu; v; kÞ¼ r; if ðu; vÞ ek� n ek ; otherwise � ð Þ in this event, r should assume a non-positive value since the strength of the connection between the nodes tends to decrease. the values of the parameters c; i and r can be determined empir- ically by evaluating the link prediction performance in a validation set and choosing the best configuration of values considered. as it will be seen, in our experiments, the number of parameters to define was reduced to two by making the values of c and r propor- tional to i (in this case, i was set to for the sake of simplicity). as it will be seen in the next subsection, given a pair of nodes, we will define as primary events the ones strictly related to both nodes in the pair, whereas secondary events happen in dyads com- posed by only one of the nodes in the pair and its neighbors. . . event-based score as previously said, many approaches for link prediction com- pute scores to pairs of nodes by deploying a chosen proximity mea- sure (see section ), aiming to determine how similar those nodes are and, consequently, how likely a connection between them will be formed in a close future. the proposed measure combines both: ( ) the rewards associ- ated to primary events, which are the temporal events strictly re- lated to the pair of nodes under analysis; and ( ) the rewards associated to secondary events, which are the temporal events ob- served in the nodes’ neighborhood. the proximity score associated to a given pair of nodes ðu; vÞ is defined as: scoreðu; vÞ¼ xn k¼ pðu; v; kÞþ a:sðu; v; kÞ ð Þ pðu; v; kÞ¼ cðu; v; kÞþiðu; v; kÞþrðu; v; kÞ ð Þ sðu; v; kÞ¼ x y cðuÞ\cðvÞ pðu; y; kÞþ pðy; v; kÞ ð Þ in eq. ( ), pðu; v; kÞ computes the reward of the event (conservative, innovative or regressive) for the pair of nodes ðu; vÞ observed in the transition from frame k � to frame k. in eq. ( ), sðu; v; kÞ indicates the aggregated reward of secondary events associated to the pair ðu; vÞ (i.e., the primary events observed in its common neighbor- hood). in this equation, cðxÞ is the set of neighbors of the node x in the network. in eq. ( ), the parameter a is an amortization factor that indi- cates how strong secondary events affect the tie between u and v. pðu; v; kÞ and sðu; v; kÞ associated to the first frame (i.e., k ¼ ) are all null since this frame is not a result of any set of events, and hence for simplicity, they are not considered in eq. ( ). fig. illustrates the application of the proposed event-based score. suppose that one wants to compute the proximity score for the pair ð ; Þ. the nodes and have node as common neighbor in the network. hence, the score of the connection between and is going to be calculated as a function of the events that occurred between themselves (primary events), plus the ones occurred in the dyads ð ; Þ and ð ; Þ (secondary events) along the network temporal structure n . from frame f to frame f , it can be noticed that a conservative and a regressive event happened in the dyad ð ; Þ and ð ; Þ respectively. so far, no event occurred in the dyad ð ; Þ. therefore, up to frame f , the partial score of the pair ð ; Þ is given by the combination of the amortized rewards associated with the second- ary events described above, that is, aðc þ rÞ. looking at the next time step (frame f ), a regressive event is associated to ð ; Þ, whereas an innovative one occurred in the pair ð ; Þ. in this step, there is also an innovative event related to the pair under analysis ð ; Þ; its presence must be counted too in the final score of the pair (its reward, however, will not be amor- tized since it is a primary event). hence, the resultant score in this frame is given by aðr þ iÞþ i. by the definition of eq. ( ), the scores are cumulative in time. the final score of a pair of nodes is, then, the sum of all its partial scores along n . this, therefore, results in a final score of aðc þ r þ iÞþ i for the pair ð ; Þ. in the current work, we also propose an extension of the event- based score that increases the importance of more recent events observed for a pair of nodes. in eq. ( ), we propose a proximity score which considers a monotonically increasing function b : z ! r to weight recent events more heavily than the old ones. in our work, we adopted a logarithmic function (see eq. ( )), although other functions can be deployed. we highlight that eq. ( ) is a special case of eq. ( ) by setting bðkÞ¼ . scoreðu; vÞ¼ xn k¼ bðkÞ:½pðu; v; kÞþ a:sðu; v; kÞ� ð Þ bðkÞ¼ logðkÞ ð Þ . . prediction finally, in order to perform the prediction of new links, we adopted an unsupervised strategy after the scores were computed for the pairs of nodes. it basically consists of ranking the pairs of nodes according to their scores, and then, selecting the top-ranked ones as the new links in the network, as explained in section . fig. . example of event-based score computation using a network composed by a click of three nodes. n represents the temporal structure of the network divided in three frames (f ; f and f ). the variables c; i and r are the rewards associated with conservative, innovative and regressive events respectively, and a is the amorti- zation factor used to deal with secondary events. p.r.s. soares, r.b.c. prudêncio / expert systems with applications ( ) – author's personal copy . experiments and results in this section we describe the experiments performed to eval- uate the proposed approach as well as the obtained results. ini- tially we describe the social network data used and the methodology of experiments (section . ), followed by the base- line methods adopted for a comparative evaluation (section . ). in section . , we present the initial results aimed to evaluate the proposed measure considering different configurations of re- wards for the temporal events. in section . , we evaluate the influence of the amortization factor a in the proposed measure. fi- nally, in section . , we present the effect of using a weighting function for recent event as well as the comparative results with the baseline methods. . . data and settings in order to evaluate the performance of the proposed scores, data from four co-authorship networks were used in our experi- ments. we highlight that co-authorship networks are highly used in the literature to evaluate sna techniques. being a social struc- ture, this kind of network is very dynamic along time, what makes it a great source of data to be explored regarding the emergence of new connections. besides, most of them are publicly available in digital environment (barabasi et al., ; newman, ). in a co-authorship network, a node represents an author and an edge indicates that two specific nodes have co-authored a paper. as seen in section , in our work an edge also stores the time when the relationship was observed. more specifically, in the co-authorship networks adopted here, each edge stores the publication year of the co-authored paper. the network data used in our experiments were collected from the e-print arxiv, which maintains a large database of electronic scientific papers in several fields, such as mathematics, physics, astronomy, among others. in our experiments, the networks were built using collaborations from four distinct sub-areas, namely: as- tro-physics (astro-ph), condensed matter (cond-mat), high energy physics – lattice (hep-lat) and theoretical high energy physics (hep- th). collaborations from to were used for astro-ph and cond-mat, and collaborations from to were used for hep- lat and hep-th. the summarized information about the adopted net- works are presented in table . in our experiments, a prediction window of size one was adopted, and hence the aim of the analysis is to investigate how the networks evolved every year. by following the methodology described in section . , a temporal structure n was built for each network. for astro-ph and cond-mat, data from to were used to build n (with a total number of n ¼ frames) and the snapshot from was used as the prediction frame. for hep-lat and hep-th, n was built with collaborations from to (with a total number of n ¼ frames), whereas the prediction frame was composed by collaborations from . as mentioned in section . , adequate values for the reward parameters can be estimated by evaluating the link prediction per- formance on a validation set. in our work, the last frame of n was adopted as the validation set in order to perform parameter selec- tion for each network: the frame of was used to validation for the astro-ph and cond-mat networks and the frame of was used to validation for hep-lat and hep-th. as usual in the link prediction field, our aim is to investigate the prediction of new links in a network. therefore, for each network, the initial list of candidate pairs to be evaluated is formed by pairs that are not connected in the last frame of n . the final list is formed by picking (from the initial list) only the dyads that were directly or indirectly affected by some event during the network evolution. table shows the distribution of the candidate pairs of nodes regarding their class labels (connected or non-connected) both in the validation and the test sets. as it can be seen, the class distribution is highly imbalanced, which is a common problem related to the link prediction task. in order to minimize the effect of imbalanced data during evalua- tion, the area under roc curve (auc) was adopted as the measure of performance. roc curves relate the sensitivity (true positive rate) and specificity (true negative rate) of a classifier. the auc has been traditionally used in imbalanced classification problems (murata & moriyasu, ; wang et al., ) due to its higher robustness to class distribution anomalies (fawcett, ). finally, for more reliable results, we performed a bootstrap over the test set by taking randomly stratified sub-samples and computing the auc. the % of the best results and % of the worst results were discarded and the average auc was computed. . . baseline methods in our experiments, the proposed measures were compared with different measures previously adopted in the literature of link prediction. initially, four classical proximity scores based on neigh- borhood were considered: pa, cn, aa and jc. these are very popu- lar measures used in the state of the art of link prediction (bringmann et al., ; lichtenwalter et al., ). each measure is explained below. the pa measure assumes that the probability of a future link between two nodes is proportional to their degrees. in a co-author- ship network, such probability is correlated to the product of the number of collaborators they have barabasi et al. ( ). hence, it is defined as: paðu; vÞ¼ jjcðuÞjj � jjcðvÞjj ð Þ the cn measure states that the bigger the number of neighbors two nodes share, the higher is their probability to form a link in the future (newman, ). formally, the measure is defined as: cnðu; vÞ¼ jjcðuÞ\ cðvÞjj ð Þ the aa measure refines the cn by increasing the scores of pairs of nodes in which the neighbors in common possess less connections (adamic & adar, ). its formal definition is: table networks statistics. astro-ph cond-mat hep-lat hep-th papers , , , , authors , , , , collaborations , , , , density ( � ) . . . . avg. degree . . . . table class distribution for the validation and test sets. astro-ph cond-mat hep-lat hep-th (a) validation set no. of pairs , , , , connected (+) , , , , non-connected (�) , , , , proportion (+/�) . % . % . % . % (b) test set no. of pairs , . , , , connected (+) , , , non-connected (�) , . , , , proportion ((+/�)) . % . % . % . % arxiv.org e-print archive – cornell university library. p.r.s. soares, r.b.c. prudêncio / expert systems with applications ( ) – author's personal copy aaðu; vÞ¼ x z cðuÞ\cðvÞ logðjjcðzÞjjÞ ð Þ finally, the jc assumes higher proximity values for pairs of nodes which share a higher proportion of common neighbors relative to to- tal number of neighbors they have salton and mcgill ( ). jcðu; vÞ¼ jjcðuÞ\ cðvÞjj jjcðuÞ[ cðvÞjj ð Þ the above measures were statically applied to the whole graph g (i.e., in the current state of the network) without considering tem- poral information. in our work, we also performed experiments with the time series approach for link prediction (potgieter et al., ; soares & prudêncio, ) described in section . as our cur- rent work, this approach also considers temporal information dur- ing the link prediction. initially the network is split into frames, using the same procedure described in section . . given a chosen proximity measure, a time series is built for each pair of nodes by applying the measure in each network frame. a forecasting model is then used in order to predict the next value of the series. such forecast value is defined as the final proximity score of the pairs of nodes. compared to the proposed work, we can point out a diffi- culty related to the time series approach, which is the choice of a good forecasting model. in fact, in the experiments performed in soares and prudêncio ( ) using co-authorship networks, the choice of the forecasting model strongly affected the link prediction performance. in our work, we performed experiments adopting aa as the measure to build the time series and the linear regression (lr) model as the forecasting model. that was the best combination of proximity measure and forecasting model in the experiments presented in soares and prudêncio ( ). finally, we highlight that similarly to the proposed measures, the baseline measures described in this section (both the static ones and the time series approach) were applied for unsupervised link prediction. . . first-round experiments as described in section . , the event-based measure is poten- tially affected by the rewards associated to each of the three events: conservative, innovative and regressive. in the first round of experiments, we aimed: ( ) to evaluate the effect of rewards’ values in the performance of the proposed measure; and ( ) to ver- ify whether these values can be properly chosen using a validation set. we deployed the validation set to empirically evaluate and select the best combination of rewards in the link prediction task. in this round, as mentioned in section . , i was fixed to . and the other two parameters c and r varied proportionally to i : c ¼f : ; : ; : ; : ; : g and r ¼f� : ;� : ;� : ;� : g. we consider here a fixed value for a, since our focus was to inves- tigate the adopted rewards’ values. the value of a was set to . through a preliminary and exploratory battery of experiments. table shows the auc value obtained by each combination of the rewards in the validation set for all networks. the best four ob- served results for each network are presented in bold. as it can be seen, the most effective combinations of reward did not vary so much from one network to another. the best results were obtained by parameters (c and r) around the same region (lower central) in the tables for all networks. besides, these results indicate that conservative and regressive rewards can balance each other. the best performance values were achieved when c i, although some good results were also obtained when we consid- ered a higher weight ( . ) to the conservative reward. in our exper- iments, if regressive reward is roughly weighted, the event-based method did not perform well (this can be noticed by analyzing results obtained with r ¼� : and � : ). as one could expect, this shows that if the aim is to predict new links in a network, then the history of innovative and conservative events assumes a higher de- gree of importance in the prediction process. table shows the auc value obtained by each combination of the rewards in the test set for all networks. the results observed in the test sets did not varied substantially from those ones ob- served in the validation set. the best configurations of parameters were also similar considering the validation and the test set (see table ). this supports the conclusions we made regarding rewards and indicates that they can be determined empirically using a val- idation set. . . second-round experiments in the second round of experiments, our aim was to investigate how the amortization factor a affects the performance of the event-based measure. as explained in section . , this parameter indicates the importance of the secondary events in relation to pri- mary events in our proposed measure. in this experiment, we eval- table aucs obtained by different combinations of rewards on the validation sets. r=c . . . . . (a) astro-ph � . . . . . . � . . . . . . � . . . . . . � . . . . . . (b) cond-mat � . . . . . . � . . . . . . � . . . . . . � . . . . . . (c) hep-lat � . . . . . . � . . . . . . � . . . . . . � . . . . . . (d) hep-th � . . . . . . � . . . . . . � . . . . . . � . . . . . . table aucs obtained for several combinations of c and r on the test sets. r=c . . . . . (a) astro-ph � . . . . . . � . . . . . . � . . . . . . � . . . . . . (b) cond-mat � . . . . . . � . . . . . . � . . . . . . � . . . . . . (c) hep-lat � . . . . . . � . . . . . . � . . . . . . � . . . . . . (d) hep-th � . . . . . . � . . . . . . � . . . . . . � . . . . . . p.r.s. soares, r.b.c. prudêncio / expert systems with applications ( ) – author's personal copy uated the proposed measure by adopting different values of a ranging from to : , with an increment of : . we highlight that when a ¼ , secondary events are in fact not taken into account. for higher values (such as a ¼ : ) in turn, possibly an excessive importance is given to secondary events, which may harm the prediction performance. in this section, the rewards c; r and i were fixed according to the best configuration observed in the validation set in the previous round of experiments. the auc values obtained as a varies are illustrated in fig. . when a assumes a null value, the performance of the model is harmed because secondary events are not being taking into account in the prediction task. in this case, the measure will only be able to predict connections between pairs affected by primary events any time in the past, that is, pairs that were connected in at least one frame of n . this shows the effectiveness of secondary events in the prediction of new connections. the best auc values were obtained by adopting a ¼ : , which is a relatively small value considering the range of values evaluated in our experiments. when it approaches to higher values, the method performs worse since surrounding pairs are becoming as important as the main pair of nodes in analysis. besides, the num- ber of secondary events is much bigger than the total amount of primary events (mainly if the network is dense). hence, isolated secondary events should contribute with small values, which when aggregated, can outstand in the predictive process without, how- ever, mask the effects of the primary events. in this section, we suggest a simple heuristic to define the value of a based on the ratio between the total number of primary events and secondary events in n (see table ). as it can be seen, the auc results obtained by adopting the heuristic value for a were similar to the best results observed in fig. . . . last-round experiments in the previous experiments we evaluated the event-based score focusing on its parameters. in this section, we evaluated it aiming to determine if the age of the temporal events (event-based score with eq. ( )) can bring any gain of performance to the predic- tion task. aditionally, we compared the event-based scores to the baseline measures described in section . . table present the table best rewards settings extracted from experiments on validation and test sets. astro-ph cond-mat hep-lat hep-th (a) validation set c . . . . i . . . . r � . � . � . � . (b) test set c . . . . i . . . . r � . � . � . � . fig. . plot of auc versus the amortization factor a. p.r.s. soares, r.b.c. prudêncio / expert systems with applications ( ) – author's personal copy auc value obtained by the proposed scores and the baseline meth- ods. in this table, we refer the function assigned to b to distinguish between the scores generated by using the two aforementioned equations. the results showed that the methods that deploy temporal information (i.e., the event-based scores and the time series approach) in general outperformed the classical static approach based on pa, cn, aa and jc. these results confirm that temporal information is actually an important aspect to consider for link prediction. the event-based scores obtained the best results from all methods considered. the performance gain was higher for the cond-mat, hep-lat and hep-th networks, but less expressive for the as- tro-ph network, in which the event-based score was quite similar to aa alone. in our results, we observed a performance gain when the event-based score using b ¼ log was compared to the event-based score using b ¼ in all networks considered, what supports our belief that recent events carry more information about the emer- gence of links.. although, the performance gain was in general small in absolute values, it was statistically verified using a t-test with a % level of confidence. finally, although both the time series approach and the event- based methods explore the temporal nature of the network, the proposed approach obtained the best comparative results. the per- formance gain was statistically verified with a % level of confi- dence. the proposed measure analyses the network by focusing on the connections as such, and for that, it was able to extract more relevant information to the prediction of ties that might emerge in a close future. the method based on time series, on the other hand, assumes that the topological measure history shows some ten- dency as the network evolves. thus, it performs an indirect analy- sis of the connections in the network. in most cases, the time series approach outperformed the static approach, but it was not able to overcome the results achieved by an analysis strictly focused on the links, as done by the method based on events. . conclusion in this work we introduced a new proximity measure for link prediction based on the notion of temporal events. differently from the classical static approach that takes the network at time t in order to predict new connections at future time t , our method takes temporal information into account by monitoring how the network evolved along time. our method consists of computing scores by aggregating rewards of the temporal events observed at each step in the network evolution, which resulted in more rep- resentative scores to the dyads under analysis. different experi- ments were performed in four co-authorship networks, initially aimed to evaluate the robustness of the proposed method concern- ing its parameters. the obtained results suggest that the method was quite stable concerning the definition of its parameters across the different networks adopted, although more experiments need to be performed in other domains of application. also, we verified the importance of considering secondary events in the proposed measure. in the performed experiments, we also compared the event- based score to other baseline approaches proposed in the literature of link prediction. the obtained results showed that the event- based measure outperformed the baseline measures considered, indicating that the combination of interactions between nodes and the dynamics of common neighborhood can bring a gain in performance to the prediction task. additionally, when the age of the temporal events was taken into account, the method per- formed slightly better, indicating that more recent events can bring more useful information to the prediction process. in our work, we adopted a log function in order to give a higher weight to more recent events. nevertheless, other functions can be evaluated in the future (e.g., linear) to weight the recent events. also, different functions can be associated to each category of temporal event if we assume for instance that recent events from one category are actually more important that the other events. finally, we highlight that the event-based score described here was specific to undirected and unweighted networks. a possible future line of research is to extend the ideas discussed here to more complex categories of networks. for instance, a variety of temporal events could be defined if the direction of the links is considered (e.g., a conservative event is observed in a direction but a regres- sive event is observed in the other direction, thus indicating a non-reciprocal interaction between the nodes at a given moment). also, in the case of weighted networks, temporal events could con- sider an increase or decrease of the weight between the nodes along time, instead of just considering the existence of the links. the rewards in such cases should be carefully defined in order to adequately reflect the important of the events in the link predic- tion process. references adamic, l. a., & adar, e. ( ). friends and neighbors on the web. social networks, ( ), – . amaral, l. a., scala, a., barthelemy, m., & stanley, h. e. ( ). classes of small- world networks. proceedings of the national academy of sciences usa, ( ), – . barabasi, a-l., jeong, h., neda, z., ravasz, e., schubert, a., & vicsek, t. ( ). evolution of the social network of scientific collaborations. physica a, ( – ), – . box, g., & jenkins, g. ( ). time series analysis: forecasting and control. san francisco: holden-day. bringmann, b., berlingerio, m., bonchi, f., & gionis, a. ( ). learning and predicting the evolution of social networks. intelligent systems ieee, ( ), – . de sá, h. r., & prudêncio, r. b. c. ( ). supervised link prediction in weighted networks. in proceedings of the international joint conference on neural networks (pp. – ). ieee. fawcett, t. ( ). an introduction to roc analysis. pattern recognition letters, , – . getoor, l., & diehl, c. p. ( ). link mining: a survey. sigkdd explorations newsletter, ( ), – . hasan, m. a., chaoji, v., salem, s., zaki, m. ( ). link prediction using supervised learning. in proceedings of sdm workshop on link analysis counterterrorism and security. hasan, m. a., & zaki, m. j. ( ). a survey of link prediction in social networks. in charu c. aggarwal (ed.), social network data analytics (pp. – ). us: springer. homans, g. c. ( ). the human group. london: routledge and kegan. zan huan, ( ). link prediction based on graph topology: the predictive value of the generalized clustering coefficient. in proceedings of th acm sigkdd international conference on knowledge discovery and data mining. table proportion between the number of primary events (p) and secondary events (s). astro-ph cond-mat hep-lat hep-th p/s . . . . auc with a ¼ p=s . . . . best auc . . . . table performance of the methods (aucs). astro-ph cond-mat hep-lat hep-th event-based (b ¼ log) . . . . event-based (b ¼ ) . . . . time series (aa + lr) . . . . aa . . . . cn . . . . jc . . . . pa . . . . p.r.s. soares, r.b.c. prudêncio / expert systems with applications ( ) – author's personal copy huang, z., & lin, dennis k. j. ( ). the time-series link prediction problem with applications in communication surveillance. informs journal on computing, ( ), – . jiang, c., coenen, f., & zito, m. ( ). a survey of frequent subgraph mining algorithms. the knowledge engineering review, ( ), – . juszczyszyn, k., musial, k., budka, m. ( ). link prediction based on subgraph evolution in dynamic social networks, in privacy security risk and trust (passar) ieee third international conference on social computing (pp. – ). katz, l. ( ). a new status index derived from sociometric analysis. psychometrika, ( ), – . liben-nowell, d., kleinberg, j. ( ). the link prediction problem for social networks. in proceedings of the international conference on information and knowledge management (pp. – ). lichtenwalter, r. n., lussier, j. t., chawla, n. v. ( ). new perspectives and methods in link prediction. in proceedings of the th acm sigkdd international conference on knowledge discovery and data mining (pp. – ). lu, l., & zhou, t. ( ). link prediction in complex networks: a survey. physica a, ( ), – . murata, t., & moriyasu, s. ( ). link prediction based on structural properties of online social networks. new generation computing, ( ), – . newman, m. e. j. ( ). clustering and preferential attachment in growing networks. physical review e, . newman, m. e. j. ( ). the structure of scientific collaboration networks. in proceedings of the national academy of sciences of the united states of america (vol. , pp. – ). potgieter, a., april, k. a., cooke, r. j. e., & osunmakinde, i. o. ( ). temporality in link prediction: understanding social complexity. journal of emergence: complexity and organization, ( ), – . qiu, b., he, q., yen, j. ( ). evolution of node behavior in link prediction. in aaai (pp. – ). salton, g., & mcgill, m. j. ( ). introduction to modern information retrieval. mcgraw-hill inc. soares, p. r. s., prudêncio, r. b. c. ( ). time series based link prediction. in proceedings of the international joint conference on neural networks (pp. – ). tylenda, t., angelova, r., bedathur, s. ( ). towards time-aware link prediction in evolving social networks. in proceedings of the third workshop on social network mining and analysis, sna-kdd ’ (pp. – ). wang, c., satuluri, v., parthasarathy, s. ( ). local probabilistic models for link prediction. in proceedings of the seventh ieee international conference on data mining (pp. – ). wasserman, s., & faust, k. ( ). social network analysis: methods and applications (structural analysis in the social sciences). cambridge university press. xiang, e. w., ( ). a survey on link prediction models for social network data. phd thesis, phd qualifying exam, department of computer science and engineering, the hong kong university of science and technology. p.r.s. soares, r.b.c. prudêncio / expert systems with applications ( ) – this article appeared in a journal published by elsevier. the attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. in most cases authors are permitted to post their version of the article (e.g. in word or tex form) to their personal website or institutional repository. authors requiring further information regarding elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright author's personal copy landmark detection from mobile life log using a modular bayesian network model keum-sung hwang, sung-bae cho * department of computer science, yonsei university, seongsanro, sudaemoon-ku, seoul - , south korea a r t i c l e i n f o keywords: modularized probabilistic reasoning mobile application a b s t r a c t mobile devices can now handle a great deal of information thanks to the convergence of diverse function- alities. mobile environments have already shown great potential in terms of providing customized ser- vices to users because they can record meaningful and private information continually for long periods of time. until now, most of this information has been generally ignored because of the limitations of mobile devices in terms of power, memory capacity and speed. in this paper, we propose a novel method that efficiently infers landmarks for users to overcome these problems. this method uses an effective probabilistic bayesian network model for analyzing various kinds of log data in mobile environments, which were modularized in this paper to decrease complexity. we also present a cooperative inference method, and the proposed methods were evaluated with mobile log data generated and collected in the real world. � elsevier ltd. all rights reserved. . introduction mobile environments have very different characteristics from desktop computer environments. first of all, mobile devices can col- lect and manage various kinds of user information, for example, by logging a user’s calls, sms (short message service), photography, music-playing and gps (global positioning system) information. also, mobile devices can be customized to fit any given user’s pref- erences. furthermore, mobile devices can collect everyday informa- tion effectively. such features allow for the possibility of diverse and convenient services, and have attracted the attention of researchers and developers. recent research conducted by nokia is a good example (nokia lifeblog). especially, the context-aware technique that has recently been widely studied is more applicable to mobile environments, so many intelligent services such as intelligent call- ing services (schmidt, takaluoma, & mntyjrvi, ), messaging ser- vices (lo, thiemjarus, & yang, ), analysis, collection and management of mobile logs (devaul, sung, gips, & pentland, ; gemmell, williams, wood, lueder, & bell, ; korpipaa, mantyjarvi, kela, keranen, & malm, ; krause, smailagic, & siewiorek, ; raento, oulasvirta, petit, & toivonen, ; siewiorek et al., ; zheng & ni, ) have been actively investigated. however, mobile devices present some limitations. they con- tain relatively insufficient memory capacity, lower cpu power (data-processing speed), smaller screen sizes, awkward input interfaces, and limited battery lives when compared to desktop pcs. in addition, they have to operate in the changeable real world, which means that they require more active and effective adapta- tion functions (dourish, ). in this paper, we propose a novel method of analyzing mobile log data effectively and extracting semantic information and mem- ory landmarks, which can be used as special ways of helping recall specific functions (horvitz, dumais, & koch, ). the proposed method adopts a bayesian probabilistic model to efficiently man- age various uncertainties that can occur when working with mo- bile environments, including real-world irregularities, like varying levels of attention and emotions, inaccuracy of sensors, and uncertain causal factors. the proposed model uses a coopera- tive reasoning method with a modular bayesian network (bn) in order to work competently in mobile environments. we also dis- cuss how to learn and update the bayesian inference model by using the proposed bn learning method with training data. the proposed method was applied to several experiments using both synthetic data and real mobile log data collected with a smart- phone for a month in the real world. in hwang and cho ( ), we proposed a method for identifying landmarks on mobile de- vices, which was a modularized bn model designed by human manually. there have already been various attempts to analyze log data and to support expanded services by using the probabilistic ap- proach. li and ji used a probabilistic model for active affective state detection of user (li & ji, ). they utilized a dynamic bayesian network and the utility theory to reason the ‘fatigue,’ ‘nervous,’ and ‘confused’ states. they showed that the probabilistic approach was good at management of uncertain information like affection. ji et al. used also a dynamic bayesian network method for real- time monitoring human fatigue (ji, lan, & looney, ). they - /$ - see front matter � elsevier ltd. all rights reserved. doi: . /j.eswa. . . * corresponding author. tel.: + ; fax: + . e-mail addresses: yellowg@cs.yonsei.ac.kr (k.-s. hwang), sbcho@cs.yonsei.ac.kr (s.-b. cho). expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a author's personal copy concerned and combined many conditions such as light, heat, humidity, time, napping, anxiety, temperature, weather, and work type. zhang and ji proposed an active and dynamic information fu- sion method for multi-sensor systems with dynamic bayesian net- works (zhang & ji, ). this showed the usefulness of bayesian approach for information fusion. these works showed the bayesian probabilistic approach was good tool for handling, reasoning, and combining uncertain information. krause et al. clustered the sensor and log data collected on mo- bile devices, discovered a context classifier that reflected a given user’s preferences, and estimated the user’s situation in order to provide smart services (krause et al., ). the context classifier was constructed using the bn model, which was based on a general learning method for a small domain of classification subjects. hor- vitz et al. proposed a method that detected and estimated land- marks by discovering a given human’s cognitive activity model from pc log data based on the bayesian approach (horvitz et al., ). their approach showed a good performance for recognizing and learning everyday-pc-life of humans. however, these methods were not suitable for mobile devices that were limited in terms of capacity and power. for larger do- mains, the general bn and bn learning method require highly com- plex computation. this is a crucial problem when it comes to modeling everyday life situations with mobile devices. to over- come these problems, a more appropriate approach was required to reduce the complexity levels. marengoni, hanson, zilberstein, and riseman ( ) tried to reduce the complexity levels of the bn model by dividing it into several multi-level modules and using procedural reasoning of the connected bns (just like chain infer- ence). however, this method required procedural and classified properties of the target functions. tu, allanach, singh, pattipati, and willett ( ) proposed a hy- brid bn model that allowed hierarchical hybridization of bns and hmms. however, it supported only links from lower level hmms to higher level bns without consideration of links between bns of the same level. hwang, kim, and zhang ( ) proposed the hierarchical probabilistic graphical modeling method, which con- structed a hierarchical and distributed bn structure using gener- ated hidden nodes and the links between them. however, this method is improper in mobile and private service environments since it leads to an increase of the number of nodes and does not keep the intuitive causal structure. . landmark inference from mobile log data the overall process of landmark extraction from the mobile log data used in this paper is shown in fig. . various mobile log data is preprocessed in advance, and then the landmark-reasoning module detects the landmarks. the preprocessing module is oper- ated by the techniques of pattern recognition and simple rule rea- soning. the bn reasoning module performs probabilistic inference. the update module learns the bayesian network mod- els and adapts to the user and environment by using the accumu- lated data. bns refer to models that can express large probability distribu- tions with relatively small costs to statistical mechanics. they have the structure of a directed acyclic graph (dag) that represents the link (arc) relations of the node, and has conditional probability ta- bles (cpts) that are constrained by the dag structure (korb & nicholson, ). fig. shows an example bn that was designed by human and used for the application of this paper. it shows a dag structure, node name, state name and inferred probabilities. . . collection and preprocessing table shows log information collected on a mobile device and on the internet. the gps log presents the places that the user vis- ited, and the call and sms logs show the user’s calling patterns. the mp (a music file format) player log offers an idea of the user’s emotions and the photograph log shows when the user wanted to memorize something. since logs have temporal properties, we considered their time spans, frequencies (per hour, daily, and weekly), and start/end times as well as their impact factors that reflect the density of gi- ven events. for example, the impact factor of time t, ipt can be cal- culated as eq. ( ), where x represents the event and a(x) and b(x) represent the increment/decrement constants and monitoring time-span for event x. we set the value a(x) as for each event x, and the values for b(x) are set manually as follows; b(gps) = h, b(call) = h, b(sms) = min,b(pic. view) = min, b(photograph- ing) = min, and b(mp playing) = min, iptðxÞ¼ ipt� ðxÞþ aðxÞ; if event x is occurred ipt� ðxÞ� aðxÞ; if event x is not occurred in bðxÞ after prior event x >< >: ð Þ the coordinates from the gps log are used to get place names. in this paper, we divided the domain area into a lattice and then la- beled each region. the user profiles and pims (personal informa- tion management system) datasets were used to find the user’s social position (student, worker), gender, the position of their home, and the names and phone numbers of their friends and relatives. weather user profile gps log preprocessing (statistics, analysis) p h o n e call / sms log pims landmark decision photo log device log log context db w eb landmark reasoning music playing log landmark db update module bnn bn … bn modular bns fig. . the process of the landmark extraction from mobile log data. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy . . cooperative reasoning of modular bayesian networks there are two general differences between the proposed bns and conventional bns. firstly, we modularized the bayesian inference models according to their separated sub-domains (fig. ). the re- quired computing power for dealing with bn models is proportional to the number of nodes and arcs. especially, since the computational complexity of bayesian inference is approximately proportional to o(cn), where c is the number of states and n is the number of causal (parent) nodes, the modularized bn is more efficient. secondly, to consider the co-causality of the modularized bn, the proposed method shows -pass inference stages (fig. ). a vir- tual linking technique is utilized to reflect the co-causal evidence more correctly. the technique is performed to add the virtual nodes and regulate their conditional probability values (cpvs) to apply the probability of the evidence. the process of the virtual linking technique is shown in figs. and . in fig. , the virtual linking technique replaces given probabilistic evidence to the prior probability or conditional probability of virtual node value to keep the connectivity of bns. for example, when the structure of bn is {a ? b ? c}, bn is its modularized version {a ? b, b ? c}, and a and c are evidence and result nodes, we can calculate the belief value of node c using the virtual link assumption and chain rule such as eqs. ( )–( ) yes % no % coffee-shop yes % no % previor:eating yes % no % previor:stroll yes % no % lunch-time yes % no % breakfast-time yes % no % dinner-time yes % no % meal-time yes % no % teaing yes % no % restaurant yes % no % eating(western) yes % no % korean-restaurant yes % no % eating(eastern) yes % no % fastfood-shop yes % no % eating(fastfood) yes % no % eating yes % no % ordinary-place yes % no % high-class-restaurant yes % no % eating-out whereprevior landmarkswhen/work fig. . the landmark inference bn ‘activity in restaurant’ designed and used in this paper. this is of bns designed. the node of gray box is landmark node. this shows the case that the ‘lunch-time’ and ‘restaurant’ evidence are set. table the log information that was collected on a mobile device. log information gps latitude, longitude, velocity, direction, date, time call caller’s phone number, call/receive/absence log, time span, start/end time sms sender’s phone number, call/receive/absence log, time span, start/end time photographing photo file name, taking time weather weather, visibility range (km), cloud degree (%), temperature (�c), discomfort index, effective temperature (�c), rainfall (mm), snowfall (cm), humidity (%), wind direction, wind velocity (m/s), barometer (hpa) mp player title, time span, start/end time charging charging status, time span, start/end time log context landmark place- activity rea- soning bn emotion/ condi- tion rea- soning bn circum- stance situation bn event rea- soning bn virtual evidence st stage landmark place- activity rea- soning bn emotion/ condi- tion rea- soning bn circum- stance situation bn event rea- soning bn nd stage log context fig. . the -pass inference process for the cooperation of modular bayesian networks. it uses the result from inference of bayesian network as the evidences at the nd inference stage. the parenthesis indicates the number of bns contained. there are four kinds of bns, and totally bns are used. the bns are as follows: place-activity bns = {houses, religion, shopping, photographs, hospitals, nature, meetings, workplaces, sports, movements, food, calls, music, schools, traffic, amusement, busy, watching, resting}, emotion/condition bns = {joy, hunger, hot temperatures, cold temperatures, nonsense, surprise, tiredness, drunk, anger, worry, gloom, sickness, boredom}, circumstance situation bns = {space, climate, time, device, group}, event bns = {anniversaries, other events}. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy where the proposed virtual link assumption is p(b) = bel(b), and bel(b) means belief value of the node b (korb & nicholson, ), bn ’s belðcÞ¼ pða; b; cÞ¼ pðc j bÞpðb j aÞpðaÞ ð Þ bn ’s belðbÞ¼ pða; bÞ¼ pðb j aÞpðaÞ ð Þ bn ’s belðcÞ¼ pðb; cÞ¼ pðc j bÞpðbÞ ð Þ ¼ pðc j bÞbelðbÞ since pðbÞ ¼ belðbÞ by virtual link assumption ¼ pðc j bÞpðb j aÞpðaÞ ) bn ’s belðcÞ¼ bn ’s belðcÞ ð Þ from the cooperative reasoning process we can get the landmarks and their probability distribution. fig. shows a pseudo code for the landmark extraction process. . . landmark decision the final landmark decision is obtained by the belief probabil- ity, thresholds and weights of each landmark node after reasoning the modular bns. the threshold is used to tune the landmark extraction model. generally, the training data is not uniform en- ough to obtain an almost-perfect model, so we need to tune the model. eq. ( ) shows the computations to obtain the selected land- mark set lms, where bi represents the belief probability of node xi, wi is the weight of node xi, and lm is the set of landmark nodes. the weights are used to apply the preferences or life-patterns of the given user, so if the landmark is a preferred by an user, the weight has higher value than others and the landmark can be se- lected more easily, lms ¼fxijbi � wi > thi ^ xi lm ^ wi g ð Þ . . complexity analysis to compare the complexities of the proposed modular bns and the ordinary bn, we combined the designed bns into a bn as shown in fig. . a total of bns were designed with nodes, arcs and cpvs, as summarized in table . on average, . nodes and . cpvs are used at the same time for the bn reasoning pro- cess since they are modularized and the computations are distrib- uted. however, even though the combined bn for them has nodes, which is decreased from after removing the duplicated nodes on the combining process, the number of parents and cpvs are increased. this means that it has much larger complexity. eq. ( ) shows the time complexity of the exact inference of the bn using the lauritzen speigelhalter (ls) algorithm (lauritzen & spiegelhalter, ), which is the most popular exact inference algorithm and a junction tree-based algorithm, where n represents the number of nodes, k represents the maximum number of par- ents for each node, r denotes the number of values for each node, and w represents the maximum clique that each parameter used in the ls algorithm (namasivayam & prasanna, ), cmpxinf ¼ oðk nk þ wn þðwrw þ rwÞnÞ ð Þ we have replaced the maximum clique size w with k, since the clique size is proportional to the parents’ size, and the number of values r is about in this paper, so eq. ( ) was reduced as shown by eq. ( ). cmpxinf ffi oðk nk þ kn þ kðk þ ÞnÞ ð Þ the ideal modularization technique roughly divides the number of nodes by the number of modules d but it has to compute d times for d modules, so the exact inference complexity of the modular bns is as shown in eq. ( ) where the new number of nodes is (n/ d). as shown in table , the proposed modularization technique roughly divides the number of nodes n by . and the number of parents by . , so the exact inference complexity is decreased significantly, cmpx inf ffi o k dk� nk þ k d n þ kðk þ Þn ! ð Þ . learning modular bayesian networks from data in this section we introduce a method to learn the proposed modular bns automatically from the training data set. fig. . virtually linked bayesian networks. dotted line denotes some nodes linked by virtual linking technique. xi'' virtual link : for root nodes; replace the prior probability as p xi xi' virtual node added xi=yes virtual link : for general nodes; set the evidence xi, and then replace the conditional probability of virtual node as p p (xi''=yes) p x xi=yesx i=no p (xi'=yes|xi) p k' probabilistic evidence: xi=yes, p (xi=yes)=p p cpt cpt p fig. . the virtual linking method using virtual evidence enabling to use proba- bilistic evidences. fig. . the pseudo codes for landmark extraction processes. l is log data set. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy . . learning bayesian network from data the bn model g can be defined as (bs, h), which means a net- work structure bs and a probability parameter set h. h = {bu, bp} is composed of the conditional probability table bu and the prior probability distribution bp. in general, learning bayesian networks from data consists of two parts. the first is to learn the structure bs and the other part is to learn the parameter set. learning the struc- ture can be solved by searching for the network structure which best matches the given data. the fitness can be measured by some scoring metrics. two popular metrics are the minimum description length (mdl) score (lam, ) and the bayesian dirichlet (bd) score (heckermann, geiger, & chickering, ). the bd score for the structure bs given the training data set d is defined as eq. ( ). to find the best structure with score metric, greedy search algorithms can be used. fig. shows a general process for the score-metric based structure learning, bdðbs; dÞ¼ pðbs; dÞ ¼ pðbsÞpðdjbsÞ¼ pðbsÞ z pðdjh; bsÞpðhjbsÞdh ð Þ parameter learning can be cast as either estimating the most likely parameter values (maximum likelihood learning) or updating the posterior probability distributions of the parameters given data (bayesian learning). the parameter h is calculated from the training data set d by using eq. ( ). h� ¼ arg max h pðdjhÞpðhÞ ð Þ p(h) means prior probability. the general discovery process is shown by eq. ( ) where zt = {z , z , . . ., zt} represents a set of t sta- tus variables, and yt is a set of t observation result variables, pðzt; y t; hÞ¼ pðy tjzt; buÞpðztjbpÞ ð Þ the conditional probability table bu is calculated from the rela- tional frequency between the observation values and the status variables, meaning the probabilistic histogram of the training data. . . discovering the modular bayesian networks the proposed learning process of the modular bns is shown in fig. . this method includes a parameter setting process (for node fig. . the complex bn structure combined from the designed bns. this has arcs and nodes. the graph is drawn by genie . (graphical bn tool). table the modular bns and the monolithic bn of them. monobn – the monolithic bn corresponding to the modular bns, nn – no. of nodes, nnr – no. of root nodes, nni – no. of intermediate nodes, nnl – no. of leaf nodes, nl – no. of arcs, npavg – avg. no. of parents, ns – no. of states, nsavg – avg. no. of states, ncpv – no. of cpvs, cptmax – maximum size of cpt. nn nnr nni nnl nl npavg ns nsavg ncpv cptmax bns . . monobn . . . initialize structure score all possible single changes any changes better? return saved structure perform the best change yes no fig. . a general process for score metric based structure learning. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy set x, the topological order set of nodes o, a level set of nodes v and the maximum size of parents p that each node had), a structure learning (for bs) and a parameter learning (for h). the node set of bn (x) is composed of the union of the collected log context set (l) and the target landmark set (lm). the values of the landmark set are defined by an expert user. the topological order set of nodes (o), the level set of nodes (v), and the maximum size of the parents (p) have been prepared for structure learning (the proposed meth- od; domain-constrained k learning). to determine the topological order (o), we ranks each variable based on its total influence on the other variables with mutual information. the domain-constrained k learning method is proposed to constrain the parent domain and the hierarchical levels of the nodes. the learning method and the mutual information computation are presented in the next section. the network structures learned are divided into several mod- ules by a modularization process as shown in fig. , where g is the network structure and d denotes the number of modules. pro- cess defines the nodes of the module-bns based on the module domain and the log context set. process defines the arcs of the module bns based on the network structure (g). process defines the virtual nodes based on the network structure (g). the virtual node usage used for virtual linking between the bns is described in section . . . . domain-constrained k algorithm the domain-constrained k method proposed in this paper is based on our prior work (hwang & cho, ) and aims to learn the structure of the bn hierarchically with a constrained parent do- main of nodes. the algorithm limits the hierarchical level at which a node can be positioned and reduces the search domain for the parent of the nodes. it can control the general direction of causality and decrease the searching complexity. in this paper, we use the algorithm to exclude the arcs between the evidence nodes (but we permits arcs between the landmarks) and maintain the direc- tion of the arcs in order to use the virtual link technique. table shows the level and domain settings used in this paper. fig. shows the proposed domain-constrained k algorithm, based on the k algorithm as proposed by cooper and herskovits ( ), which is the most popular bn algorithm and the basis of many advanced learning algorithms. this algorithm adopts a score metric known as the k metric, which calculates scores based on the difference between the bn graph and the training data distri- bution. the search process of the k algorithm is greedy and heu- ristic, and the k metric is calculated using eq. ( ), pðg; dÞ¼ pðgÞ yn i¼ yqi j¼ ðri � Þ! ðnij þ ri � Þ! yri k¼ nijk! ð Þ where ri represents the number of status variables, xi, qi means the number of combinations of variables included in pi, nijk denotes the count in the dataset d, pi is the jth combination, and xi is the kth preprocessed log l={l , l , …} target landmark lm={lm , lm , …} defining the topological order o={o , o , …} sorted by mi=Σj mij module domain m ={x , x , …} md={x , x , …} modularization of bn parameter learning of bn learned modular bns structure learning of bn g = structure_learn (x, o, v, p) defining the node set of bn x={x , x , …}=l lm calculating the link strength between nodes m(xi, xj): mutual information defining the level of nodes v={v , v , …| vi= ,if xi l, vi= , if xi lm, i= ,…} fig. . the learning procedure of modular bayesian network. fig. . the modularization process of bn. the right side shows an example, in which the gray nodes denote evidence nodes and the white nodes mean landmark nodes or their virtual nodes. table the level and domain definition. level node domain parent domain lm = {lm , lm , . . .} d = u l = {l , l , . . .} d = l [ lm lm is landmark set, l is log context set. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy value. the nij value is computed by eq. ( ) and the function g( ) is calculated by eq. ( ), nij ¼ xri k¼ nijk! ð Þ gðxi; paðxiÞÞ¼ yqi j¼ ðri � Þ! ðnij þ ri � Þ! yri k¼ nijk! ð Þ the k algorithm uses a topological order to maintain the graph as the dag by maintaining that the prior node cannot be the child of the posterior node without any other dag checking rules. how- ever, we have to optimize the topological order since a different topological order will lead to a different bn structure. in this paper, we compute the influence score of all the nodes by using mutual information (su & zhang, ), and sort out the topological order with the score. eqs. ( ) and ( ) show the influence score and the mutual information calculation, mi ¼ xmðxi ;xjÞ j ð Þ mðx; yÞ¼ x x;y pðx; yÞ log pðx; yÞ pðxÞpðyÞ ð Þ . . threshold and weight the weight value presents the importance of landmarks and can be set by humans in terms of their preferences. in this paper, we set all the weight values as . to observe the general performance of landmark extraction. the performance of a classifier depends on its threshold since it determines the result status of the node. especially in bn classifi- ers, it is changeable because the belief probability distribution is based on the number and distribution of training data values, so we have defined the threshold as the value that shows the best hit rate for the training data. fig. shows an example of the clas- sification results of the node ’moving.’ in the results, the hit rate fig. . the domain-constrained bn structure learning algorithm. dk denotes the parent domain of kth level, and level(i) means level of xi and maxparentnum is a limitation of the number of parents. fig. . the classification performance of the node ‘moving’ with threshold transitions (from a modular bn trained with the parameter p = ). the rest threshold (more than . ) is omitted since there is no change. the value . shows the best performance. table the category, id and names of nodes. category ids node names landmark nodes ( ) n , n , . . ., n moving, using vehicle, school-club activity, eat out, lecture, sleeping, surprising, fret, joy, throb, tired, meet friend, cold, meal (korean), meet family, busy, hungry, drinking (alcohol), take a walk, troublesome, overflowing joy, bored, ride a vehicle, weight-training, home activity, with leisure, hair-cut, study, yearning, school activity, traffic jam, run, walk, go to school, meal (western), eat (tee), ready to go out, date with my date, sad, test in school, sleepy, meet kin, employment counsel, extracurricular lecture, late for school, concentration, eat (chinese), supper log context nodes ( ) n , n , n , . . ., n , n , n sms sender = unknown, take a picture, listening music, calling with kin or friend, receive a call, listening music for a long time, charging, good scenery place, gps is operating, average velocity is less then km/h, speed is from km/h to km/ h, amusement grounds, velocity is from km/h to km/h, shopping place, lecture building, velocity is less than km/ h, town, front of school, department store, rest place, bus station, soccer play ground, park, traffic place, a post office, street, public institution, home, o’clock, o’clock, o’clock, . . ., o’clock, night, dawn, morning, am, pm, evening, daytime, sunshine-time, week, calling, ignore of call, velocity is faster then km, out of town, merrymaking place, subway station, tourist resort, terminal of exp. bus, express bus, student building, place for meal, church, place for special meal, school, coffee shop, out of town, front of ys univ., main load of ys univ., front of lecture building, main entrance of school, wedding place, museum, street basket-ball court, coffee shop, japanese restaurant, playground, library, lecture building, theater, middle school, weather = fine, thursday, spring, frequent call, many call, electronic and electric products center, commercial center, subway station, restaurant, many photo, high-school, dong-dae-moon place, weather = fine, hospital, dining room, korean restaurant, house of kin, house of friend, chinese restaurant the parenthesis denotes the number of nodes. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy was the best when the threshold was . , and we adopted the va- lue for the threshold of the nodes. this procedure was conducted automatically during bn training stage. . experiments in this section, we analyze the proposed method with experi- ments with generated and collected data. we evaluate the reason- ing performance of the landmark extraction model designed and learned. then, we observe the experimental results. finally, we compare the auto-trained modular bns and their monolithic versions. . . experimental data the log data used in this paper include gps log, call log, sms log, picture log, music-playing log, device charging log, and weather log obtained from a website. the total number of designed bns is ( activity reasoning bns for kinds of places, emotion/con- dition reasoning bns for the users, circumstances or situation bns, and event reasoning bns). we collected data from three college students (women) with smartphones for a month. these users performed subtasks (such as writing activity diaries, shopping, walking and calling) to make the data more useful. the experimental data was segmented into units of ten min- utes, so there were entries. redundant data were excluded resulting in examples remained. we defined landmarks and used life log contexts (as evidence) as shown in table . to learn the modular bns, we divided the landmarks into four modules based on four categories (emotion and status, everyday life, events, and school life) as shown in table . . . a case study we tested the proposed landmark-reasoning model with the following scenario in order to confirm performance. the left side of fig. shows the scenario used. the bn set strongly related to table module domain definition of landmark nodes. domain name number landmarks emotion and status bored, busy, cold, concentration, fret, hungry, joy, overflowing joy, sad, sleepy, surprising, throb, tired, troublesome, with leisure, yearning everyday life eat (chinese), eat (western), eat (korean), home activity, meet family, moving, ready to go out, ride a vehicle, run, sleeping, supper, using vehicle, walk event date with my date, drinking(alcohol), eat (tee), eat out, hair-cut, meet friend, meet kin, take a walk, traffic jam, weight-training school life employment counsel, extracurricular lecture, go to school, late for school, lecture, school activity, school-club activity, study, test in school total home park school restaurant downtown coffee shop home hr hr hr hr hr hr hr hr hr hr hr hr hr hr hr hr moving moving activity lecture building student hall lecture building scenario today is school day. morning lecture time. simple lunch. walking for a park. photo with spring flowers. met friends. going to the restaurant wanted going a coffee shop. chatting. funny & joyful day. target landmark going-out- preparation walking-for joyful-photo eating-out teaing overflowing- joy lecture lecture eating sleeping sleeping (a) (b) fig. . (a) a scenario of an everyday life with mobile device of an undergraduate student for experiments. (b) the observation of the probability values of target landmarks. the denoted time is from o’clock to o’clock (equal to o’clock of the next day). abbreviations used for landmarks: a – overflowing-joy, b – photo (scenery), c – joyful-photo, d – walking-for, e – tea, f – eating-out, g – eating (western style), h – eating, i – going-out-preparation, j – shower, k – sleeping. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy the scenario is {food, photo, movement, nature, joy, home}. the probabilities were calculated when the related evidence was given. the log context data used as evidence includes artificial log con- texts: going out (in h), lecture-building, daytime, restaurant, pho- tography, eating, entertainment, before breakfast, condition, morning, going-out (before), moving, daylight-hours, sleeping (in h), nature, device-in-use, dinner(before), good-weather, joyful day, home, charging, coffee-shop, ordinary-place, on campus, stu- dent-hall, gps-running. after generating the log contexts during the day, we tested them. the right side of fig. shows the inference results. we can see the high probability values of the related landmarks at the corresponding times. for example, there are ‘going-out-prepa- ration’ and ‘shower’ landmarks at – o’clock, ‘eating’ at – and – o’clock, ‘walking’ at – and – o’clock, ‘photography at – o’clock, and ‘eating’ landmarks at – o’clock. . . performance evaluation of the designed bns to evaluate performance of the designed bns, we operated our landmark extraction model. table shows the results of landmark extraction using data from eleven days collected from human . the data of the underlined date was also used in the learning experiment. in the experiment, we set all landmark thresholds as % to avoid worthless landmarks, since a human designed the bn models by focusing on true positive relations naturally and without regarding negative relations between the logs and the landmarks. since the log data were uncertain and the activity diary did not cover all possible life contexts, we evaluated the results (rhit’) using a wide viewpoint (permitting partial hits), which means that the landmarks could be reasonably detected. we decided the table the real life log data collected in smartphone. we have excluded the data of unstable gps log. ncon – the number of input contexts (evidences), nlm – the number of extracted landmark candidates, n lm – the number of selected landmarks, n hit – the number of hitting landmarks, n lm – the number of partial hitting landmarks, nerr – the number of landmarks, rhit – the rate of nhit , r hit – the rate of n hit , rerr – the error rate. date ncon nlm n lm nhit n hit nerr rhit r hit rerr - . . . - . . . - . . . - . . . - . . . - . . . - . . . - . . . - . . . - . . . - . . . total . . . table the matching results of extracted landmarks. they are evaluated based on the activity diaries and gps trajectories of users. date hit landmark partial hit landmark missing (omission) - meal (korean), meeting, shopping studying, lecture, lecture time, busy, unpleasant sms, viewing, eat out, traffic jam, singing, dancing at club – - study, joyful photograph, meeting, shopping lecture time, busy, unpleasant sms, photograph (food), photograph (product), photograph (scene), viewing, meal (korean), eat out, singing, dancing at club – - busy, joyful call, meeting lecture time, unpleasant sms, viewing, studying shopping - lecture time, busy, joyful call, meal (korean), eat out, meeting, take a walk unpleasant sms, photograph (food), photograph (product), photograph (scene), viewing, disappointment, shopping, singing, dancing at club, studying, studying till late – - viewing computer working, washing face, busy, unpleasant sms, traffic jam lecture time - meeting, unpleasant sms, lecture time, busy, meal (korean) take a walk, viewing, traffic jam eat out, shopping, singing, dancing at club - lecture time, meal (korean), eat out unpleasant sms, busy, traffic jam, meeting, shopping, singing, dancing at club viewing, take tea - study, meeting, lecture time, viewing, meeting busy, traffic jam shopping - lecture time, mean (korean), eat out, shopping cleaning, cooking, dishwashing, busy, viewing, meeting, singing, dancing at club, take a walk – - studying, take a walk, lecture time busy, disappointment, meeting shopping, viewing - study till late, lecture time, meal (korean), shopping busy, meeting, singing, dancing at club viewing, eat out . . . . . . . . . ea t o ut joy mo vin g ha ir- cu t ea tin g( ko re an ) lec tur e stu dytire d tak e a w alk tes t in sc ho ol me et frie nd us ing ve hic le em plo ym en t c ou ns el sc ho ol ac tiv ity ea t ( we ste rn ) wa lk da te wi th my da te sc ho ol- clu b a cti vit y dr ink ing al co ho l) ru n fre t tra ffic ja m ea t ( tee ) sa d sle ep y me et kin bo re d ye ar nin g ex tra cu rri cu lar le ctu re tro ub les om e thr ob co nc en tra tio n wi th lei su re sle ep ing su rp ris ing co ld me et fam ily we igh t-t ra ini ng ho me ac tiv ity go to sc ho ol re ad y t o g o o ut bu sy ov er flo wi ng jo y ea t ( ch ine se ) rid e a ve hic le hu ng ry lat e f or sc ho ol su pp er landmark nodes o pt im iz ed th re sh ol d yes no fig. . the optimized threshold set for node. the thresholds are sorted by the value. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy partial hits by majority of persons. as a result, the rate of the complete hit landmarks (rhit) was mostly low ( . % in total) but the rate of the partial hit landmarks (rhit’) was as high as . %. especially, when the activities were plenty, the results were better since the contexts were sufficient. table shows the specific results of landmark reasoning. for example, the user’s real activity on - was practicing music until late at night, but the land- mark’studying till late’ was detected. we classified landmarks like this as partial hit landmarks. . . performance evaluation of the trained bns in this section, we describe the test result of the landmark extraction model using days of mobile life log data obtained fig. . the monolithic bn trained with parameter p = . the brighter nodes are landmark nodes and the others are evidence nodes. out of town front of ys univ. main load of ys u... house of friend home o'clock express bus weddin g place subway station take a picture front of school am tourist resort school house of kin joy lecture buildin g japane se resta... playgro und v_meet friend v_eat out v_sho ol-club activity lecture buildin g v_lectur e daytime surprisi ng fret o'clock throb street o'clock out of town middle school electon ic and electr... tired o'clock cold v_ea ting (kore an) v_studyv_sch ool activity comme rcial center high-sc hool busy hungry v_drink ing(alc ohol) bored v_rid e a vehic le troubles ome v_weig ht-traini ng v_hom e activity overflo wing joy v_eat (wester n) with leisure yearningsleepy o'clock sad concent ration out of town f ront of y s univ . v_mov ing hom e f ront of scho ol tourist resort scho ol house of kin v_joy lecture bui ldin g town japan ese restau rant play gr ound meet f riend mai n entra nce of school eat out o'clock sunshi ne-t ime many photo gra phi ng cof f ee sho p v_studyv_busy rest place dri nking( al coho l) o'clock restau rant take a walk park wei ght-tr ai nin g o'clock hair- cut v_go to scho ol traf f ic jam cof f ee sho p eat (tee) date with my date meet kin gps is operating out of town f ront of y s univ . main load of y s univ . house of f riend mov ing traf fic place home o'clock pm express bus wedding place museum subway station using v ehicle tourist resort v_joy lecture building town japanese restaurant v_eat out weather=f i ne night day timesleeping o'clock v_throb v_tired o'clock o'clock eating (korean) o'clock speed is f aster then ... terminal of exp. bus meet family speed is less than km/h public institution walk o'clock restaurant v_bored park ride a v ehicle home activity eat (western) o'clock week run v_sleepy o'clock ready to go out chinese restaurant eat (chinese) o'clock supper front of ys univ. mai n load of ys univ. house of friend pm wedding place am school v_joy lect ure building v_meet friend v_eat out o'clock sunshine- time shool-clu b acti vity lect ure building weather =f ine many photogra phing lect ure street middle school evening coff ee shop library frequent call weather =f ine st udy v_wal k theater many call school act ivit y go to school test in school calling st udent building place for meal empl oym ent counsel extrac urri cular lect ure v_conc e ntrat ion late for school emotion & status bn everyday life bn event bn school life bn fig. . the bn modules trained using the proposed learning method with parameter p = . the dark nodes: virtual nodes, the lighter nodes are landmark nodes, and the others (medium brightness) are evidence nodes. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy by the proposed modular bn learning method. we set the maxpar- entnum parameters (p) as and in the experiment. the automat- ically-discovered threshold values for the landmarks of the modular bns are shown in fig. , where the landmarks ‘hungry’, ‘late for school’, and ‘supper’ cannot be detected when the thresh- old is . this means the training data do not contain the related evidences. the other landmarks show various threshold values. figs. and show the monolithic bn and modular bns learned with (parameter p = ). the average number of nodes, par- ents, and conditional probability values and the level of complexity are shown in table . the complexities are calculated by eq. ( ) and we can observe the decrement of the complexity of the mod- ular bns. table shows the results of the landmark reasoning evaluation. because the number of training data was small, we used the leave- one-out validation method. we compared the monolithic bn and modular bns with the parameters p = and p = . the computation of the precision rate is (tp/(tp + fp)), and the hit rate is ((tp + tn)/ (tp + tn + fp + fn)). as shown by the results, the performance of the modular bns is similar to that of the monolithic bn. the measured values are same when the parent size parameter p is . this means the proposed method is valuable since it is used to reduce the bn model and in- crease efficiency. the results with p = is better than with p = . we think that the bigger p increase complexity by allowing more parents and cause lack of training. how many parameters of bn are not trained from data can be measured by the number of cpvs that is . since an initial untrained cpv is . . the number of the trained mono bn with p = is and with p = is , so it is reasonable that the bn with p = is better than with p = . . . comparison of the execution speed in order to compare the computation complexities between the ordinary monolithic bn model and the proposed model, we have tested their execution time on mobile device simulation. the sub- jects are designed bns, their monolithic bn, trained bn, and their monolithic bn. in this experiment, since we used -step infer- ence method for modular bn model, the reasoning process is con- ducted twice for each bn. we conducted times of inference for each bn model. the test environment for mobile device is as follows: � simulation tool: microsoft pocket pc sdk. � operating system: pocket pc . � memory: mb. table shows the experimental results. the execution time denotes cpu clock time-span during reasoning processes. the monolithic and combined bn took more time than modular bns on pocket pc environments, but some experiments with mono- lithic bn failed to run because of out-of-memory error. this means that the ordinary bn cannot handle the landmark detection task on the constrained environment like pocket pc where the domain is too large while the proposed method can manage the task well. to find the executable size of bn for monolithic bn (for the learned), we tested the bn with progressive deletion of nodes. fig. shows the experimental result. the learned monolithic bn could operate after deleting the rd node. it means that the rd node requires much computation complexity. actually, the rd node had parents and has conditional probability param- eters. it can be understood that some complex nodes cause much computation and make troubles. in the experiments, we can observe that the proposed model re- duced the computation complexity of bn and enabled to handle more complex probabilistic model on the constrained environment even when the original bn model cannot work. . concluding remarks in this paper, we proposed a modular landmark inference mod- el, which was more efficient and suitable to mobile environments. we introduced the modularized bn model for efficient operations in mobile environments, and proposed the -step inference meth- od by applying the virtual node concept, and then learned the modular bns automatically from the given training data. in exper- imental results with artificial and real mobile life log data, the in- tended landmarks were well-extracted and the proposed method was able to reduce the level of complexity. we also discussed how to define the learning parameters and thresholds however, in this paper, we did not sufficiently cover the temporal properties of human landmarks. in the future, we need to continue research using a dynamic bn model that manages temporal features table the comparison of the number of nodes, parents, and conditional probability values and the complexity of learned bns. the complexities are calculated based on eq. ( ). bn nodes # nodes #avg parents # parents #avg complexity mono bn . o ( . � ) modular bns . o ( . � ) # – the number of, #avg – the average number of. table the correctness comparison of extracted landmarks. same values are underlined. bn p tp tn fp fn precision hit rate mono bn , , . . modular bns , , . . mono bn , , . . modular bns , , . . tp – true positive, tn – true negative, fp – false positive, fn – false negative. table the comparison of the execution time on pocket pc emulator. bn environment number of bn execution time (clk) modular bns pocket pc combined bn pocket pc modular bns pocket pc monolithic bn pocket pc n/a – out of memory error fig. . the execution time of the learned monolithic bns with node deletion on pocket pc emulation. k.-s. hwang, s.-b. cho / expert systems with applications ( ) – author's personal copy well. also, experiments with sufficient real world data should be conducted for a longer period of time. acknowledgement this work was supported by mke, korea under itrc iita- - (c - - ) and ucn c -t - t. references cooper, g. f., & herskovits, e. ( ). a bayesian method for the induction of probabilistic networks from data. machine learning, , – . devaul, r., sung, m., gips, j., & pentland, a. ( ). mithril : applications and architecture. in proceedings of the th ieee international symposium on wearable computers (pp. – ). dourish, p. ( ). what we talk about when we talk about context. personal and ubiquitous computing, ( ), – . gemmell, j., williams, l., wood, k., lueder, r., & bell, g. ( ). pervasive capture and ensuing issues for a personal lifetime store. in proceedings of the st acm workshop on continuous archival and retrieval of personal experiences, october (pp. – ). genie . . decision systems laboratory, university of pittsburgh. . heckermann, d., geiger, d., & chickering, d. m. ( ). learning bayesian networks: the combination of knowledge and statistical data. machine learning, ( ), – . horvitz, e., dumais, s., & koch, p. ( ). learning predictive models of memory landmarks. in cogsci : th annual meeting of the cognitive science society (pp. – ). hwang, k.-s., & cho, s.-b. ( ). constrained learning method of bayesian network structure for efficient context classification. proceedings of korea information science society, ( ), – (in korean). hwang, k.-s., & cho, s.-b. ( ). modular bayesian networks for inferring landmarks on mobile daily life. in the th australian joint conference on artificial intelligence (pp. – ). hwang, k.-b., kim, b.-h., & zhang, b.-t. ( ). learning hierarchical bayesian networks for large-scale data analysis. international conference on neural information processing, , – . ji, q., lan, p., & looney, c. ( ). a probabilistic framework for modeling and real- time monitoring human fatigue. ieee transactions on systems, man, and cybernetics, part a, ( ), – . korb, k. b., & nicholson, a. e. ( ). bayesian artificial intelligence. chapman & hall/ crc. korpipaa, p., mantyjarvi, j., kela, j., keranen, h., & malm, e.-j. ( ). managing context information in mobile devices. ieee pervasive computing, ( ), – . krause, a., smailagic, a., & siewiorek, d. p. ( ). context-aware mobile computing: learning context-dependent personal preferences from a wearable sensor array. ieee transactions on mobile computing, ( ), – . lam, w. ( ). bayesian network refinement via machine learning approach. ieee transactions on pattern analysis and machine intelligence, ( ), – . lauritzen, s. l., & spiegelhalter, d. j. ( ). local computations with probabilities on graphical structures and their application to expert systems. journal of the royal statistical society, – . li, x., & ji, q. ( ). active affective state detection and user assistance with dynamic bayesian networks. ieee transactions on systems, man, and cybernetics, part a, ( ), – . lo, b. p. l., thiemjarus, s., & yang, g.-z. ( ). adaptive bayesian networks for video processing. international conference on image processing, ( ), – . marengoni, m., hanson, a., zilberstein, s., & riseman, e. ( ). decision making and uncertainty management in a d reconstruction system. ieee transactions on pattern analysis and machine intelligence, ( ), – . namasivayam, v. k., & prasanna, v. k. ( ). salable parallel implementation of exact inference in bayesian networks. international conference on parallel and distributed systems. nokia lifeblog. . raento, m., oulasvirta, a., petit, r., & toivonen, h. ( ). contextphone: a prototyping platform for context-aware mobile applications. ieee pervasive computing, ( ), – . schmidt, a., takaluoma, a., & mntyjrvi, j. ( ). context-aware telephony over wap. personal technologies, ( ), – . siewiorek, d. p., smailagic, a., furakawa, j., krause, a., moraveji, n., reiger, k., et al. ( ). sensay: a context-aware mobile phone. in proceedings of the th international symposium of wearable computers, october (pp. – ). su, j., & zhang, h. ( ). full bayesian network classifiers. in proceedings of the international conference on machine learning (pp. – ). tu, h., allanach, j., singh, s., pattipati, k. r., & willett, p. ( ). information integration via hierarchical and hybrid bayesian networks. ieee transactions on systems, man, and cybernetics, part a: systems and humans, ( ), – . zhang, y., & ji, q. ( ). active and dynamic information fusion for multisensor systems with dynamic bayesian networks. ieee transactions on systems, man, and cybernetics, part b, ( ), – . zheng, p., & ni, l. m. ( ). the rise of the smartphone. ieee distributed systems online, ( ). k.-s. hwang, s.-b. cho / expert systems with applications ( ) – automatic expert system based on images for accuracy crop row detection in maize fields automatic expert system based on images for accuracy crop row detection in maize fields j.m. guerrero , m. guijarro , m. montalvo , j. romeo , l emmi , a. ribeiro , g. pajares a b s t r a c t this paper proposes an automatic expert system for accuracy crop row detection in maize fields based on images acquired from a vision system. different applications in maize, particularly those based on site specific treatments, require the identification of the crop rows. the vision system is designed with a defined geometry and installed onboard a mobile agricultural vehicle, i.e. submitted to vibrations, gyros or uncontrolled movements. crop rows can be estimated by applying geometrical parameters under image perspective projection. because of the above undesired effects, most often, the estimation results inaccurate as compared to the real crop rows. the proposed expert system exploits the human knowledge which is mapped into two modules based on image processing techniques. the first one is intended for separating green plants (crops and weeds) from the rest (soil, stones and others). the second one is based on the system geometry where the expected crop lines are mapped onto the image and then a correction is applied through the well-tested and robust theil-sen estimator in order to adjust them to the real ones. its performance is favorably compared against the classical pearson product-moment correlation coefficient. . introduction . . problem statement machine vision systems onboard robots are being increasingly used for site specific treatments in agriculture. with such arrange- ments, the robot navigates and acts over a site-specific area of a larger farm (davies, casady, & massey, ), where the vision sys- tems can supply abundant information. an important issue related with the application of machine vi- sion methods is that concerning the crop row and weed detection, which has attracted numerous studies in this area (burgos-artizzu, ribeiro, tellaeche, pajares, & fernández-quintanilla, ; guerrero, pajares, montalvo, romeo, & guijarro, ; lópez- granados, ; montalvo et al., ; onyango & marchant, ; sainz-costa, ribeiro, burgos-artizzu, guijarro, & pajares, ; tellaeche, burgos-artizzu, pajares, & ribeiro, ; tellaeche, burgos-artizzu, pajares, ribeiro, & fernández-quintanilla, ). the goal is to eliminate weeds to favor the growth of crops. the vision system consists of a ccd-based calibrated camera with known intrinsic parameters, i.e. focal length, lens distortion, image center and ccd sensor sizes and pixel resolutions. the cam- era is located in front of the robot, inclined with a tilt angle (pitch) and at a high from the ground. yaw and roll angles are also known. this allows determining the rotation and translation matrices defining the extrinsic parameters. thus, areas in the field can be identified onto the image plane. this means that given an element in the field, with its spatial location, we can determine its relative positioning on the image. the vehicle navigates on a real terrain presenting irregularities and roughness. this produces vibrations and also swinging mainly in the pitch and roll angles. the yaw angle is assumed to be correct because otherwise the robot navigates erroneously out of the crop rows. moreover, the spacing of crop rows in the field is also known. because of the above, most often, the mapped expected crop rows in the image do not match with the real ones, this inaccurate esti- mation impedes the application of correct site specific treatments. on the other hand the discrimination of crops and weeds in the im- age is a very difficult task because their red, green and blue spec- tral components display similar values. this means that no discrimination is possible between crops and weeds based on the spectral signatures. thus, the best option is to locate the crop rows in the image with the maximum accuracy as possible. indeed, if the crop rows are well located, we can accurately identify those pixels along and around the detected line as crops and the remainder, which are moved away, can be considered as weeds. to achieve this goal, we propose an automatic expert system, which exploits the human knowledge, with two main modules based on image processing techniques, as described later. . . revision of methods several strategies have been proposed for crop row detection. fontaine and crowe ( ) tested the abilities of fourth line detec- tion algorithms to determine the position and the angle of the cam- era with respect to a set of artificial rows with and without simulated weeds. these were stripe analysis, hough transform, blob analysis and linear regression. the following is a list of crop row detection methods grouped into different categories including the above. . . . methods based on the exploration of horizontal strips sogaard and olsen ( ) apply rgb color image transformation to gray scale. this is done by first dividing the color image into its red, green and blue channels and then by applying the well-tested methods to extract living plant tissue described in woebbecke, meyer, von bargen, and mortensen ( ). after this, the gray scale image is divided into horizontal strips where maximum gray values indicate the presence of a candidate row, each maximum deter- mines a row segment and the center of gravity of the segment is marked at this strip position. crop rows are identified by joining marked points through a similar method to the one utilized in the hough transform or by applying linear regression. sainz-costa et al. ( ) have developed a strategy based on analysis of video sequences for identifying crop rows. crop rows persist along the directions defined by the perspective projection with respect the d scene in the field. exploiting this fact, they apply gray scale transformation based on the approach proposed by ribeiro, ferná- ndez-quintanilla, barroso, and garcía-alegre ( ) and then the image is binarized applying a thresholding technique. each image is divided into four horizontal strips. rectangular patches are drawn over the binary image to identify patches of crops and rows. the gravity centers of these patches are used as the points defining the crop rows and a line is adjusted considering these points. the first frame in the sequence is used as a lookup table that guides the full process for determining positions where the next patches in subsequent frames are to be identified. hague, tillet, and wheeler ( ) transform the original rgb image to gray scale. the transformed image is then divided into eight horizontal bands. the intensity of the pixels across these bands exhibits a periodic variation, due to the parallel crop rows. since the camera character- istics, pose and the crop row spacing are known a priori, the row spacing in image pixels can be calculated for each of the horizontal bands using a pinhole model of the camera optics. a band-pass filter can then be constructed which will enhance this pattern, and has a given frequency domain response. sometimes horizontal patterns are difficult to extract because crops and weeds form a unique patch. . . . methods based on the hough transformation according to slaughter, giles, and downey ( ), one of the most commonly used machine vision methods for identifying crop rows is based upon the hough ( ) transform. it was intended to deal with discontinuous lines, where the crop stand is incomplete with gaps in crop rows due to poor germination or other factors that result in missing crop plants in the row. it has been intended for real- time automatic guidance of agricultural vehicles (astrand & baerveldt, ; hague, marchant, & tillett, ; leemans & destain, ; marchant, ). it is applied to binary images, which are obtained by applying similar techniques to the ones explained above, i.e. rgb image transformation to gray scale and binarization (tellaeche, pajares, burgos-artizzu, & ribeiro, ; tellaeche et al., ; tellaeche, burgos-artizzu, pajares, ribeiro, et al., ). gee, bossu, jones, and truchetet ( ) apply a double hough transform under the assumption that crop rows are the only lines of the image converging to the vanishing point, the remainder lines are rejected, additional constraints such as inter-row spacing and perspective geometry concepts help to identify the lines. it is required to deter- mine the threshold required by the hough transform to determine maximum peaks values (jones, gée, & truchetet, a, b) or predominant peaks (rovira-más, zhang, reid, & will, ). depending on the crop densities several lines could be feasible and a posterior merging process is applied to lines with similar parame- ters (tellaeche et al., ; tellaeche, burgos-artizzu, pajares, ribeiro, et al., ; tellaeche et al., ). although intended for real-time, as mentioned before, in our images, where crop and weed plants contribute on the hough parameter estimation, this method becomes computationally expensive (ji, & qi, ). on the other hand, the randomized hough transform requires selecting pairs of points to be considered as a line, i.e. pairs of points belonging to a crop row. if we apply this technique in images where edge points have been extracted, the selection of those pairs becomes highly complex because weeds are also involved. . . . vanishing point-based pla, sanchiz, marchant, and brivot ( ) propose an approach that identifies regions (crops/weeds and soil) by applying color im- age segmentation. they use the skeleton of each defined region as a feature to work out the lines that define the crop. the resulting skeletons and their properties, defined as chains of connected con- tour points, allow the identification of crop rows oriented toward the vanishing point. this process is highly dependent of skeletons, which are not always easy to extract, specially taking into account that weed patches are present. romeo et al. ( ) apply also knowledge concerning the position of the vanishing point and the crop rows arrangement in the field to detect the expected crop rows. the process is based on the identification of maximum accu- mulation of green pixels along lines oriented toward the vanishing point. a supervised fuzzy clustering method is the proposed strat- egy for greenness identification. this makes the method highly dependent on the training phase unlike the one prosed in this ap- proach which is automatic, i.e. unsupervised. . . . stereo-based approach kise, zhang, and rovira-más ( ) or kise and zhang ( ) developed a stereo vis ion-based agricultural machinery crop-row tracking navigation system. stereo-image processing is used to determine d locations of the scene points of the objects of interest from the obtained stereo image. those d positions, determined by means of stereo image disparity computation, provide the base information to create an elevation map that uses a d array with varying intensity to indicate the height of the crop. this approach re- quires crops with significant heights with respect the ground. be- cause in maize fields, during the treatment stage, the heights are not relevant, it becomes ineffective in our application. rovira-más, zhang, and reid ( ) have applied and extended stereovision techniques to other areas inside precision agriculture. only feasible if crops or weeds in the d scene display a relevant height. . . . methods based on blob analysis this method finds and characterizes regions of contiguous pix- els of the same value in a binarized image (fontaine & crowe, ). the algorithm searches for white blobs (inter-row spaces) of more than pixels, under the assumption that smaller blobs could represent noise in the crop rows. once the blobs were iden- tified, the algorithm determined the angle of their principal axes and the location of their centre of gravity. for a perfectly straight white stripe, the centre of gravity of the blob was over the centre line of the white stripe, and the angle was representative of the angle of the inter-row spaces. the algorithm returned the angle and center of gravity of the blob closest to the center of the image. identification of blobs in areas with weed patches does not distin- guish between blobs caused by weeds and crops. . . . methods based on the accumulation of green plants olsen ( ) proposed a method based on the consideration that along the crop row appear an important accumulation of green parts in the image. the image is gray scale transformed where green parts appear clearer that the rest. a sum-curve of gray levels is obtained for a given rectangular region exploring all col- umns in the rectangle. it is assumed that vertical lines follow this direction in the image. the images are free of perspective projec- tion because they are acquired with the camera in orthogonal po- sition. a sinusoidal curve is fitted by means of least squares to the sum-curve previously obtained. local maxima of the sinusoid pro- vide row centers locations. . . . methods based on frequency analysis because crop rows are vertical in the d scene, they are mapped under perspective projection onto the image displaying some behavior in frequency domain. vioix et al. ( ) exploit this fea- ture and apply a bi-dimensional gabor filter, defined as a modula- tion of a gaussian function by a cosine signal. the frequency parameter required by the gabor filter is empirically deduced from the d-fast fourier transform (bossu, gee, guillemin, & truchetet, ). bossu, gee, jones, and truchetet ( ) apply wavelets to discriminate crop rows based on the frequency analysis. they exploit the fact that crop rows are well localized in the frequency domain; thus selecting a mother wavelet function with this fre- quency the crop rows can be extracted. crops, in the images we have studied, do not display clear frequency contents in the fourier space, therefore the application of filters based on the frequency becomes a difficult task. . . . methods based on linear regression some techniques above apply this approach. billingsley and schoenfisch ( ) reported a crop detection system that is rela- tively insensitive to additional visual 'noise' from weeds. they used linear regression in each of three crop row segments considered and a cost function analogous to the moment of the best-fit line to detect lines fitted to outliers (i.e., noise and weeds) as a means of identifying row guidance information. montalvo et al. ( ) ap- ply a linear regression for crop row detection in images containing high weeds densities. some templates are used to guide the detec- tion. linear regression is also applied in sogaard and olsen ( ). linear regression is highly sensitive to isolated weeds patches placed on the inter crop rows and also for weeds patches over- lapped with crops. in this paper we also apply linear regression based on the theil sen estimator (sen, ; theil, ) which is free of the above sensitivity and has been proven in statistics with satisfactory results. . design of the automatic expert system . . system architecture the system architecture is inspired on the human expert knowl- edge about the specific application and also considering the require- ments that must be fulfilled. astrand ( ) and slaughter et al. ( ) propose a list of requirements for guidance systems that can be also considered for crop row detection, which in essence is a similar problem. knowledge and requirements are mapped as fol- lows to build the architecture of the proposed automatic expert sys- tem for the accuracy crop row detection based on images. (a) both crop and weeds display similar color spectral compo- nents and during the treatment their growth stages are sim- ilar, i.e. with similar height in the plants. (b) crop rows are accumulations of green plants following spe- cific alignments oriented to the vanishing point. crops are sown, not manually planted, and the inter-line distances in the field are known. (c) weeds appear on isolated or overlapped patches with respect crops with irregular distributions. (d) crop rows must be located with the most accuracy as possi- ble, regardless the distribution of weeds patches around crop and also considering that crop plants could miss along crop lines, as a common situation. original ¡mage image segmentation apply combined vegetation indices & greenness reinforcement otsu thresholding binary image (crop lines & weeds patches) expert system crop row detection crop row estimation trace expected crop lines according to the camera system geometry apply theil-sen estimator fig. . automatic expert system architecture. (e) camera s y s t e m g e o m e t r y is k n o w n , i.e. t h e intrinsic a n d extrinsic p a r a m e t e r s . (f) the robot navigates on u n e v e n t e r r a i n s w i t h p e r h a p s a b u n - d a n t irregularities. (g) the s y s t e m m u s t w o r k on r e a l - t i m e . this r e p r e s e n t s a t r a d e - off b e t w e e n t h e speed of t h e robot a n d t h e c o m p u t a t i o n a l cost. based on this k n o w l e d g e a n d r e q u i r e m e n t s a n d also considering a d v a n t a g e s a n d shortcomings of t h e different crop r o w d e t e c t i o n m e t h o d s , t h e a u t o m a t i c e x p e r t s y s t e m is designed consisting of t w o m a i n m o d u l e s : image segmentation a n d crop rows estimation. fig. displays schematically t h e s e t w o m o d u l e s w i t h t h e c o r r e - s p o n d i n g processes. this results in a r o b u s t e x p e r t system, m a k i n g t h e c o n t r i b u t i o n of this paper. . . image segmentation image s e g m e n t a t i o n is focused on t h e s e p a r a t i o n of g r e e n plants (crops a n d w e e d s ) from t h e rest (soil, s t o n e s a n d o t h e r s ) . according t o point (a) in t h e list of k n o w l e d g e a n d r e q u i r e m e n t s above, t h e b e s t o p t i o n t o identify w e e d s a n d c r o p is t h e application of v e g e t a t i o n indices i n s t e a d of m e t h o d s b a s e d on height discrim- ination. vegetation indices are well t e s t e d m e t h o d s , guijarro et al. ( ) p r o p o s e a c o m b i n a t i o n of v e g e t a t i o n indices, w h i c h is t h e o n e c h o s e n in this p a p e r b e c a u s e its p e r f o r m a n c e in m a i z e fields. . . . combination of vegetation indices given a n original i n p u t i m a g e in t h e rgb color space, w e apply t h e following n o r m a l i z a t i o n s c h e m e , w h i c h is usually applied in a g r o n o m i c i m a g e s e g m e n t a t i o n (gee et al., ), rn c„ , b„ r„ b„ r„ b„ b = rn b„ ( ) w h e r e r, g a n d b are t h e normalized rgb coordinates ranging from t o and are o b t a i n e d as follows: r„ rn b„ bn ( ) w h e r e rm x = gm a x = b m a x = for our -bit color images. vegetation indices t o b e c o m b i n e d are c o m p u t e d as follows guijarro et al. ( ), excess green ( w o e b b e c k e et al., ; ribeiro et al., ) color index of vegetation extraction (kataoka, kaneko, okamoto, & hata, ) vegetativen ( h a g u e et al., ) ( ) exg = g -r -b c/ve = . r - . g ( ) + . b + . veg = ( ) r"b w i t h a = . as in its reference excess green minus excess red ( ) (meyer & camargo-neto, = _ ; neto, ) w h e r e excess red is c o m p u t e d as follows (meyer, hindman, & lakshmi ): exr = . r - g . according t o guijarro et al. ( ) t h e above four indices are c o m b i n e d t o obtain t h e resulting value com as follows, com = wexgexg + wexgexgr + w q v e c / v e + wvegveg ( ) w h e r e wexg = . , wexgr = . , wcjv£ = . a n d wveg = . are t h e w e i g h t s for each index, r e p r e s e n t i n g t h e i r relative relevance in t h e combination. the resulting c o m b i n e d image com, is linearly m a p p e d t o range into t h e interval [ , ]. . . . greenness reinforcement romeo et al. ( ) p r o p o s e a fuzzy clustering strategy w h e r e t h e cluster c o n t a i n i n g pixels belonging t o g r e e n plants has b e e n analyzed. clusters c o n t a i n pixels w i t h t h e t h r e e spectral c o m p o - n e n t s in t h e rgb m o d e l as features. obviously, a n d as expected, t h e g r e e n spectral c o m p o n e n t is d o m i n a n t . on a v e r a g e this c o m p o - n e n t in t h e cluster c e n t e r for g r e e n plants r e p r e s e n t s values a b o v e t h e % w i t h r e s p e c t t h e o t h e r t w o c o m p o n e n t s . exploiting this k n o w l e d g e a n d applying t h e trivial reasoning t h a t pixels c o m i n g from plants should have t h e i r g r e e n c o m p o n e n t d o m i n a n t , w e a c c e n t u a t e t h e g r e e n n e s s in com by multiplying t h e i r values by g in eq. ( ), i.e. a n e w g r e e n n e s s is o b t a i n e d a s : ga = com*g. the multiplication is carried o u t pixel by pixel a n d ga is linearly m a p p e d to range in [ , ]. because g r e p r e s e n t s t h e p e r c e n t a g e of t h e g r e e n c o m p o n e n t , t h e result o b t a i n e d r e p r e s e n t s t h e e m p h a s i s in t h e g r e e n n e s s . . . . thresholding given t h e t r a n s f o r m e d i m a g e ga, t h e n e x t s t e p is its binariza- tion for posterior processing. an easy t h r e s h o l d b a s e d on t h e m e a n gray level of t h e i m a g e ( h i s t o g r a m ) has b e e n i m p l e m e n t e d in gee et al. ( ) w h e r e t h e living plant m a t e r i a l (crop or w e e d ) a p p e a r s as w h i t e spots a n d t h e rest (i.e. soil surface, stones, s h a d o w s ) as black. also in guijarro et al. ( ) t h e w e l l - k n o w n otsu's ( ) m e t h o d , traditionally applied for binarization, has b e e n applied. more c o m p l e x a p p r o a c h e s h a v e b e e n also applied such as t h e o n e u s e d in bossu et al. ( ), b a s e d on t h e fe-means clustering m e t h o d . w e h a v e c h o s e n t h e otsu's m e t h o d for its w e l l - k n o w n p e r f o r m a n c e as r e p o r t e d in meyer a n d camargo-neto ( ) a n d also b a s e d on t h e s t u d y of sezgin a n d sankur ( ) w h e r e its per- formance has b e e n t e s t e d in images w h e r e t h e n u m b e r of pixels in b o t h p a r t s of t h e i m a g e h i s t o g r a m t h a t otsu's p r o d u c e s is close t o e a c h o t h e r . fig. (a) displays a n original i m a g e in t h e rgb color space of a m a i z e crop field. the color space t r a n s f o r m a t i o n by a p p l y i n g ga is displayed in fig. (b). fig. (c) displays t h e i m a g e t r a n s f o r m a t i o n from i m a g e in fig. (b) by applying t h e otsu's m e t h o d . note t h e l a n d m a r k s in t h e image, w h i c h a r e explained later in section . . . crop row estimation this m o d u l e is i n t e n d e d t o apply t h e k n o w l e d g e e m b e d d e d in points (b)-(f), section . , at t h e s a m e t i m e it provides specific solutions for t h e r e q u i r e m e n t s e x p r e s s e d in such points. . . . tracing expected crop lines the robot navigates on u n e v e n t e r r a i n s w i t h p e r h a p s a b u n d a n t irregularities, t h e k n o w l e d g e of extrinsic p a r a m e t e r s of t h e vision s y s t e m does not suffice b e c a u s e t h e c a m e r a is c o n t i n u o u s l y in- volved in a p e r m a n e n t swinging. w e p r o p o s e t h e c u s t o m i z a t i o n of t h e theil-sen regression e s t i m a t i o n a p p r o a c h , b e c a u s e of its w e l l - t e s t e d p e r f o r m a n c e in statistics. because t h e c r o p rows a r r a n g e m e n t are k n o w n in t h e field a n d also t h e extrinsic a n d intrinsic c a m e r a s y s t e m p a r a m e t e r s , t h e ex- p e c t e d c r o p r o w locations in t h e i m a g e can b e e s t i m a t e d a n d m a p p e d as k n o w n lines o n t o t h e i m a g e (fu, gonzález, & lee, ; hartley & zisserman, ). u n d e r t h e a s s u m p t i o n of ideal s y s t e m g e o m e t r y t h e e x p e c t e d lines should m a t c h a n d overlap t h e imaged real crop r o w s . nevertheless, d u e t o u n e v e n t e r r a i n s a n d errors in t h e c r o p r o w a l i g n m e n t d u r i n g t h e sowing, this often does not occur. (a) t \ '% (c) fig. . (a) original image; (b) ga index extracted from the image in (a); (c) binary image after otsu thresholding. therefore, under the above consideration, two cases can appear with respect to the expected and the imaged real crop lines: (a) they match; (b) they do not match. in the first case, the detection method needs to verify this matching. in the second case, a line location correction must be applied until the real crop row is lo- cated. under this approach the system geometry through the intrinsic and extrinsic parameters guides the crop row detection process. now the question is: how can we verify the expected lines match or not with the real crop ones? because we have available white pixels representing green plants in the binary image, we can adjust a straight line for specific pixel alignments that are ex- pected to identify crop rows. this will represent the real crop line. so, because we have both, the expected straight line equation and the adjusted one, we are able to verify the correct or incorrect match for both lines. thus, we focus the effort in methods for esti- mating the parameters defining real crop lines. . . . correction of the expected crop lines: theil-sen estimator an important problem to be addressed in our approach is that the method selected can cope with specific pixel alignments but also must be robust enough to avoid significant deviations caused by weeds that are not aligned and placed more or less near the main crop row alignments. this is the main issue addressed in this work. stewart ( ) provides a tutorial oriented toward robust parameter estimation in computer vision. two frequently tech- niques used are least-median of squares (lms) (rousseeuw, ) and m-estimators (hampel, rousseeuw, ronchetti, & stahel, ; huber, ), but a huge volume of data implies that param- eter estimation techniques in computer vision are heavily over constrained, even for problems where low-level feature extraction, such as edge detection, are applied. this in turn implies that parameter estimation problems in vision should be solved by least squares or, more generally, maximum likelihood estimation (mle) techniques. unfortunately, computer vision data are rarely drawn from a single statistical population as required for effective use of mle. in our approach we must estimate two parameters, defining the straight line equations associated to the corresponding crop rows, they are the slope a and the intercept /¡. for the linear regression approach, after several studies, we observed that the least squares estimator of a regression, coefficient a is vulnerable to gross errors and the associated confidence interval is, in addition, sensitive to non-normality of the parent distribution. with other measures, for example, the breakdown point (rousseeuw & leroy, ) a small number of outlying data can cause an estimate to diverge arbitrarily far from the true estimate. from the point of view of our approach, this means that few weed pixels can move the least squares fit far from the true fit, i.e. far of the real crop line. a second measure of robustness is the influence function (hampel et al., ; huber, ), in which the change in an estimate caused by insertion of outlying data, using a function of the distance, also causes false estimations because it should tend to zero with increasing distance to achieve robustness. alternative estimators for the regression coefficient, a, based on suitable rank tests are proposed by mood ( ). they apply the estimation of both parameters a and ¡ simultaneously using the statistical median by trial and error. adichie ( ) proposes a more restrictive method under the assumption that the set of points to be adjusted is an absolutely continuous and symmetric distribution function with also an absolutely continuous and square integrable density function; theil ( ) proposes a very simple estimator for a using also the statistical median; dytham ( ) and sen ( ) study a simple and robust estimator for a based on kendall's ( ) tau rank correlation, a simple non- parametric test that can be used instead of normal regression. hence, the estimator for a and ¡ is the median of the set of slopes, where a simple slope is computed between every possible pair of pixels i and j with image coordinates (x„ y,) and (x,-, y¡) respec- tively, finally the median slope is then selected as the best estimate for a. based on the above considerations, we select the theil-sen esti- mator as proposed in massart ( ), because of its statistical effi- ciency and its robustness, even for low image resolutions, resulting in a promising approach in agricultural images containing crop rows. nevertheless, as we will see later, its effectiveness from the real-time point of view is relatively low. this means that further analysis or new software and hardware implementations should be required for real-time processing. a straight line is represent by its slope a and its intercept ¡ as follows, y = ax- ( ) given a distribution of n pixels the goal is to adjust a straight line to such distribution. the theil-sen estimator evaluates pairs of pixels i and j and compute the slope over the set of all possible pairs of such pixels, i.e. over the n(n - l)/ possible combinations. this is carried out as follows, a = med[sij\sij = y¡-y¡ x¡ t^xj, i,j = , , ( ) the estimation of the intercept, /¡, is computed as the statistical median of the intercepts obtained with the robust slope a in ( ). the e parameter, set to ~ , is introduced to avoid exactly vertical lines with slope toward oo. polar coordinates could be used to avoid this problem. nevertheless, vertical lines do not appear in our real application. this is carried out as follows, • med(y¡ - ~p � ~a ð Þ� ~a �� �� . similarly, the projection ratio of the data vector ~p in c on ~a is larger than the projection ratio on ~a , i.e., ~p � ~a ð Þ � ~a �� �� < ~p � ~a ð Þ� ~a �� �� . where c(pi) is cluster assignment for pi, and ~pi is the vector from ref- erence center to pi. according to the cluster assignments, there are m side-clusters, and how data points are partitioned into side clus- ters will be described in the following paragraph. fig. illustrates the basic concept of how a side-cluster is formed. let ~a and ~a be normal vectors of boundary edges b adn b , respectively. the decision boundary between two side- clusters c and c is drawn in dot line, which is a hyperplane pass- ing through the origin point o and the intersection point i of the two boundary edges b and b . for a vector ~p from reference cen- ter to a point p in c , its projection ratio on ~a is larger than the projection ratio on ~a . that is, p.-s. lai, h.-c. fu / expert systems with applications ( ) – ~p � ~a ~a �� �� > ~p � ~a ~a �� �� : by segmenting the data points of a cluster into side clusters corre- sponding to the line segments between cluster centers and by pro- jecting the data points in a side cluster onto the corresponding line segment between cluster centers, an -d data distribution can be created. the -d data distribution can be used to measure the var- iance along the line segment between cluster centers or adjust the location of decision boundary between clusters. however, since the coordinates of data points for k-medoid model is absent, a new method is proposed in section . to estimate the projection length without knowing the coordinate of data points. . . projections on the center-to-center link computing variance along normal vector using projection ratio requires the coordinate values of data points in feature space. in or- der to computing the projection ratio on normal vectors without knowing the coordinates of data points in feature space, a new pro- jection method is proposed. as shown in fig. , given a point p, and two clusters with center points at m and m . m pi !��� ���; m pi! ��� ���; and m m !��� ��� are the norms from p to m , p to m , and m to m , respectively. let d be the pro- jection of p on m ; m ! . m d ! ���� ���� is the projection length of m pi! on m m ! . as shown in fig. , the following relation holds: m pi !��� ��� ¼ m d! ���� ���� þ pi d ! ���� ���� ; m pi !��� ��� ¼ m m !��� ���� m d! ���� ���� � � þ pid ! ���� ���� : then, the length of m d ! ���� ���� can be derived as follows: m m !��� ���� m d! ���� ���� � � ¼ m pi !��� ��� � m pi! ��� ��� þ m d! ���� ���� ; m m !��� ��� � m m !��� ��� m d! ���� ���� ¼ m pi! ��� ��� � m pi! ��� ��� ; m d ! ���� ���� ¼ m m !��� ��� � m pi! ��� ��� þ m pi! ��� ��� m m !��� ��� : by projecting data points onto the line segment between two clus- ter centers, the -d data distribution along the line segment be- comes available. based on the -d data distribution, the decision boundary can therefore be adjusted to maximize the decision mar- gin between two estimated clusters. . . decision boundaries adjustment using the projection method proposed in section . , data points in a cluster can be segmented into several side-clusters cor- fig. . m and m are two cluster centers. p is a data point. d is the projection point on m m . responding to the line segments between each pair of cluster cen- ters. the projection of data points in a side-cluster on the corresponding line segment between cluster centers forms an -d data distribution. based on the -d data distribution, the var- iance of data points along the line segment between cluster centers can be measured. since the purpose of measuring variance is to ad- just the location of decision boundary, the estimation or the mea- suring of variance can be avoid by directly adjusting the location of decision boundary according to projected -d data distribution. as shown in fig. , given two cluster centers m and m . by pro- jecting data points pi (shown in black square dots) onto the line segment m m , a series of -d location values (shown in black tri- angle dots) are measured. the following iterative procedure is pro- posed to find a decision boundary, such that the distance between the means of two partitioned clusters can be maximized. these -d location values are partitioned into two groups namely g and g according to the decision boundary b . the decision boundary b is determined to maximize the distance between l and l , where l and l are the means of data groups g and g , respectively. let a be the target location of decision boundary, and the pro- jection ratio h(pi) for point pi is defined as hðpiÞ¼ m pi ! �m m ! m m ! �m m ! : actually, h(pi) is the projection length of vector m pi ! on m m divid- ing by the distance between two cluster centers. then, a can be evaluated according to the following formula: a ¼ arg max a p hðpiÞ>a hðpiÞp hðpiÞ>a � p hðpiÞa hðpiÞp hðpiÞ>a is the average projection ratio that is larger than a, and p hðpiÞ i, data points in ci and cj are projected onto mimj ! , where mi and mj are medoids of ci and cj, respectively. then, the line segment between ci and cj is parti- tioned into intervals. by counting the number of points locating in one interval, a histogram is drawn. based on the histogram, the decision boundary can be properly located so that the distance be- tween two means of data sets, which are separated by the decision boundary, is maximized. all data points are then redistributed using the adjusted decision boundaries. the process repeats until the location of decision boundaries converged, such that the dis- tance between l and l is maximized. fig. . data points pi are projected to the line segment between cluster centers m and m . the decision boundary is located at the place which partition the projected data points (triangle points) into two groups with farthest mean location l and l . fig. . the flow chart of variance enhanced k-medoid clustering. at first, general k-medoid algorithm is applied. according to the resulted clusters of general k-medoid algorithm, the cluster-to-cluster variance is estimated. then, the weights for each data points are updated and new cluster is estimated based on the weighted data points. these steps are repeated until all weights are smaller than a given threshold q. p.-s. lai, h.-c. fu / expert systems with applications ( ) – . . variance enhanced k-medoid clustering method the flow chart of variance enhanced k-medoid clustering is shown in fig. . at first, one medoid is estimated. then, the num- ber of medoid is gradually increased until the distance from each data point to the closest cluster center is smaller than a given threshold. for each medoid, its positions are estimated using gen- eral k-medoid algorithm first. then, the -d data distributions can be established by projecting data points in side-clusters to the cor- responding line segment between cluster centers. according to these -d data distributions, the location of decision boundaries is adjusted to redistributed data points to each clusters. after the position of medoid converged, weights wi of each points pi is calcu- lated as the following formula: wi ¼ � þ min pm m simðpi; pmÞ � �� ; where m is a set of estimated medoids, and sim (pi, pm) is the pro- jected similarity defined as follows: simðpi; pmÞ¼ arg max pj m;pj – pm ~pi � ~pj � ~pm � a ~pj � ~pm � � a ~pj � ~pm � ; where a ~pj � ~pm � is the vector from pm to the decision boundary be- tween the clusters associated to pm and pj. based on these weights, a new medoid is created to include high weighted data points. by repeating these processes, the number of medoid gradually increas- ing until the maximal weight is lower than a given threshold. using the proposed variance enhanced k-medoid, the location of decision boundaries between each pair of cluster centers is re- lated to the data distribution along a line segment between cluster centers now. a real-world application, photo album preview with variable-sized thumbnails, is proposed and implemented in sec- tion to illustrate the performance of the proposed variance en- hanced k-medoid clustering method. . real-world application in this section, variance enhanced k-medoid is used to intelli- gently create image thumbnails, where ‘‘intelligently” means that the size and the display order of image thumbnail are decided without human intervention. generally, image gallery (pao et al., ) uses thumbnail to provide a quick preview for each photo (pao et al., ). however, even though the thumbnail is down- sized from the original photos, the display area is often still too small to contain all thumbnails. to contain all thumbnails of same size can be very difficult in selecting a proper size for all the thumbnails. sometimes, it may be too crowded to clearly represent the original image, or it may be too large to put all thumbnails in a page. therefore, variable-sized thumbnail seems a better alterna- tive to fit all thumbnails into a page without losing too much details. since most of images in an image gallery may be similar to each other, these images can be partitioned into groups according to their similarity. for each group, a most representative image can be displayed by a normal sized thumbnail, and the rest of similar images can be shown by smaller thumbnails. the detail description of the proposed system is depicted in the following sections. color information of an image is used to create four bitmaps to show the distribution of dominant color compo- nents. based on these bitmaps, the similarity between two images is measured. then, the proposed variance enhanced k-medoid is applied to cluster all images into groups. . . color clustering giving an image i, k’s dominant color components are selected to establish k’s bitmaps which show the coverage of data distribu- tion of dominant color components. k-means algorithm is used to cluster the color value of each pixel px,y i. let k = , then four dominant colors mi, for i = , . . ., , are estimated by k-means algo- rithm. then, four bitmaps i(x, y)i are created as follows: iðx; yÞi ¼ ; if arg min mi normðmi; ciÞ¼ mi ; if arg min mi normðmi; ciÞ – mi >< >: for i ¼ ; . . . ; ; where i(x, y)i is the bitmap created using dominant color mi, and ci is the ith color component at (x, y). an example is shown in fig. , the size of each image is nor- malized to � at first. then, k-means algorithm is used to cluster pixels of original image to four color component groups according to the color values of pixels. and, four bitmaps are estab- lished to show the distribution of data points in each groups. these four bitmaps will be used to measure the similarity between images (in sections . and . ). . . similarity measurement of two bitmaps giving two normalized bitmaps that show the coverage area for data distribution of specified color components, the similarity s(a, b) between two bitmaps a and b is defined as follows: sða; bÞ¼ a\bk k a[bk k ; where a and b are the sets containing the pixels with value in bitmap a and b, respectively. as shown as fig. , the similarity is defined according to the ratio between overlapped (d) and overall (a [ d [ b) area of a ða [ dÞ and b ðb [ dÞ, where a, b, and d repre- sent the shaded areas in fig. . since four bitmaps are extracted to represent the color charac- teristics of an color image, the similarity between two images can be estimated based on the group similarity sg between two group of bitmaps. the computation of group similarity can be performed on a best match between the elements in bitmap groups. the best match can be searched by the method proposed in section . . according to the best match, the group similarity measurement sg is defined as follows: sg ¼ x i kaikþkbikp j kajkþkbjk � � sðai; biÞ; fig. . using k-means algorithm, pixels of original image, for example in (a), are clustered into four dominant groups according its color value. then, four bitmaps, for example (b), are established to show the data distribution of pixels in each dominant groups. fig. . giving two coverage area of data distribution a and b, the similarity between these two coverage area is defined according to how much portion of these two coverage area are overlapped (d). p.-s. lai, h.-c. fu / expert systems with applications ( ) – whereai and bi are the sets containing the pixels with value in bit- map ai and bi, respectively. that is, the group similarity measurement sg is the weighted sum of similarities between each pair of bitmaps, among the total number of pixels in each pair of bitmaps. . . similarity matching between bitmap groups for each image, four bitmaps are extracted to represent the col- or characteristics of an image. to measure the similarity between two images, a one-to-one and onto match (liu, ) between ele- ments of two bitmap groups can be performed by the following proposed method. giving two bitmap groupsa¼ a ; . . . ; anf g and b¼ b ; . . . ; bnf g, where ai and bi are bitmaps. an n � n matrix m is formed in the fol- lowing manner: m ¼ m � � � m n .. . . . . .. . mn � � � mnn ; where mij ¼ sðai; bjÞ: p.-s. lai, h.-c. fu / expert systems with applications ( ) – since an element mij represents the similarity value when ai and bi are considered to be matched with each other. finding a one-to-one and onto match between a and b is equivalent to selecting n ele- ments from each column of m, and no two elements are selected from the same row. to have the largest overall similarity sgða;bÞ of a one-to-one and onto match between two bitmap groups a and b, the n elements should be selected such that the sum of se- lected elements is maximized. in order to regularize the computing procedure, row exchanges are performed to have the n selected ele- ments aligned on the diagonal position. thus, the similarity value between bitmap groups can be computed as the sum of diagonal elements. fig. shows the flow chart of the proposed method that max- imizes the sum of diagonal elements by properly exchanging rows. giving a similarity matrix m. first, find an element mij with the largest value in a similarity matrix m. then, swap the ith row and jth row such that the element mij is at (j, j). then, mark all ele- ments in jth row and jth column. then, find elements with maxi- mal value from un-marked elements and swap the corresponding raw and column, such that the maximal element is moved to diag- onal position. repeat the same process, until all elements are marked. although the procedures stated above are efficient, its results may not be good enough. therefore, a checking step is applied on each rows and columns to see whether any possible row exchange fig. . the flow chart of the proposed matching method. at the beginning, a similarity matrix is constructed. the order of rows is then initialized by greedily exchanging the largest element to diagonal position. then, further checking steps are applied on the diagonal elements to maximize the sum of diagonal elements. can further maximize the sum of diagonal elements in the similar- ity matrix m. the checking step is described as follows: fig co se sim th th ro res if mii + mjj > mij + mji then exchange the ith and jth rows. this checking step is performed repeatedly until no further row exchange can be performed to increase the sum of diagonal ele- ments. in other words, the near maximum similarity value be- tween bitmap group a and b is achieved. an example of the one-to-one onto match method is shown in fig. . at first, since m = is the largest value in the � ma- trix, the elements in first row and first column are marked. since the value m = , the largest value among the unmarked ele- ments is located in the third row, and the second row and the third row are exchanged. then, the elements in second row and second column are marked too. similar process is repeated until all col- umns are marked. according to the checking step, the second and the fourth rows are exchanged, since + > + . in order to evaluate the performance of the proposed one-to- one and onto match algorithm, synthesis testing data sets are gen- erated by a random number generator. to evaluate different size of similarity matrices, groups of test data sets with row dimension from to are generated. for each group, random seeds from to are used to generate test data. the testing results are listed in table . label ost (optimizing sorting test) is performed by the proposed method, and label tat (testing all test) indicates an exhaustive search method that vali- dates all the possible solutions. the tat results are considered as a b c . . an example of the proposed matching method. the similarity matrix that is nstructed from the pair-wise similarity values is shown in (a). at first, by lecting an element with the largest value among unmarked elements in a ilarity matrix, swapping the selected element to diagonal position, and marking e elements in the same row and column iteratively until all elements are marked. e result is shown as (b). then, the checking steps are applied to see whether any w exchanges can be applied to maximize the sum of diagonal elements. the final ults are shown in (c). table the performance comparison of the proposed one-to-one match algorithm (ost) and the exhaustive search method (tat). to evaluate different size of similarity matrices, groups of test data sets with row dimension from to are generated. the data listed in table are the averaged computing time in milliseconds spent by the matching process. tat/ost is the ratio of the spending time by these two methods. matrix dimension computing time (ms) tat/ost tat ost . . . . , . , . , , . , , . , , . , , , . p.-s. lai, h.-c. fu / expert systems with applications ( ) – the ground truth. the performance is measured in milliseconds of the total computing time to process all the test data. apparently, the proposed method is quite efficient. besides, based on the , ( group of random seeds) testing data, the proposed method generates the same results as the ground truth. . . images clustering by enhanced k-medoid method giving an image set i ¼ i ; . . . ; inf g, a matrix m is constructed by measuring the similarity between two images using the meth- ods proposed in sections . , . , and . . the image similarity ma- trix m for image comparison is defined as follows: m¼ m � � � m n .. . . . . .. . mn � � � mnn ; where mij ¼ sgðii; ijÞ: fig. . according to the web-based image clustering prototype system, the proposed similarity threshold is set as . . that is, the similarity between a data object and the rep then, the proposed variance enhanced k-medoid model clusters a set of images using the pair-wise image similarities sg represented in the image similarity matrix m. the image corresponding to the medoid of a data cluster is selected as the representative image for each image cluster. . system demonstration a web-based prototype system (http://www.csie.nctu.edu.tw/ pslai/askmddemo) is implemented to demonstrate the proposed method for public trial and testing. fig. shows a web-based user interface for image clustering. as shown in fig. , a user may select a test media (http://www.youtube.com/watch?v = mgpzuo_ kbfy) set from the selector at the top-right corner of the web interface. the threshold, which represents the maximal allowed difference between data objects and its cluster center, is selected by clicking at a proper position on the ruler at the top-left corner. then, the images in the test media set are clustered. and, the re- sulted clusters are presented in the order of cluster size. for each cluster, the represented image of each cluster is displayed in larger size, and the rest of images in a cluster are shown in smaller-sized thumbnail. figs. – show the testing results using three different thresholds on the same media data set. in general, using higher threshold may reduce the number of clusters. more specifically, enlarging similarity threshold may reduce the similarity among data elements in certain clusters, so as to receive more not so sim- ilar data elements in a cluster. that is, the maximum allowed var- iance is increased. however, for data clusters containing quiet similar data elements, enlarge the similarity threshold will only slightly increase the number of less similar data elements, thus the maximal allowed variance has little inference on these data clusters. in addition, noisy data elements are clustered into certain less-similar clusters instead of spreading to all clusters. method successfully clusters similar images together. in this demonstration, the resentative object (cluster center) of a cluster should be larger than . (= � . ). fig. . by increasing the similarity threshold to . , certain data clusters receive several not so similar data objects (red frame a). however, the data clusters, which are previously contained quiet similar data elements, still have almost the same similarity among each other (clusters marked with b). (for interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.) fig. . by further increasing the similarity threshold to . , data clusters previously contained some not so similar objects will grow to accept more dis-similar objects (red frame a). however, these data clusters, which previously contained quiet similar objects, will still have very similar objects (clusters marked with b). (for interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.) p.-s. lai, h.-c. fu / expert systems with applications ( ) – . conclusion generally speaking, grouping similar data together is useful to help user browse data efficiently. however, the similarity of data are difficult to define. often, the definition of similarity between data objects depends on their applications. on the other hand, most clustering algorithms are developed for data sets which are represented as data points in feature space. that is, before p.-s. lai, h.-c. fu / expert systems with applications ( ) – clustering similar data, a feature extraction process is applied to each data objects to represent data objects as points in feature space. however, since the definition of similarity is different for various applications, and mapping the data objects into a feature space reasonably can be difficult and also complicated. most of the time, proper feature extraction becomes a bottleneck for appli- cation system development. instead of mapping data objects into points in feature space, an- other way to show the similarity among data objects is to measure the pairwise similarity among data objects. describing the similar- ity between two data objects is more instinctively than mapping data objects into a feature space since computing similarity be- tween pair of data objects only need the special relationship be- tween data objects. a popular clustering algorithm, called k-medoid, is often used to cluster data objects using the pairwise similarities between data objects. k-medoid clustering algorithm groups data objects into clusters by maximizing the sum of similarity between each data objects and their cluster center. that is, the variance of each cluster still has the same value. in this paper, variance enhanced k-medoid clustering is pro- posed to introduce the idea of data variance into the original k- medoid algorithm. besides, multiple variances are used to describe data variances along line segments between cluster centers. that is, the variance of cluster in different orientation is considered. an intelligent image thumbnail system is prototyped to demon- strate the proposed variance enhanced k-medoid clustering meth- od. for each image, the covered regions of four dominant colors are segmented at first. between two images, the differences in four dominant color regions are measured as the similarity between images. then, the variance enhanced k-medoid model is applied on these pairwise image similarities to group images into clusters. and, the image at cluster center (medoid) is used as the key image for this group. the web-based thumbnail demonstration show that the system successfully groups similar images together, and the key image is a good representative of the images in the cluster. by increasing the similarity thresholds of the proposed clustering method, the number of clusters will decrease. and, the clusters that contained quiet similar images before the changes will only changes slightly, i.e., receive a few more images, although the clus- ter that contained some not so similar images will grow bigger due to more outlier images are received. such properties of the pro- posed model are useful such as selecting representative frame-sets from video for various video summary applications. references bilmes, j. ( ). a gentle tutorial on the em algorithm and its application to parameter estimation for gaussian mixture and hidden markov models, tech. rep., icsi-tr- - , university of berkeley. comaniciu, d., & meer, p. ( ). mean shift: a robust approach toward feature space analysis. ieee transactions on pattern analysis and machine intelligence, ( ), – . fu, h.-c., lai, p.s., lou, r.s., & pao, h.t. ( ). face detection and eye localization by neural network based color segmentation. in: neural networks for signal processing x, . proceedings of the ieee signal processing society workshop (vol. , pp. – ). guyon, i., & elisseeff, a. ( ). an introduction to variable and feature selection. the journal of machine learning research, , – . hastie, t., tibshirani, r., & friedman, j. ( a). the elements of statistical learning:data mining, inference, and prediction. springer. hastie, t., tibshirani, r., & friedman, j. ( b). the elements of statistical learning:data mining, inference, and prediction. springer. hastie, t., tibshirani, r., & friedman, j. ( c). the elements of statistical learning:data mining, inference, and prediction. springer. hastie, t., tibshirani, r., & friedman, j. ( d). the elements of statistical learning:data mining, inference, and prediction. springer. haykin, s. ( ). neural networks, a comprehensive foundation. macmillan college publishing company, inc.. lai, p.-s. & fu, h.-c. ( ). a polygon description based similarity measurement of stock market behavior. in: ieee congress on evolutionary computation, (cec ) singapore (pp. – ). lai, p.-s., & fu, h.-c. ( ). a polygon description based similarity measurement of stock market behavior. expert system with applications, ( ), – . liu, c. l. ( ). elements of discrete mathematics. new york: mcgraw-hill. ch. , p. . pao, h., chen, y., lai, p., xu, y., & fu, h.-c. ( ). constructing and application of multimedia tv-news archives. expert systems with applications, , – . pao, h., chuang, s., xu, y., & fu, h.-c. ( ). an em based multiple instance learning method for image classification. expert systems with applications, , – . schapire, r. e., & singer, y. ( ). improved boosting algorithms using confidence- rated predictions. machine learning, – . the demonstration of variance enhanced k-medoid model. available from http:// www.csie.nctu.edu.tw�pslai/askmddemo/. the test video. available from http://www.youtube.com/watch?v=mgpzuo_kbfy. http://www.csie.nctu.edu.tw~pslai/askmddemo/ http://www.csie.nctu.edu.tw~pslai/askmddemo/ http://www.csie.nctu.edu.tw~pslai/askmddemo/ http://www.youtube.com/watch?v=mgpzuo_kbfy variance enhanced k-medoid clustering introduction k-medoid the learning algorithm of k-medoid clustering self-growing k-medoid discussion variance enhanced k-medoid clustering polygon-shaped data model side-cluster segmentation projections on the center-to-center link decision boundaries adjustment variance enhanced k-medoid clustering method real-world application color clustering similarity measurement of two bitmaps similarity matching between bitmap groups images clustering by enhanced k-medoid method system demonstration conclusion references solid/liquid separation equipment simulation & design – an expert systems approach browse explore more content tarleton .pdf ( . kb) solid/liquid separation equipment simulation & design – an expert systems approach citedownload ( . kb)shareembed journal contribution posted on . . , : by richard j. wakeman, steve tarleton published texts on solid/liquid separation technology have allowed a limited amount of knowledge to be available widely, but none put forward the rules of thumb in such a manner that they are readily assimilable by the non-expert. there are many unpublished techniques and approaches. the computer technologist might target solid/liquid separation as a technology ripe for the application of expert systems. it is argued here that expert systems on their own are inadequate in this subject area, but the most effective software utilises a well-chosen mix of algorithm, graphics, expert system, and interactive input from the engineer. in this paper some examples are given of both public and private knowledge and an example of the application of the combined approach to equipment selection demonstrates that efficient software can save time and enable decisions to be made rapidly. equipment selection using the pc-select software is demonstrated. categories chemical engineering not elsewhere classified keywords untagged history school aeronautical, automotive, chemical and materials engineering department chemical engineering citation wakeman, r.j. and tarleton, e.s., . solid/liquid separation equipment simulation & design – an expert systems approach. filtration & separation, ( ), pp. - publisher © elsevier version am (accepted manuscript) publication date notes this article was published in the journal, filtration & separation [© elsevier] and the definitive version is available at: http://dx.doi.org/ . / - ( ) -o doi https://doi.org/ . / - ( ) -o issn - language en administrator link https://repository.lboro.ac.uk/account/articles/ licence cc by-nc-nd . exports select an optionrefworksbibtexref. managerendnotedatacitenlmdc categories chemical engineering not elsewhere classified keywords untagged licence cc by-nc-nd . exports select an optionrefworksbibtexref. managerendnotedatacitenlmdc hide footeraboutfeaturestoolsblogambassadorscontactfaqprivacy policycookie policyt&csaccessibility statementdisclaimersitemap figshare. credit for all your research. untitled the knowledge engineering review, vol. : , – . & cambridge university press, doi: . /s from expert systems to context-based intelligent assistant systems: a testimony p a t r i c k b r é z i l l o n lip , university pierre et marie curie - upmc, place jussieu, , paris, france e-mail: patrick.brezillon@lip .fr abstract this paper presents a personal interpretation of the evolution of artificial intelligence (ai) systems during these last years. this evolution is presented along five generations of ai systems, namely expert systems, joint cognitive systems, intelligent systems, intelligent assistant systems, and the coming generation of context-based intelligent assistant systems. our testimony relies on different real-world applications in different domains, especially for the french national power company, the subway companies in paris and in rio de janeiro, in medicine, a platform for e-maintenance, road safety, and open sources. our main claim is to underline that the next generation of ai systems (context-based intelligent assistant systems) requires a radically different consideration on context and its relations with the users, the task at hand, the situation, and the environment in which the task is accomplish by the user; the observation of users through their behaviors and not a profile library; a robust conceptual framework for modeling and managing context; and a computational tool for representing in a uniform way pieces of knowledge, of reasoning, and of contexts. introduction this paper is a testimony of years of research in artificial intelligence (ai). my initial interest for ai presented several facets. first, coming from the domain of mathematical modeling and having faced the limits of this approach in real-world applications, i was curious to study how ai tools (especially expert systems) could be a modeling tool for some problems in which human reasoning played a key role. second, ai belongs to cognitive sciences and has to deal with different disciplines such as cognitive psychology or neurosciences far from the engineering viewpoint. third, the initial paradigm of separation of the representation of the knowledge from its use was a powerful alternative to the usual approach in software engineering, at least in the design stage. fourth, ai has been (and is always) a battlefield where ‘hard’ science (i.e. theory-based science) was opposed to ‘soft’ science (i.e. cognition-based science). for example, the mathematical logic bloomed in a large variety of logics attempting each to deal with a specific aspect of cognition. fifth, ai is a domain where the success of an ai system is measured by its disappearance in classical software. the first concept that plays an important role in ai is knowledge, and more specifically its nature. education provides students with theoretical knowledge, when enterprises ask for their employees to have operational knowledge. for example, the ohm’s law is learned through instantiations where r has an exact value (say, . v) when real resistances have a value with a limited accuracy that may play a crucial role in an electronic circuit. the transformation of theoretical knowledge into operational knowledge is a process of contextualization during which we adapt our knowledge to the task, the situation, and the local environment. expertise is contextualized knowledge. a second point is the difference between data, information, and knowledge. information is data with meaning that results from a process of interpretation. for example, a temperature of c (the data) will be interpreted differently by a person living in the polar circle ( c is hot) and a person living near the equator ( c is cold). this is due to the fact that their mental models to interpret the data are different. relevant information enriches the person’s knowledge (the person can establishes a link between the information and his mental model). a third point concerns procedures and practices. organizations develop procedures that are collections of secure action sequences to address a given focus in any case. as a consequence, actors prefer to plan their action in real time rather than to rely on procedures for two main reasons. first, the procedure is never perfectly adapted to any specific situation and can lead to improper actions or sub-optimal incident resolution strategies. second, a practice captures advantages and disadvantages of the situation at hand when a procedure tries to capture only the abstracted solution. given a practical unit of reasoning: if p ;p ; . . . then c in a logical and theoretical reasoning (a procedure), the action leads to the conclusion c that is added to the base alone. a practical reasoning is an inductive probabilistic reasoning, the con- clusion cannot be detached from the premises, that is, take meaning out of premises: c is added to the base linked to p , p , etc. this means that knowledge is defined in a context of use that is lost in procedures because procedures are supposed to be applicable to a large class of problems when a practice is a contextualized version of the procedure in a specific context. as a side effect, this implicitly shows that there is not always a unique and optimal solution, but a set of best solutions in given contexts. it is a problem clearly identified in different disciplines: prescribed task versus effective task, task versus activity, logic of functioning versus logic of use. the difference between procedure and practice is a question of knowledge organization: parallel structures for procedures and sequential structures (and simplification of the domain knowledge from the user’s viewpoint) for practices. the role played by context between procedure and practice in real-world applications is, in my view, at the origin of successive mutations in ai that occurred during these past years. the time of expert systems the paradigm of separation of the representation of knowledge from its use, is born from the attack of ill-defined problems (expertise, heuristics, combinatorial, etc.) with no algorithmic solutions. this corresponds to a shift of the interest from procedural knowledge (algorithm, software, etc.) to declarative knowledge (fact base, knowledge base, rule engine), although expert systems face both types of knowledge. knowledge was represented in flat knowledge bases, easy to update without a need to recompile the software because an inference engine made the use of the knowledge removed from the representation. the rule base frame is employed to represent rules constraining potential courses of action, which are treated as non-negotiable (‘facts of nature’, prescriptions from authorities). however, this was a source of problem because part of the knowledge comes often from technical domains where knowledge organization and structure were not represented explicitly. this leads to the introduction of some implicit knowledge on the use of the knowledge. a first option concerned the introduction of screening clauses. figure presents an example of a rule in mycin. buchanan and shortliffe ( ) proposed to add screening clauses to the condition part of rules, which were then activated only in some kind of context. for example, the th clause in the rule above (taken from mycin) constrains the activation of the rule, but is not a clause for identifying the infection. clause acts to restrict the rule application to the context of adults (the implicit reason in the rule is that tetracycline gives a violet color to teeth that is a problem for children). another option was the introduction of rule packets as a structuring factor of rule bases. the reason was, say in equipment diagnosis as in figure , that one does not need to consider the knowledge on a protective relay when one diagnoses a circuit breaker. p . b r é z i l l o n here, it is the whole if part that acts as the context of a circuit breaker. the explicit use of context in such a rule offers also two advantages: ( ) the rule packet, although valid for all the circuit-breakers (about in this application for the french national power company, electricité de france), is triggered for the unique circuit breaker ‘$cb’ in the cut-off system ‘$cos’; and ( ) the triggering of this rule allows to inhibit the test of other equipment pieces (freeze ‘check_pw’ and ‘check_teac’) of the cut-off system. three lessons learned at the end of this phase are: ( ) the need of a user-centered approach, ( ) need to deal explicitly with context, and ( ) need of new methodologies different from classical engineering. the time of joint cognitive systems the expression ‘joint cognitive system’ was coined to stress the fact that neither the system nor the user is able to solve the problem at hand alone (woods, ). the emphasis was on the word ‘cognitive’. in a broad sense, this implies that the system and the user are able to understand each other (or more realistically make compatible their interpretations (karsenty and brézillon, )) during problem solving. this entails strong requirements about the design of the system as: the sharing of the cognitive representations, an agreement about the contextual knowledge of the problem solving, the ability to follow each others reasoning, the capability of exchanging expla- nations, the capability to acquire incrementally new knowledge and learn new practices from the other, and a sharing of the interaction control. a key point here is the user and system com- plementarity to exploit their agent asymmetry. the complementarity implies that the system may contain either some knowledge (model or data) or functions (e.g. database exploitation) that really complements the user’s skills. however, joint cognitive systems focused on the task at hand, when some tasks such as knowledge acquisition, practice learning and explanation are parts of the task at hand and managed by the system as well as the user, because you cannot anticipate or ‘design away’ all the problems that might arise during user–system interaction. thus, the system must know how to acquire knowledge incrementally when needed. this was not possible as long as context was not made explicit. if . the infection which requires therapy is meningitis, . only circumstantial evidence is available for this case, . the type of meningitis is bacterial, . the age of the patient is greater than years old, and . the patient is an alcoholic, then there is evidence that the organisms, which might be causing the infection, are diplococcus-pneumoniae (. ) or e.coli (. ) figure example of screening clause ! check_cb "test of the circuit breaker in the control system $name" > if failure freeze check_pw, check_teac if equipment_piece ($cos) := ($cb) , nature ($cb) := circuit_breaker . then call the rule packet 'circuit-breaker_diagnosis($cos, $cb)' figure a rule calling a rule packet from expert systems to context-based intelligent assistant systems the interest here is not on the probability of a branch but about the possibility to determine, as soon as possible, on which path the operator is, to determine what is the next action to undertake. in other words, each state of nature (a sequence of events is a state of nature) describes a state of the world, or using our words, a context for action. as many contextual elements may intervene in several scenarios (e.g. traffic activity, position of the next train), operators prefer to take them into account as soon as possible to get a general picture of the best path to choose. at this step, contextual knowledge is proceduralized and in the meantime, operators postpone actions in sequences like macro-actions. the proceduralized context corresponds to the needed knowledge for addressing the focus of the operators. the main objective is to eliminate event nodes and to use contextual information for the choice of the actions. previously, users were considered as data gatherers with a total lack of understanding of the system behavior (i.e. not on the expert knowledge in the machine, but the way in which the system used this knowledge). the system relies on a logic of functioning often far from the logic of use of users. thus, for completing or modifying knowledge in the system, users prefer to intervene superficially by adding new rules to control the firing of existing rules. as a consequence, the system rapidly becomes intractable. humans usually function in an anticipative mode; operators explicitly check hypotheses to eliminate discovery or surprise. an anticipatory system would use knowledge (as a model itself and the relevant part of its environment) about future states to decide what action to take at the moment. an intelligent system should predict what will probably happen and preadapt itself for the occurrence of a crucial or time-critical event. an anticipatory capability supposes that the system has a simulation component. the time of intelligent systems in frederik pohl’s ( ) science fiction novel, the age of pussyfoot, each person needs a mobile device—a ‘satisfactor’—to buy products, communicate with others, watch television, control the children, and even receive medications. in this future world, computing-centered humans cannot function in society without their satisfactors. this type of computing is now everywhere: mobile phones, intelligent houses, and soon communicating objects. however, such (new) systems pro- pose services, integrate few contextual cues (gps, weather, etc.) and a user model often coming from statistical consideration. however, it is not the right way because the line of reasoning of the user does not deal with the line of reasoning of the system. an intelligent system must present different properties like: > provide users with a first approximation of environmental trends and events, > point out useful information implicit in large volumes of data to alert users to sudden changes, > develop multiple scenarios and perspectives on a given line of action, > attract user attention to existing and emerging strategic issues, > support users in sharing and communicating their views and perspectives, > guide user attention to specific data and its interpretation in relation to particular issues. with such qualities, an intelligent system must be more considered as an intelligent assistant system, like a secretary with her boss, solving tactical and operational problems and preparing elements to her boss for political and strategic problems. for example, context-aware applications can be assimilated to intelligent systems. note that the role of the system with respect to the user is more precise than in joint cognitive systems. there is an evolution from the problem-solver view to the information-processing view. in the e-business arena, one speaks of the push model. the time of intelligent assistant systems an intelligent assistant system (ias) has two sides: the process side and the human side. an ias has to understand: (a) the real-world process, (b) the tasks at hand, and (c) the operator’s p . b r é z i l l o n behavior. the ias may learn (i.e. acquire knowledge) by observing the behavior of the operators and the process. this is an incremental knowledge acquisition process where knowledge is acquired when needed and, above all, in its context of use. only a kernel of knowledge has to be elicited directly, mainly to select the right formal representation for the knowledge. depending on the operator’s experience, this cooperation can take one of two forms: a waking state if there is no discrepancy between the operator’s actions and those of the system, and a cooperative state if a difference is noted relative to either the operator’s action or its possible consequences in the future. in the wake state, the ias must follow the operators’ actions on the process, observe the corresponding effect on the process’ behavior and compare it to a simulated behavior. the ias does not intervene, if there is no discrepancy between the observed and simulated behaviors of the process. when a problem is detected, the ias first waits to allow time for the operator to react and, only after a while, alert the operator about the problem. the main reason is to not disturb the operator in a critical situation. if the operator does not see the problem, then, the system enters a cooperative state. however, more frequently, it is the operator that triggers the cooperative state. as the operator’s assistant, the system must automatically solve minor problems and prepare elements for complex problems that the operator will have to solve. it must also benefit from operator interaction to learn new practices and acquire new knowledge when it faces failure. for example, when the user overrides a decision for various reasons, the assistant will not understand this action if they have incompatible internal representations. if the same decision must be made in altered circumstances (i.e. in other contexts), the assistant must know when it is out of its domain and again ask the user to acquire the current context of use for the decision. in the e-business arena, one speaks of the pull model. generally, there are different methods for executing a task, and the choice of a method relies heavily on contextual cues and operators’ experience. clearly, an ias must be designed and developed in a formalism providing a uniform representation of knowledge, reasoning, and context elements. moreover, the formalism must allow incremental knowledge acquisition and practice learning. thus, the human side begins to be addressed, but not yet the context side. the time of context-based intelligent assistant systems as discussed above, the importance of context was clear since a long time. however, it is at the beginning of the s that a community appeared in ai with mccarthy ( ) as a leader. he formalized context as first class objects and introduced the basic relation ist(c, p) (the proposition p is true in the context c), the formula being always asserted in a context. as a consequence, mccarthy showed that: > context is always relative to another context, > contexts have an infinite dimension, > contexts cannot be described completely, > when several contexts occur in a discussion, there is a common context above all of them in which terms and predicates can be lifted. now, a number of disciplines have a community of context, about context-aware applications, ubiquitous computing, pervasive computing, context-driven testing, etc. for our purpose in this paper, we retain that context is the key factor of intelligent assistant systems. making context explicit allows us to use knowledge in its context of use, to capture variants in the reasoning (e.g. recording practices effectively developed by operators), to generate relevant explanations. after years, there is a conceptual framework for context modeling and management and its implementation in a piece of software called contextual graphs. the working definition of context is that ‘context constrains a focus without intervening in it explicitly’. as a consequence, ( ) context is relative to the focus, ( ) as the focus evolves, its context evolves too, and ( ) context is highly domain-dependent. linking context to a focus, one can deduce that the focus divides from expert systems to context-based intelligent assistant systems context into external knowledge and contextual knowledge. the latter constitutes a kind of tank where contextual elements are related to the focus in a flat way, whereas the former has nothing to do with the focus at the moment. the focus evolves because a new event occurs (e.g. an unpre- dicted event) or as a result of a decision made at the previous stage of the focus. for addressing the focus, the actor selects a subset of contextual knowledge that is organized, assembled, and structured in a proceduralized context that is used at the current focus. a following step concerns how to represent knowledge in context with the notion of ‘focus dressing’, and a three-layer model of context (a meta-model, a domain model, and instantiation). a practice representing the development of an actor’s reasoning as a kind of contextualization of a procedure, there is a formalism that allows us to represent in a uniform way, the elements of knowledge and the context in which they are to be used in a reasoning. the formalism called contextual graphs has been used in several real-world applications in different domains. con- textual graphs are the building blocks of ‘experience bases’ on which an ias can reason and accompany a user in the realization of his task. conclusion evolution of ai systems comes from a move of the focus on the task alone (expert system time), through tasks and user (joint cognitive system time), after to the user accomplish a task in a situation (intelligent system time) to a holistic view on the user, the task, the situation, and the immediate environment with an emphasis on context. thus, this view of ai systems is one that travel smoothly from an engineering viewpoint to a cognitive science viewpoint in which context plays a central role in the modeling and the representation of the knowledge and the reasoning. a first consequence is that it is not possible to reason in an abstract way: context is deeply embedded in the actor, the task at hand, the situation and the environment, and context is intimately related to a focus of attention. a second consequence is that most of the communities interested in context management now prefer to study context by a bottom-up approach instead of the previous top–down approach. however, it is always important to work in a coherent conceptual frame- work, with an operational definition of context, a modeling of context linked to other elements (focus, task, user, etc.), and to use this solid framework for developing relevant tools for analysis knowledge and reasoning in context. we do this with contextual graphs for representing users’ behaviors (i.e. practices instead of procedures), but there are other aspects to study, like a description of a focus or a situation in terms of contextual elements. finally, the current approaches try to catch what humans are doing, but one will soon have to face some challenge like representing time in context and develop ai systems that are able to predict new behaviors. references buchanan, b. g. & shortliffe, e. h. . rule based expert systems: the mycin experiments of the stanford heuristic programming project. addison-wesley. karsenty, l. & brézillon, p. . cooperative problem solving and explanation. international journal of expert systems with applications ( ), – . mccarthy, j. . notes on formalizing context. in proceedings of the th international joint conference on artificial intelligence , – . pohl, f. . the age of pussyfoot. ballantine. woods, d. d. . cognitive technologies: the design of joint human-machine cognitive systems. ai magazine ( ), – . p . b r é z i l l o n neural network ensemble operators for time series forecasting kourentzes, n. , barrow, d. k. and crone, s. f. author post-print (accepted) deposited in curve january original citation & hyperlink: kourentzes, n. , barrow, d. k. and crone, s. f. ( ) neural network ensemble operators for time series forecasting. expert systems with applications, volume ( ): – http://dx.doi.org/ . /j.eswa. . . doi . /j.eswa. . . isbn - - - - publisher: elsevier copyright © and moral rights are retained by the author(s) and/ or other copyright owners. a copy can be downloaded for personal non-commercial research or study, without prior permission or charge. this item cannot be reproduced or quoted extensively from without first obtaining permission in writing from the copyright holder(s). the content must not be changed in any way or sold commercially in any format or medium without the formal permission of the copyright holders. this document is the author’s post-print version, incorporating any revisions agreed during the peer-review process. some differences between the published version and this version may remain and you are advised to consult the published version if you wish to cite from it. curve is the institutional repository for coventry university http://curve.coventry.ac.uk/open http://dx.doi.org/ . /j.eswa. . . http://curve.coventry.ac.uk/open neural network ensemble operators for time series forecasting nikolaos kourentzesa,∗, devon k. barrowa, sven f. cronea alancaster university management school department of management science, lancaster, la yx, uk abstract the combination of forecasts resulting from an ensemble of neural networks has been shown to outperform the use of a single “best” network model. this is supported by an extensive body of literature, which shows that combining generally leads to improvements in forecasting accuracy and robustness, and that using the mean operator often outperforms more complex methods of combining forecasts. this paper proposes a mode ensemble operator based on kernel density estimation, which unlike the mean operator is insensitive to outliers and deviations from normality, and unlike the median operator does not require symmetric distributions. the three operators are compared empirically and the proposed mode ensemble operator is found to produce the most accurate forecasts, followed by the median, while the mean has relatively poor performance. the findings suggest that the mode operator should be considered as an alternative to the mean and median operators in forecasting applications. experiments indicate that mode ensembles are useful in automating neural network models across a large number of time series, overcoming issues of uncertainty associated with data sampling, the stochasticity of neural network training and the distribution of the forecasts. keywords: time series, forecasting, ensembles, combination, mode estimation, kernel density estimation, neural networks, mean, median ∗correspondance: n kourentzes, department of management science, lancaster uni- versity management school, lancaster, lancashire, la yx, uk. tel.: + - - email address: n.kourentzes@lancaster.ac.uk (nikolaos kourentzes) preprint submitted to expert systems with applications december , . introduction with the continuing increase in computing power and availability of data, there has been a growing interest in the use artificial neural networks (nns) for forecasting purposes. nns are typically used as ensembles of several net- work models to deal with sampling and modelling uncertainties that may otherwise impair their forecasting accuracy and robustness. ensembles com- bine forecasts from the different models that comprise them. this paper proposes a new fundamental ensemble operator for neural networks that is based on estimating the mode of the forecast distribution, which has appeal- ing properties compared to established alternatives. although the use of ensembles is nowadays accepted as the norm in fore- casting with nns (crone et al., ), their performance is a function of how the individual forecasts are combined (stock and watson, ). improve- ments in the ensemble combination operators have direct impact on the re- sulting forecasting accuracy and the decision making that forecasts support. this has implications for multiple forecasting applications where nn ensem- bles have been used. some examples include diverse forecasting applications such as: economic modelling and policy making (mcadam and mcnelis, ; inoue and kilian, ), financial and commodities trading (zhang and berardi, ; chen and leung, ; versace et al., ; bodyanskiy and popov, ; yu et al., ), fast-moving consumer goods (trapero et al., ), tourism (pattie and snyder, ), electricity load (hippert et al., ; taylor and buizza, ), temperature and weather (roebber et al., ; langella et al., ), river flood (campolo et al., ) and hydrolog- ical modelling (dawson and wilby, ), climate (fildes and kourentzes, ), and ecology (araújo and new, ) to name a few. zhang et al. ( ) lists multiple other forecasting applications where they have been em- ployed successfully. as nn ensembles are fundamental for producing accurate nn forecasts for these various applications; hence, improvements in the construction of the ensembles are important. in this paper, the performance of the proposed mode operator is investigated together with the two existing fundamental ensemble operators: the mean and the median. two different datasets, in- cluding , real time series, are used to empirically evaluate the different operators. furthermore, ensembles of both training initialisations and sam- pling (bagging) are used to investigate the performance of the operators. the proposed operator is found to be superior to established alternatives. moreover, the robustness and good performance of the median operator is validated. the findings provide useful insights for the application of nns in large scale forecasting systems, where robustness and accuracy of the fore- casts are equally desirable. the rest of the paper is organised as follows: section discusses the benefits of nn ensembles and the limitations of the established ensemble operators. section introduces multilayer perceptrons that will be used for this paper and section discusses the three fundamental ensemble operators and presents the proposed method for mode ensembles. sections and discuss the experimental design and the results respectively, followed by a discussion of the findings in section . . forecasting with neural networks over the last two decades there has been substantial research in the use of nns for forecasting problems, with multiple successful applications (zhang et al., ). adya and collopy ( ) found that nns outperformed estab- lished statistical benchmarks in % of the papers reviewed. nns are flexible nonlinear data driven models that have attractive properties for forecasting. they have been proven to be universal approximators (hornik et al., ; hornik, ), being able to fit to any underlying data generating process. nns have been empirically shown to be able to forecast both linear (zhang, ) and nonlinear (zhang et al., ) time series of different forms. their attractive properties have led to the rise of several types of nns and appli- cations in the literature (for examples see connor et al., ; zhang et al., ; efendigil et al., ; khashei and bijari, ). while nns powerful approximation capabilities and self-adaptive data driven modelling approach allow them great flexibility in modelling time series data, it also complicates substantially model specification and the es- timation of their parameters. direct optimisation through conventional min- imisation of error is not possible under the multilayer architecture of nns and the back-propagation learning algorithm has been proposed to solve this problem (rumelhart et al., ), later discussed in the context of time series by werbos ( ). several complex training (optimisation) algorithms have appeared in the literature, which may nevertheless be stuck in local optima (hagan et al., ; haykin, ). to alleviate this problem, training of the networks may be initialised several times and the best network model selected according to some fitting criteria. however, this may still lead to suboptimal selection of parameters depending on the fitting criterion, result- ing in loss of predictive power in the out-of-sample set (hansen and salamon, ). another challenge in the parameter estimation of nns is due to the uncertainty associated with the training sample. breiman ( b) in his work on instability and stabilization in model selection showed that subset selection methods in regression, including artificial neural networks, are un- stable methods. given a data set and a collection of models, a method is defined as unstable if a small change in the data results in large changes in the set of models. these issues pose a series of challenges in selecting the most appropriate model for practical applications and currently no universal guidelines exist on how best to do this. in dealing with the first, the nn literature has strongly argued, with supporting empirical evidence, that instead of selecting a single nn that may be susceptible to poor initial values (or model setup), it is preferable to consider a combination of different nn models (hansen and salamon, ; zhang and berardi, ; versace et al., ; barrow et al., ; crone et al., ; ben taieb et al., ). naftaly et al. ( ) showed that ensembles across nn training initialisations of the same model can improve accuracy while removing the need for identifying and choosing the best training initialisation. this has been verified numerous times in the literature (for example see zhang and berardi, ). these ensembles aim at reducing the parameter uncertainty due to the stochasticity of the training of the networks. instead of relying on a single network that may be stuck to a local minima during its training, with poor forecasting performance, a combination of several networks is used. in the case of uncertainty about the training data, breiman ( a) proposed bagging (bootstrap aggregation and combination) for generating ensembles. the basic idea behind bagging is to train a model on permutations of the original sample and then combine the resulting models. the resulting ensemble is robust to small changes in the sample, alleviating this type of uncertainty. recent research has lead to a series of studies involving the application of the bagging algorithm for forecasting purposes with positive results in many application areas (inoue and kilian, ; lee and yang, ; chen and ren, ; hillebrand and medeiros, ; langella et al., ). apart from improving accuracy, using ensembles also avoids the problem of identifying and choosing the best trained network. in either case, neural network ensembles created from multiple initial- isations or from the application of the bagging algorithm, require the use of an ensemble combination operator. the forecast combination literature provides insights on how to best do this. bates and granger ( ) were amongst the first to show significant gains in forecasting accuracy through model combination. newbold and granger ( ) showed that a linear com- bination of univariate forecasts often outperformed individual models, while ming shi et al. ( ) provided similar evidence for nonlinear combinations. makridakis and winkler ( ) using simple averages concluded that the forecasting accuracy of the combined forecast improved, while the variabil- ity of accuracy among different combinations decreased as the number of methods in the average increased. the well known m competitions provided support to these results; model combination through averages improves ac- curacy (makridakis et al., ; makridakis and hibon, ). elliott and timmermann ( ) showed that the good performance of equally weighted model averages is connected to the mean squared error loss function and under varying conditions optimally weighted averages can lead to better ac- curacy. agnew ( ) found good accuracy of the median as an operator to combine forecasts. stock and watson ( ) considered simple averages, me- dians and trimmed averages of forecast, finding the average to be the most accurate, although one would expect the more robust median or trimmed mean to perform better. on the other hand, mcnees ( ) found no sig- nificant differences between the performance of the mean and the median. kourentzes et al. ( ) showed that combining models fitted on data sam- pled at different frequencies can achieve better forecasting accuracy at all short, medium and long term forecast horizons and found small differences in using either the mean or the median. there is a growing consensus that model combination has advantages over selecting a single model not only in terms of accuracy and error variability, but also simplifying model building and selection, and therefore the forecast- ing process as a whole. nonetheless, the question of how to best combine different models has not been resolved. in the literature there are many dif- ferent ensemble methods, often based on the fundamental operators of mean and median, in an unweighted or weighted fashion. barrow et al. ( ) ar- gued that the distribution of the forecasts involved in the calculation of the ensemble prediction may include outliers that may harm the performance of mean-based ensemble forecasts. therefore, they proposed removing such elements from the ensemble, demonstrating improved performance. jose and winkler ( ) using a similar argument advocated the use of trimmed and winsorised means. on the other hand, median based ensembles, are more robust to outliers and such special treatment may be unnecessary. however, the median, as a measure of central tendency is not robust to deviations from symmetric distributions. the median will merely calculate the middle value that separates the higher half from the lower half of the dataset, which is not guaranteed to describe well the location of the distribution of the forecasts that are used to construct the ensemble. taking a different perspective, ensembles provide an estimate of where most forecasts tend to be. mean and median are merely measures of the cen- tral tendency of the forecast distribution. in the case of normal distribution these coincide. outliers and deviations from normality harm the quality of the estimation. an apparent alternative, that in theory is free of this prob- lem, is the mode. this measure of central tendency has been overlooked in the combination literature because of its inherent difficulty in estimating it for unknown distributions. this paper exploits the properties of the mode to propose a new fundamental ensemble operator. in the following sections this operator is introduced and evaluated against established alternatives. . multilayer perceptrons the most commonly used form of nns for forecasting is the feedforward multilayer perceptron. the one-step ahead forecast ŷt+ is computed using inputs that are lagged observations of the time series or other explanatory variables. i denotes the number of inputs pi of the nn. their functional form is: ŷt+ = β + h ∑ h= βhg ( γ i + i ∑ i= γhipi ) . ( ) in eq. ( ), w = (β, γ) are the network weights with β = [β , . . . , βh], γ = [γ , . . . , γhi] for the output and the hidden layers respectively. the β and γ i are the biases of each neuron, which for each neuron act similarly to the intercept in a regression. h is the number of hidden nodes in the network and g(·) is a non-linear transfer function, which is usually either the sigmoid logistic or the hyperbolic tangent function. nns can model interactions between inputs, if any. the outputs of the hidden nodes are connected to an output node that produces the forecast. the output node is often linear as in eq. ( ). w w m s e − − − − . . . . figure : contour plot of the error surface of a neural network. the initial (⊕) and ending (•) weights for six different training initialisations are marked. in the time series forecasting context, neural networks can be perceived as equivalent to nonlinear autoregressive models (connor et al., ). lags of the time series, potentially together with lagged observations of explanatory variables, are used as inputs to the network. during training pairs of input vectors and targets are presented to the network. the network output is compared to the target and the resulting error is used to update the network weights. nn training is a complex nonlinear optimisation problem and the network can often get trapped in local minima of the error surface. in order to avoid poor quality results, training should be initialised several times with different random starting weights and biases to explore the error surface more fully. figure provides an example of an error surface of a very simple nn. the example network is tasked to model a time series with a simple autoregressive input and is of the form ŷt+ = g (w g (w yt− )), where g(·) is the hyperbolic tangent and w and w its weights. six different training initialisations, with their respective final weights, are shown. observe that minor differences in the starting weights can result in different estimates, even for such a simple model. in order to counter this uncertainty an ensemble of all trained networks can be used. as discussed before, this approach has been shown to be superior to choosing a single set of estimated weights. note that the objective of training is not to identify the global optimum. this would result in the model over-fitting to the training sample and would then generalise poorly to unseen data (bishop, ), in particular given their powerful approximation capabilities (hornik, ). furthermore, as new data become available, the prior global optimum may no longer be an optimum. in general, as the fitting sample changes, with the availability of new information, so do the final weights of the trained networks, even if the initial values of the network weights were kept constant. this sampling induced uncertainty can again be countered by using ensembles of models, following the concept of bagging. . ensemble operators let ŷmt be a forecast from model m for period t, where m = , . . . , m and m the number of available forecasts to be combined in an ensemble forecast ỹt. in this section the construction of ỹt using the mean, median and the proposed mode operators is discussed. to apply any of these operators reliably a unimodal distribution is assumed. . . mean ensemble the mean is one of the most commonly used measures of central tendency and can be weighted or unweighted. let wm be the weight for the forecasts from model m. conventionally ≤ wm ≤ and ∑m m= wm = . the ensemble forecast for period t is calculated as: ỹmeant = m ∑ m= wmymt. ( ) if all wm = m − the resulting combination is unweighted. the properties of the mean are well known, as well as its limitations. the mean is sensitive to outliers and unreliable for skewed distributions. to avoid some of its prob- lems one might use a winsorised or truncuted mean (jose and winkler, ). in this case the mean behaves more closely to the median. for distributions with finite variance, which is true for sets of forecasts, the maximum distance between the mean and the median is one standard deviation (mallows, ). . . median ensemble similarly the median can be unweighted or weighted, although the latter is rarely used. the median ensemble ỹmediant is simply calculated sorting wmymt and picking the middle value if m is odd or the mean of the two middle values otherwise. although the median is more robust than the mean it still suffers with non-symmetric distributions. . . mode ensemble the mode is defined as the most frequent value in a set of data. the mode is insensitive to outliers, in contrast to the mean and median. there is no formula to calculate the mode of an unknown distribution for continuous variables. there are two common ways to calculate it: either by discretising the data and identifying the most frequent bin, or by kernel density esti- mation. in this work the second approach is preferred in order to avoid the discretisation of the data. furthermore, kernel density estimation lends itself well to the continuous-valued nature of forecasts. kernel density estimation is a non-parametric way to estimate the proba- bility density function of a random variable, in this case the forecasts. given forecasts of a distribution with unknown density f, we can approximate its shape using the kernel density estimator f̂th(x) = (mh) − m ∑ m= k ( x − ŷmt h ) , ( ) where k(·) is a function with the property ∫ k(x)dx = that is called kernel and h > is its bandwidth. the kernel is often chosen to be a unimodal symmetric density function, therefore making f̂h(x) a density function itself, which is often, for computational reasons, the gaussian kernel φ(x): φh(x) = √ πh e − x h . ( ) figure shows an example of the calculation of kernel density. a kernel with bandwidth h is fitted around each observation and the resulting sum approximates the density function of the sample. a number of alternative kernel functions have been proposed in the litera- ture, however the choice of kernel has been found to have minimal impact on the outcome for most cases (wand and jones, ). the bandwidth of the kernel h controls the amount of smoothing. a high bandwidth results in more d e n si ty observations figure : example calculation of kernel density estimation. smoothing. therefore, the choice of h is crucial, as either under-smoothing or over-smoothing will provide misleading estimation of the density f (sil- verman, ). the approximation by silverman ( ) is often used in practice h = ( σ̂ m ) , ( ) where σ̂ is the standard deviation of the sample of the forecasts. this approx- imation is often adequate for gaussian kernels. botev et al. ( ) propose a methodology to automatically select the bandwidth that is free from the arbitrary normal reference rules used by existing methods. this is preferred in the calculation of the mode ensemble as the resulting bandwidth h allows fast convergence and good performance of the ensemble, as it is discussed in the results. the value x that corresponds to the maximum density approximates the mode of the true underlying distribution for a set of forecasts, which is also the value of the mode ensemble ŷmodet+h . this is true as long as the estimated distribution is unimodal. although the probability of facing non-unimodal distributions when dealing with forecasts is low, the following heuristic is proposed to resolve such cases. since there is no preference between the modes, the one closer to the previous (forecasted or actual) value is retained as the mode. this results in smooth trace forecasts. eq. ( ) results in unweigthed ŷmodet+h . it is trivial to introduce wm individual weights for each model. for kernel density estimation to adequately reveal the underlying density a relatively large number of observations are required. a small number of observations will lead to a bad approximation. this is illustrated in figure . it shows the mean, median, mode ensembles as well as the selected “best” model forecast, as selected using a validation sample for four different forecast horizons. furthermore, the estimated kerned density for each horizon is plotted. it is apparent by comparing a and b that the kernel density estimation using only models is very poor. while in a the shape of the distribution is successfully approximated, in b there are not enough forecasts to identify the underlying shape of the distribution of the forecasts. furthermore, in figure a it is easy to see that neither the mean, median or the “best” model are close to the most probable value of the forecast distribution. the mode ensemble offers an intuitive way of identifying where forecasts from different models converge and provide a robust forecast, independent of distributional assumptions. . empirical evaluation . . datasets to empirically evaluate the performance of the mean, median and the proposed mode ensemble for nns, two large datasets of real monthly time series are used. the first dataset comes from federal reserve economic data (fred) of st. luis. from the complete dataset , monthly time series that contain or more observations ( years) were sampled. long time series were preferred to allow for adequate training, validation and test sets. the second dataset comes from the uk office for national statistics and contains monthly retail sales time series. again, only time series with or more observations were retained for the empirical evaluation. a summary of the characteristics of the time series in each dataset is provided in table . to identify the presence of trend in a time series the cox- stuart test was employed on a -period centred moving average fitted to each time series. the test was performed on the centred moving average to smooth any effects from irregularities and seasonality. to identify the presence of seasonality, seasonal indices were calculated for the de-trended time series and then these were tested for significant deviations from each other by means of a friedman test. this procedure, based on non-parametric tests, is robust, the dataset can be accessed at http://research.stlouisfed.org/fred /. the dataset can be accessed at http://www.ons.gov.uk/ons/rel/rsi/ retail-sales/january- /tsd-retail-sales.html. (a) models (b) models figure : example of the distribution of nn forecasts of different number of models, as estimated by gaussian kernel density estimation, for the first four steps ahead. the forecasts by model selection, mean, median and mode ensembles are provided. however different tests may provide slightly different percentages to those in table . table : dataset characteristics. series length series patterns dataset series min mean max level trend season trend-season fred . % . % . % . % retail . % . % . % . % the last observations from each time series are withheld as test set. the prior observations are used as validation set to accommodate nns training. . . experimental design a number of nn ensemble models are fitted to each time series. two are based on mean, two on median and two on mode ensembles. hereafter, these are named nn-mean, nn-median and nn-mode respectively. all combina- tion operators are applied in their unweighted version, as the objective is to test their fundamental performance. in each pair of ensembles, the first is a training ensemble, combining multiple training initialisations and the sec- ond is based on bagging, as described by kunsch ( ). this moving block bootstrap samples the original time series while preserving the temporal and spatial covariance structure, as well as the serial correlation of the time series data. by assessing the operators using different types of ensembles we aim to assess the consistency of their performance. furthermore, different sizes of ensembles are evaluated, from members up to members, with steps of . results for single nn models, based on selecting the best one, are not provided as there is compeling evidence in the literature that ensembles are superior (for example see zhang and berardi, ; barrow et al., ). this was validated in our experiments as well. the individual neural networks have identical setup. following the sug- gestions of the literature, if trend is identified in a time series it is removed through first differencing (zhang and qi, ). the time series is then li- nearly scaled between - . and . to facilitate the nn training. the inputs are identified through means of stepwise regression, which has been shown to perform well for identifying univariate input lags for nns (crone and kourentzes, ; kourentzes and crone, ). all networks use the hy- perbolic tangent transfer function for the hidden nodes and a linear output node. the number of hidden nodes was identified experimentally for each time series. up to hidden nodes were evaluated for each time series and the number of hidden nodes that minimised the validation mean squared error (mse) was chosen. each network was trained using the levenberg-marquardt (lm) algo- rithm. the algorithm requires setting a scalar µlm and its increase and decrease steps. when the scalar is zero, the lm algorithm becomes just newton’s method, using the approximate hessian matrix. on the other hand, when µlm is large, it becomes gradient descent with a small step size. newton’s method is more accurate and faster near an error minimum, so the aim is to shift toward newton’s method as quickly as possible. if a step would increase the fitting error then µlm is increased. here µlm = − , with an increase factor of µinc = and a decrease factor of µdec = − . for a detailed description of the algorithm and its parameters see hagan et al. ( ). mse was used as the training cost function. the maximum training epochs are set to . the training can stop earlier if µlm becomes equal or greater than µmax = . the mse error at the validation set is tracked while training. if the error increases consequently for epochs then training is stopped. the weights that give the lowest validation error are selected at the end of each training. this is common practice in the literature and helps to achieve good out-of-sample performance, since it avoids over-fitting to the training sample (haykin, ). following the suggestions of the forecasting literature (adya and col- lopy, ) two statistical benchmarks are used in this study, namely the naive forecast (random walk) and exponential smoothing. this is done to assess the accuracy gains of using nns against established simpler statistical methods. the naive requires no parameterisation or setup, hence is used as a baseline that any more complex model should outperform. the appro- priate exponential smoothing model is selected for each time series, depend- ing on the presence of trend and/or seasonality using akaike’s information criterion. model parameters are identified by optimising the log-likelihood function (hyndman et al., , ). exponential smoothing was selected as a benchmark based on its widely demonstrated forecasting accuracy and robustness (makridakis and hibon, ; gardner, ) and will be named ets in this work. the use of these benchmarks can help establish the rel- ative performance of the nn models. in total, eight forecasting models are fitted to each time series, six nns and two statistical benchmarks. rolling trace forecasts of months are produced using each model. the rolling origin evaluation enables collecting a large sample of forecasts and their errors, while being robust to irregular forecast origins and outliers, thus providing reliable error measurements. based on the long test set, trace forecasts from t+ up to t+ months are collected for each time series. the reader is referred to tashman ( ) for a detailed description of the evaluation scheme and its advantages. the forecasting accuracy is assessed using the mean absolute scaled er- ror (mase). this is preferred due to its favourable statistical properties. mase is calculated for each trace forecast as: mase = m− ∑m j= |yj − ŷj| (n − )− ∑n r= |yr − yr− | , ( ) where yj and ŷj are the actual and forecasted value for j = , . . . , m out- of-sample observations. the denominator is the mean absolute error of the random walk in the fitting sample of n observations and is used to scale the error. mase, being a scaled error, permits summarising model performance across time series of different scale and units, which mean squared or absolute errors cannot do, and is less biased from errors like the mean absolute per- centage error and its symmetric equivalent. another advantage of this error is that it is very improbable that the denominator is zero, therefore making it easy to calculate in several scenarios and robust to time series with several values equal or close to zero (hyndman and koehler, ). note that the retail dataset contains several time series that do not permit the calculation of conventional percentage errors, due to zero values in the denominator. to summarise the results across the time series of each dataset the mean and median mase across all series are calculated. . results table presents the results for the fred time series. numbers in brack- ets refer to median mase, while the rest to mean mase. the table provides results for ensembles from to members. the results for bagging and training ensembles are presented separately to assess the impact of the ensem- ble type on the different ensemble operators. in each row the best performing method according to mean and median mase is highlighted in boldface. table : mean (median) mase for fred dataset. ensemble size nn-mean nn-median nn-mode naive ets bagging . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) training ensemble . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) overall, the difference between the mean and median mase results in- dicates that there are several difficult time series, particularly affecting the less robust mean mase. focusing on the bagging results, all nn-mean, nn- median and nn-mode are more accurate than the benchmarks when con- sidering mean mase. furthermore, as the ensembles increase in size their accuracy improves. in particular, for nn-mode after there are or more members the forecasts are very accurate. this was to be expected since the kernel density estimation becomes reliable once there is an adequate num- ber of observations, as discussed in section . for ensembles of or more members nn-mode provides consistently the best accuracy, closely followed by nn-median. note that achieving large numbers of ensemble members is trivial with nns, as this merely implies that more training initialisations or bootstrapped samples are used. therefore, the requirement of the mode op- erator for or more ensemble members is not a limiting factor. in contrast, nn-mean underperforms to the extent that ets is more accurate for median mase. this is an interesting finding, given how common is the mean oper- ator for ensembles in the literature. the more robust behaviour of median and the in-sensitive to outliers nature of the mode result in more accurate ensemble forecasts. looking at mean mase, all nns behave more robust than ets, the latter being severely affected by outliers. the results of the training ensembles are very similar. again, as the number of members in the ensemble increases nn-mode performs better and is the most accurate model for or more ensemble members. nn-median ranks second with small differences, while nn-mean is substantially worse. comparing the results between bagging and training ensembles we can see that the former is marginally more accurate for nn-median and nn-mode when mean mase is considered. however, the same is not true for nn-mean, indicating that the robustness and performance of this ensemble operator is affected by the type of ensemble. table presents the results for the retail dataset. its structure is the same as in table . the differences between mean and median mase are smaller than the fred results, showing that the time series in this dataset are better behaved. considering the bagging results, nn-median consistently outperforms the statistical benchmarks, while the same is true for nn-mode, once there is an adequate number of members in the ensemble (again or more). nn-mode is the most accurate model with the lowest mean and median mase. this is followed closely by nn-median. on the other hand, nn-mean often fails to outperform the benchmark ets, although it is always better than the naive. looking at the accuracy of the training ensembles nn-mode is overall more accurate for mean mase, nn-median is the most accurate for median mase. although all nn models outperform the naive benchmark, the dif- ferences between either nn-mode or nn-median and ets are very small. nn-mean is worse than ets in terms of mean mase, while occasionally it is marginally better in terms of median mase. comparing accuracies between bagging and training ensembles there are differences in favour of the former when looking at nn-median and nn-mode, while the accuracy for nn-mean is almost identical for both types of ensembles. across both datasets nn-mode and nn-median are the most accurate models. nn-mode seems to perform better when the size of ensemble is large enough. nn-median has slightly lower accuracy. while large ensem- bles benefit nn-median, it can perform well for small ensembles too. both table : mean (median) mase for retail dataset. ensemble size nn-mean nn-median nn-mode naive ets bagging . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) training ensemble . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) . ( . ) these models are on average more accurate than the statistical benchmarks ets and naive. on the other hand, nn-mean provides mixed results. in both datasets it outperforms naive, but not always ets. it is substantially outperformed by both nn-mode and nn-median. . discussion the value of ensembles for nns has been argued theoretically and demon- strated empirically. the combination of the models has often involved some type of mean operator. the empirical evaluation in this paper found that the less commonly used median operator and the proposed mode operator are more accurate and thus preferable. the size of the ensemble was found to be important for the accuracy of all operators. both mode and median, for the two datasets investigated here, seemed on average to converge for ensem- bles of or more members, with any additional members offering minimal changes in the forecasting performance. in particular, the mode, due to its reliance on kernel density estimation, required at least members. however, after that point it was found to be on average the most accurate ensemble op- erator. this is illustrated in figures and that present the mean mase for different number of ensemble members for the fred and the retail datasets respectively. the results for the different type of ensembles have been pooled together, since they had only small differences. note that there is little evi- dence that the mean ensembles had converged even with members. even larger ensembles were not calculated due to the substantial computational resources required, especially when the objective is to forecast a large num- ber of time series, which is common in supply chain and retailing forecasting applications. . . . . . . . . ensemble members m e a n m a s e fred dataset mean median mode figure : mean mase for different number of ensemble members for the fred dataset. in order to assess whether these differences are significant or not, we employ the testing methodology suggested by koning et al. ( ) that is appropriate for comparing forecasts from multiple models. the comparison is done across all different ensemble sizes to highlight if an operator is consis- tently statistically different. first, a freidman test is used to assess whether the accuracy of any model is significantly different from the rest. subse- quently, the mcb test is used to reveal the exact ordering of the different operators, and any significant differences between them. for both datasets the mode operator was significantly better than the median, which in turn was significantly different than the mean, at % significance level. at this point, it is useful to comment on the associated computational . . . . . . . . . ensemble members m e a n m a s e retail dataset mean median mode figure : mean mase for different number of ensemble members for the retail dataset cost of the nn ensembles. the main cost comes from training the networks. therefore, the more ensemble members that need to be trained, the less scalable forecasting with nns becomes, and the operator that achieves good forecasting performance with the least amount of members is preferable. table provides an overview of the average time required for forecasting across all series, for each dataset. as a different number of hidden nodes are used for each time series, the complexity of nn training changes, requiring different amount of time. to keep the presentation of the values simple, we summarise the training time over different series into the reported average time. the ensemble size that gave the minimum error for each operator in figures and is used as reference for the comparison. the average time in seconds, as well as the percentage difference over the time needed for the mode ensembles, are provided in the table. the networks were trained in parallel on an i - k cpu clocked at . ghz with logical cores. the mode operator needed the least number of ensemble members, re- quiring from % up to % less time than the mean or median operators across both datasets. therefore, apart from the significant gains in forecast- ing accuracy, the proposed ensemble operator required the least computa- tional resources. in particular for the retailing dataset, the run-time was more than halved. in figures and it is clear that similar performance is achieved for a large range of ensemble sizes for the median operator. this allows exchanging marginal differences in accuracy for smaller run-times, thus improving its scalability as well. on the other hand, this is not the case with the mean table : average computational time comparison. ensemble operator fred retail ensemble mean difference ensemble mean difference size time size time mean . secs + . % . secs + . % median . secs + . % . secs + . % mode . secs - . secs - operator, the accuracy of which improves with bigger ensembles. in the experiments, two types of ensembles were considered, bagging and training ensembles. each one tackles a different type of parameter uncer- tainty. we examined whether the performance of the operators was affected by the type of ensemble. again median and mode had very similar perfor- mance, favouring bagging. for the mean this behaviour was not evident. we attribute this different behaviour to the sensitivity of the mean to ex- treme values, which both median and mode are designed to avoid, albeit with different success. . conclusions this paper evaluates different fundamental ensemble operators. the well known mean and the less commonly used median were compared, together with a proposed mode operator that is based on kernel density estimation. all these three different operators attempt to describe the location of the distribution of the forecasts of the members of an ensemble. however, they deal with outlying extreme values differently, with the mean being the most sensitive and the mode the least. furthermore, distributional asymmetries can affect both the mean and the median, while the mode is immune. the findings in this paper suggest that both median and mode are very useful operators as they provided better accuracy than mean ensembles con- sistently across both datasets. the mode was found to be the most accurate, followed by the median. based on this finding, we recommend investigating the use of the mode and median operators further in ensembles research and applications, which have been largely overlooked in the literature that has mainly focused on the mean. furthermore, this work demonstrated that mode ensembles can robustly and accurately forecast automatically a large number of time series with neural networks, while the commonly used mean ensembles were often out- performed by exponential smoothing forecasts. moreover, mean ensembles required a very large number of members, which neither mode or median needed, with apparent implications for computational costs. in particular, the mode operator was found to require the least computation resources, due to the relatively small number of ensemble members that needed to be trained. we have already mentioned a number of applications that can benefit from improved nn ensemble forecasts, ranging from economic and business forecasting to climate modelling. most of these applications are characterised by forecasting a few, yet important, time series. the improved scalability of mode ensembles over the commonly used mean ensembles allows applying nns to areas that routinely require large scale automatic forecasting, which can benefit from the nonlinear modelling capabilities of nns. one such ex- ample is retailing, where one has to forecast a large number of products, the sales of which are affected by multiple factor that interact in a nonlin- ear fashion, such as pricing, promotional and temperature effects. the im- proved scalability of mode ensembles, compounded with the ever increasing computing capabilities provides opportunities for novel important forecasting applications of nn ensembles. this paper found significant savings in com- puting time from the proposed operator, which over the complete number of time series accounts for several hours of computations. such reduction will also help using nn ensembles in high frequency forecasting cycles, where the computational speed has been a limiting factor. future work should explore these potentials. the empirical evaluation, in this paper, focused on the unweighted version of all these operators, trying to assess their fundamental properties. although their differences are attributed to their robustness to extreme values, future research should extend this work to weighted versions of the operators. this will allow considering their use on further ensemble types, such as boosting. references adya, m., collopy, f., . how effective are neural networks at forecasting and prediction? a review and evaluation. journal of forecasting ( - ), – . agnew, c. e., . bayesian consensus forecasts of macroeconomic vari- ables. journal of forecasting ( ), – . araújo, m. b., new, m., . ensemble forecasting of species distributions. trends in ecology & evolution ( ), – . barrow, d., crone, s., kourentzes, n., . an evaluation of neural net- work ensembles and model selection for time series prediction. in: neural networks (ijcnn), the international joint conference on. ieee, pp. – . bates, j. m., granger, c. w. j., . the combination of forecasts. oper- ational research quarterly ( ), – . ben taieb, s., bontempi, g., atiya, a. f., sorjamaa, a., . a review and comparison of strategies for multi-step ahead time series forecasting based on the nn forecasting competition. expert systems with applications ( ), – . bishop, c. m., . neural networks for pattern recognition, st edition. oxford university press, usa. bodyanskiy, y., popov, s., . neural network approach to forecasting of quasiperiodic financial time series. european journal of operational research ( ), – . botev, z. i., grotowski, j. f., kroese, d. p., . kernel density estimation via diffusion. the annals of statistics ( ), – . breiman, l., a. bagging predictors. machine learning ( ), – . breiman, l., b. heuristics of instability and stabilization in model selec- tion. the annals of statistics ( ), – . campolo, m., andreussi, p., soldati, a., . river flood forecasting with a neural network model. water resources research ( ), – . chen, a.-s., leung, m. t., . regression neural network for error correc- tion in foreign exchange forecasting and trading. computers & operations research ( ), – . chen, t., ren, j., . bagging for gaussian process regression. neurocom- puting ( ), – . connor, j. t., martin, r. d., atlas, l. e., . recurrent neural networks and robust time series prediction. ieee transactions on neural networks , – . crone, s. f., hibon, m., nikolopoulos, k., . advances in forecasting with neural networks? empirical evidence from the nn competition on time series prediction. international journal of forecasting ( ), – . crone, s. f., kourentzes, n., . feature selection for time series prediction - a combined filter and wrapper approach for neural networks. neurocom- puting ( - ), – . dawson, c., wilby, r., . hydrological modelling using artificial neural networks. progress in physical geography ( ), – . efendigil, t., önüt, s., kahraman, c., . a decision support system for demand forecasting with artificial neural networks and neuro-fuzzy models: a comparative analysis. expert systems with applications ( ), – . elliott, g., timmermann, a., . optimal forecast combinations under general loss functions and forecast error distributions. journal of econo- metrics ( ), – . fildes, r., kourentzes, n., . validation and forecasting accuracy in mod- els of climate change. international journal of forecasting ( ), – . gardner, e. s., . exponential smoothing: the state of the art - part ii. international journal of forecasting ( ), – . hagan, m. t., demuth, h. b., beale, m. h., . neural network design. ma: pws publishing, boston. hansen, l. k., salamon, p., . neural network ensembles. ieee trans- actions pattern analysis and machine intelligence ( ), – . haykin, s., . neural networks and learning machines. pearson educa- tion, inc. hillebrand, e., medeiros, m. c., . the benefits of bagging for forecast models of realized volatility. econometric reviews ( - ), – . hippert, h. s., pedreira, c. e., souza, r. c., . neural networks for short- term load forecasting: a review and evaluation. power systems, ieee transactions on ( ), – . hornik, k., . approximation capabilities of multilayer feedforward net- works. neural networks ( ), – . hornik, k., stinchcombe, m., white, h., . multilayer feedforward net- works are universal approximators. neural networks ( ), – . hyndman, r. j., koehler, a. b., . another look at measures of forecast accuracy. international journal of forecasting , – . hyndman, r. j., koehler, a. b., ord, j. k., snyder, r. d., . forecasting with exponential smoothing: the state space approach. springer. hyndman, r. j., koehler, a. b., snyder, r. d., grose, s., . a state space framework for automatic forecasting using exponential smoothing methods. international journal of forecasting ( ), – . inoue, a., kilian, l., . how useful is bagging in forecasting economic time series? a case study of us consumer price inflation. journal of the american statistical association ( ), – . jose, v. r. r., winkler, r. l., . simple robust averages of forecasts: some empirical results. international journal of forecasting , – . khashei, m., bijari, m., . an artificial neural network (p, d, q) model for timeseries forecasting. expert systems with applications ( ), – . koning, a. j., franses, p. h., hibon, m., stekler, h. o., . the m competition: statistical tests of the results. international journal of fore- casting ( ), – . kourentzes, n., crone, s. f., . frequency independent automatic input variable selection for neural networks for forecasting. in: neural networks (ijcnn), the international joint conference on. ieee, pp. – . kourentzes, n., petropoulos, f., trapero, j. r., . improving forecasting by estimating time series structural components across multiple frequen- cies. international journal of forecasting. kunsch, h. r., . the jackknife and the bootstrap for general stationary observations. the annals of statistics ( ), – . langella, g., basile, a., bonfante, a., terribile, f., . high-resolution space–time rainfall analysis using integrated ann inference systems. journal of hydrology ( ), – . lee, t.-h., yang, y., . bagging binary and quantile predictors for time series. journal of econometrics ( ), – . makridakis, s., andersen, a., carbone, r., fildes, r., hibon, m., lewandowski, r., newton, j., parzen, e., winkler, r., . the ac- curacy of extrapolation (time series) methods: results of a forecasting competition. journal of forecasting ( ), – . makridakis, s., hibon, m., . the m -competition: results, conclusions and implications. international journal of forecasting ( ), – . makridakis, s., winkler, r. l., . averages of forecasts: some empirical results. management science ( ), – . mallows, c., . another comment on ocinneide. the american statisti- cian , . mcadam, p., mcnelis, p., . forecasting inflation with thick models and neural networks. economic modelling ( ), – . mcnees, s. k., . the uses and abuses of ‘consensus’ forecasts. journal of forecasting , – . ming shi, s., da xu, l., liu, b., . improving the accuracy of nonlinear combined forecasting using neural networks. expert systems with appli- cations ( ), – . naftaly, u., intrator, n., horn, d., . optimal ensemble averaging of neural networks. network: computation in neural systems , – . newbold, p., granger, c. w. j., . experience with forecasting univariate time series and the combination of forecasts. journal of the royal statis- tical society. series a (general) ( ), – . pattie, d. c., snyder, j., . using a neural network to forecast visitor behavior. annals of tourism research ( ), – . roebber, p. j., butt, m. r., reinke, s. j., grafenauer, t. j., . real-time forecasting of snowfall using a neural network. weather and forecasting ( ), – . rumelhart, d. e., hinton, g. e., williams, r. j., . learning internal rep- resentations by error propagation. in: rumelhart, d. e., mcclelland, j. l. (eds.), parallel distributed processing: explorations in the microstructure of cognition. vol. . mit press, cambridge, ma, usa, ch. learning in- ternal representations by error propagation, pp. – . silverman, b., . density estimation for statistics and data analysis. chapman & hall/crc, london. silverman, b. w., . using kernel density estimates to investigate multi- modality. journal of the royal statistical society. series b (methodologi- cal) ( ), – . stock, j. h., watson, m. w., . combination forecasts of output growth in a seven-country data set. journal of forecasting ( ), – . tashman, l. j., . out-of-sample tests of forecasting accuracy: an ana- lysis and review. international journal of forecasting ( ), – . taylor, j. w., buizza, r., . neural network load forecasting with weather ensemble predictions. power systems, ieee transactions on ( ), – . trapero, j. r., kourentzes, n., fildes, r., . impact of information ex- change on supplier forecasting performance. omega ( ), – . versace, m., bhatt, r., hinds, o., shiffer, m., . predicting the exchange traded fund dia with a combination of genetic algorithms and neural networks. expert systems with applications ( ), – . wand, m. p., jones, m. c., . kernel smoothing. chapman & hall. werbos, p. j., . backpropagation through time - what it does and how to do it. proceedings of the ieee ( ), – . yu, l., wang, s., lai, k. k., . forecasting crude oil price with an emd- based neural network ensemble learning paradigm. energy economics ( ), – . zhang, g., patuwo, b. e., hu, m. y., . forecasting with artificial neural networks: the state of the art. international journal of forecasting ( ), – . zhang, g. p., . an investigation of neural networks for linear time-series forecasting. computers and operations research ( ), – . zhang, g. p., berardi, v. l., . time series forecasting with neural net- work ensembles: an application for exchange rate prediction. journal of the operational research society , – . zhang, g. p., patuwo, b. e., hu, m. y., . a simulation study of artifi- cial neural networks for nonlinear time-series forecasting. computers and operations research ( ), – . zhang, g. p., qi, m., . neural network forecasting for seasonal and trend time series. european journal of operational research ( ), – . neural cover neural towards expert systems for the selection of search keys towards expert systems for the selection of search keys* raya fidel graduate school of library and lnforma tion science, university of washington, seattle, wa intermediary expert systems are designed to mediate between end-users and complex information retrieval systems. however, since most of these expert systems are based on text analysis rather than on models of hum man searching, they cannot process requestrelated cri- teria, such as precision or recall requirements. analysis of the searching behavior of human intermediaries re- vealed a routine for the selection of search keys-free- text or controlled vocabulary-along a decision tree. examples of decision rules demonstrate that although further research is required, these rules can be auto- mated to significantly enhance the adaptability of inter. mediary expert systems. introduction it is believed that end-users will very likely search their own requests online when search processes are simplified or made friendlier. the prevailing approach to providing easier and friendlier user-system communication is to de- velop “interface” or “intermediary” systems. indeed, systems such as cite [l] are already available for public access, and others, such as conit [ ] or cansearch [ ], are being tested in experimental settings. through such systems, users are freed from encoun- ters with many peculiarities in databases and search sys- tems, and yet can benefit from a large range of capabili- ties. in particular, users can enter a request in a loosely structured format, preferably in a natural language, sen- tence-like expression. an intermediary system processes the request terms, displays information to users, and asks for some sort of feedback. the information displayed may be in the form of a list of subject areas, databases, search keys (i.e., strings of characters that are entered to *this study was supported by the graduate school of the university of washington. received march i, ; revised may , ; accepted june , be searched for occurrence in pre-defined fields), or ac- tual citations from which users are asked to make a selec- tion, possibly in ranked order. interactions of this nature usually proceed until users terminate the session. some intermediary systems are actually helper sys- tems: they support shortcuts in end-user training by pro- viding menu-driven interaction or online help programs. typically, helper systems require end-users to make most of the decisions during a search process, or they drasti- cally simplify searching by reducing the number of op- tions to a minimum. intermediary expert systems, on the other hand, can be quite powerful. expert systems replicate the perfor- mance of an expert in a particular area by incorporating the knowledge of an expert with rules for making infer- ences on the basis of this knowledge. systems knowledge may or may not be intended to model actual processes as they would be performed by human experts. this article illustrates how to model searching behav- ior of human intermediaries by demonstrating that for- mal rules for the selection of search keys can be extracted from human experts. these rules cannot be incorporated yet into a knowledge base of an intermediary expert sys- tem because they are incomplete and were derived from searching behavior that is limited in its subject area. the work presented here, however, clearly indicates that with more research a complete set of rules can be established. it also provides guidance for future exploration and points to various issues that could be readily considered by designers of intermediary expert systems. modeling the selection of search keys search processes consist of three basic intellectual components: ( ) definition of query structure; ( ) selec- tion of search keys; and ( ) evaluation of feedback. we focus here on the second component. in a database that journal of the american society for information science. ( ): - , ccc - / / - $ . offers the capability, an intermediary system must exam- ine each term of a request and consider its representa- tion: as a controlled vocabulary key, as a free-text key, or as both. expert systems vary in the degree of freedom they provide users in this selection: some dutifully search only those search keys designated by a user; some use search keys designated by a user to generate additional search keys; and some automatically generate search keys without user participation. intermediary expert systems use mapping algorithms to generate search keys. whether user support and in- volvement are high or low, a term may be mapped to a descriptor (a search key from a controlled vocabulary) through an exact or other kinds of match, or it may not be mapped to a descriptor at all. while some systems, such as cansearch [ ], use subject-specific ap- proaches, most existing intermediary systems, expert and helper, use mapping algorithms that are based on text characteristics, such as word-occurrence frequency or statistical associations. the most apparent drawback of text analysis al- gorithms is their lack of sensitivity to request-specific re- quirements that cannot be directly deduced from a re- quest statement. consider, for example, a request about “the attitudes of anorexic students toward themselves during examinations periods.” one user is interested only in anorexic students, another wants to get all the in- formation available on the topic but is primarily inter- ested in students and is willing to look at material about student behavior during examinations periods. the sys- tem decides, say, to use anorexic students as a single search key, but may or may not suggest the term students as a search key, following its own algorithm. however, when experienced online searchers select search keys, they examine not only the degree of term- descriptor match, but they also consider other factors. moreover, these additional factors are quite frequently essential to the success of a search. one wonders why the experience of human intermediaries has been neglected by system designers when it is an important source of knowledge. there may be two explanations. first, online search- ing behavior was not being investigated and thus could not contribute knowledge. second, research and develop- ment in information storage and retrieval is familiar with text analysis because of the long and established experi- ence in automated indexing. although methods and ap- proaches to automated indexing vary greatly, most of them rely on text analysis and are thus text-oriented rather then user-oriented. only recently, in an experi- ment at the american petroleum institute, a first attempt has been made to develop an automated indexing system which models indexing behavior [ ]. as a first step toward knowledge engineering in inter- mediary expert systems, i analyzed the online searching behavior of several human intermediaries and found that online searchers do indeed follow some rules [s]. in their selection of search keys, they utilize informal and some- times highly intuitive decision rules. moreover, these rules can be detected, examined, and presented in a for- mal structure which can be processed by computers. this formal presentation incorporates knowledge of multiple experts and with further research can be used in a knowl- edge base for intermediary expert systems. the study method to examine online searching behavior, eight searchers were observed performing their regular, job-related searching [ ]. searchers who have been searching for more than two years were recruited from among informa- tion specialists in scientific areas, primarily in the life sci- ences. they were studied one at a time, and were asked to verbalize their thought processes during their searching to the degree that speaking out loud would not interfere with their performance. these verbalizations, including the creation of search strategy, were recorded. at the end of the observation period (approximately to searches), each searcher was interviewed to reveal and clarify information not accessible to observation. data collected for analysis were about one hundred printed search protocols with transcriptions of verbalized thought processes and additional explanations from the final interviews. each instance in which searchers had selected a search key was then identified and the reason for the specific choice was explicitly noted. analysis of the first ten search protocols generated a preliminary list of condi- tions under which a particular selection was made. for example, a condition for choosing a free-text key is: to enter straightforwardly a specific concept which might not be a trustworthy descriptor. all the search protocols were then analyzed against this preliminary list of conditions. each instance in which searchers had selected a search key was listed under the condition to which it applied. instances whose condition could not be found suggested a new condition to be added to the list. this analysis revealed that most conditions were considered by most searchers, and only a few combi- nations reflect searchers’ individual idiosyncrasies. the list of conditions for the selection of search keys is presented in figure in the form of a decision tree. this set of decision rules is called here “the selection routine.” the selection routine specifies conditions which are necessary for a searcher to be able to select a particular type of search key. it describes the most commonly se- lected path, but there might be complications. as such, the selection routine is not deterministic: it cannot always accurately predict the selection of search keys unless other factors and their impact are known. this routine groups together similar conditions so it could be pre- sented in a decision tree, but it is not meant to represent a necessary sequence in the selection process. a list of the options in search key selection and the conditions necessary for each option is given in table . journal of the american society for information science-january journal of the american society for information science-january table . a list of options and the associated conditions. the selection routine option conditions use descriptors add the next broader descriptor in the hierarchy use generic descriptors in an inclusive mode limit to retrieval by descriptors limit to major descriptors specify document type use free-text terms use free-text terms to probe indexing use descriptors as free- text terms in other databases use free-text terms for an inclusive search use free-text terms to introduce uncommon types of search keys descriptor searching a term is a common term + it is mapped to a descriptor [a]. a term is a single meaning term + it is mapped to a descriptor + the descriptor is an exact match [d]. the concept has many synonyms [e ]. the concept is not clear to the searcher e . the concept may not be explicitly mentioned [e ]. the descriptor is a partial match [f]. the descriptor is a broader term [.i]. a term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [e ]. a term is a single-meaning term + it is mapped to a descriptor -i- recall needs to be improved [e ]. a term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [e ]. a term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [elo]. a term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [e ]. free-text searching a term is a single-meaning + it is mapped to a descriptor + the concept is not “trustworthy” as an index term [el]. the descriptor is a broader term [hi. it cannot be mapped to a descriptor [k]. a term is a common term + it cannot be mapped to a descriptor [b]. a term is a single-meaning term + it cannot be mapped to a descriptor [l]. a term is a single-meaning term + it is mapped to a descriptor + a request needs to be searched on several databases [e ]. a term is a single-meaning term + it is mapped to a descriptor + the descriptor is a partial match [cl. a term is a single-meaning term + it cannot be mapped to a descriptor [ml. other combinations use free-text terms in combination with descriptors add free-text synonyms to descriptors add truncated free-text terms to descriptors add role indicators a term is a single-meaning term + it is mapped to a descriptor + the descriptor is a broader term [i]. a term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [es]. a term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [e ]. a term is a single-meaning term + it is mapped to a descriptor + change database precision needs to be improved [el i]. a term is a common term + it cannot be mapped to a descriptor [cl. before searchers decide how to represent a request term in a query formulation they must answer two central questions: (a) can a term be mapped to a descriptor, and (b) is it a “good” term for free-text retrieval. a searcher maps a term to a descriptor when she/he has decided that a particular descriptor best represents a request term, whether or not there is an exact match between the term and the descriptor. the second question is a little more complex and re- quires some explanation. searchers consider a term to be a “good” term for free-text searching if it: ( ) usually oc- curs in a particular context, ( ) is uniquely defined, and ( ) is specific in the concept it represents. such a term will be called here a “single-meaning” term. on the other hand, a term that occurs in more than one context will be called a “common” term. for example, in the request about the attitude of anorexic students toward them- selves during examinations periods, terms such as ano- rexia and students are single-meaning terms. the term examination, on the other hand, is a common term; it can occur in a subject-related context (“the best way to take student examinations”), or in a descriptive capacity (“ex- amination of students’ responses”), or still further, it can be used very loosely to represent the concept of an inquiry of any kind. when a term is a common term, searchers do not have much choice in the selection of search keys: if it can be mapped to a descriptor, searchers almost always enter the descriptor as the search key [a] (i.e., option [a] in figure . and table .) because, by definition, it is not desirable to use a common term as a free-text search key. a common term that cannot be mapped to a descrip- tor almost always results in unsatisfactory retrieval. searchers have no choice but to enter a free-text key, preferably in combination with other search keys, in or- der to retrieve citations, select some relevant ones and re- view their indexing in an attempt to find descriptors that might possibly be relevant [b]. for example, if the term examination cannot be mapped to a descriptor, one can devise a formulation (using the and operator) that com- bines the terms students, anorexia, and the free-text term examination. reviewing a sample of retrieved cita- tions, one may find that all the relevant citations include the descriptor instructional tests in their indexing, thus suggesting that this descriptor is an appropriate choice for the representation of the concept examinations. such probing does not always further a search and searchers may then decide to select a different database: one which does allow the common term to be mapped to a descriptor [cl. single-meaning terms provide more options. if a sin- gle-meaning term cannot be mapped to a descriptor, searchers may enter a free-text term as the only search key [k], but they may also probe indexing of relevant ci- tations to make sure that no adequate descriptor is over- looked [l]. journal of the american society for information science-january in some cases, searchers may use a free-text key to search for a single-meaning term that cannot be mapped to a descriptor in a special way: they require that it occurs in a field other than the common ones, such as the jour- nal title field [ml. suppose a user is interested only in the psychological aspects of anorexic students, and suppose that the term psychology cannot be mapped to a descrip- tor. searchers may predict that searching for the occur- rence of psychology in the text would retrieve a large number of irrelevant citations, and decide instead to re- trieve citations to articles whose authors are affiliated with organizations which include the stem psych in their titles, or articles that were published in sources whose ti- tles include this stem. the least problematic term is one that is single-mean- ing and also can be mapped to a descriptor. such a term can be entered either as a controlled vocabulary, as a free-text key, or as both. here, searchers are free to deal with other factors when selecting search keys. it is useful to show the conditions under which searchers select free- text keys, and those under which they choose to use de- scriptors. selection of free-text search keys a single-meaning term can be mapped to a descriptor through an exact match, partial match, or a term might be mapped to a broader descriptor. when a single-mean- ing term is mapped to a descriptor through an exact match, searchers may use a descriptor [d], or elect to consider a variety of factors [el. partial match usually implies mapping a request term to a narrower descriptor, or to a group of narrower de- scriptors. if suitable, searchers use a free-text key to in- clusively search concepts that are not grouped together by the hierarchy of the controlled vocabulary [g]. if, for ex- ample, the request term students is mapped to descrip- tors such as foreign students, college students, or un- dergraduates, and a descriptor students does not exist, the free-text key can be used to retrieve information about any type of student. it should be noted that in some search systems use of the free-text key students also would retrieve citations that are indexed with descriptors which include the term. this is a source for constant con- fusion for searchers because the routine changes from one search system to another. when a single-meaning request term is mapped to a broader descriptor, searchers may prefer to preserve the specificity of the request and use free-text search keys [hi. if they are concerned with the precision of the set to be retrieved, they enter free-text terms in combination (using the and operator) with the broader descriptor to which it is mapped [i]. regardless of the degree of match between a single- meaning term and a descriptor, searchers may still prefer to use free-text search keys for three reasons. first, if re- call needs to be improved, searchers use both descriptors and free-text terms as search keys [es]. for a further in- crease in recall, free-text keys are entered in a truncated form [e ]. second, if searchers think that a particular descriptor may be assigned inconsistently by indexers, they consider the use of a free-text key to be more trust- worthy [el]. suppose, for instance, a controlled vocabu- lary includes the descriptor nutrition and also the de- scriptor diet-each one to be assigned for distinct representation of a subject. when looking for nutrition- related problems of anorexic students, searchers may find the distinction confusing and thus assume that in- dexers are likely to be inconsistent in assigning these de- scriptors. to compensate for indexers’ errors, they may use both a descriptor and free-text keys, or only free-text keys. lastly, if a request is to be searched on several data- bases, searchers may map single-meaning terms to de- scriptors in only a few of the relevant controlled vocabu- laries. running the same query formulation against several databases, they then, in fact, search some of the terms as free-text keys in some of the databases [e ]. selection of controlled vocabulary search keys the most straightforward use of a descriptor to repre- sent a single-meaning term is when a term is exactly matched with a descriptor and no other apparent con- straints exist. however, searchers may elect to enter a re- quest term as a descriptor when it is mapped to a descrip- tor through a partial match [f], in which case it is mapped to a narrower descriptor, or when the term is mapped to a broader descriptor [.i]. such decisions de- pend on the nature of the request, and when searchers suspect that precision might not be satisfactory, they may combine free-text terms with a descriptor [i]. even single-meaning terms may have some attributes that will make them unattractive for free-text searching. regardless of the degree or nature of a term-descriptor match, searchers most often prefer to enter a descriptor when: ( ) a term has many synonyms [e ]; ( ) a concept and its use is not clear to a searcher [e ]; or ( ) when a concept is likely to be implied rather then explicitly men- tioned in the searched text [e ]. the previously mentioned request about anorexic stu- dents provides a clear example of the last condition. the request concept attitudes toward themselves can indeed be entered and searched as a free-text phrase in most search systems. however, this concept can be expressed, directly or indirectly, in various other phrases, such as “students displayed attacks of self-hatred”, or “narcis- sism level dropped with time.” searchers, then, consider descriptors such as self image, or se&esteem to be more reliable than free-text terms for searching. in addition to providing search keys for “problematic” terms, controlled vocabularies provide special means to improve precision and recall. when searchers perceive precision and/or recall to be unsatisfactory, they may de- cide to take advantage of these means and elect to use a descriptor to represent a single-meaning term. although journal of the american society for information science-january these routines are quite well known, it might be helpful to mention them. when searchers decide that recall needs to be improved, they select a controlled vocabulary key because they can add the next broader descriptor in the hierarchy [e ], or use generic concepts in an inclusive mode [e ]. controlled vocabularies readily suggest broader descriptors, and thus make it convenient to in- deed broaden a concept. moreover, broadening the meaning of a concept is not always possible in free-text searching since the broader concept may be a common term. inclusive searching-which is quite straightfor- ward in descriptor searching-facilitates retrieval of cita- tions indexed under a descriptor as well as those indexed under its narrower terms. lastly, when searchers predict that precision may not be satisfactory, they may select a controlled vocabulary key because they can exercise various ways of adding weights to search keys such as: to limit to retrieval by de- scriptor only [e ]; to limit a descriptor to be a major de- scriptor [elo]; to add role indicators [eli]; or to specify document type [e ]. discussion the selection routine clearly shows that the process of selecting search keys as performed by online searchers can be formalized into a decision tree. moreover, several suggestions for improvements in existing systems are im- mediately apparent; others will require more research. first, the pattern of the selection routine illustrates the significance of decisions made during search key selec- tion to the success of a search. this pattern shows that when a term is “not adequate” for searching, i.e., it is a common term and/or it cannot be mapped to a descrip- tor, only a few options are available for searching. only six of the twenty-five options in the selection routine are suggested for such terms, and some, such as the use of free-text terms to probe indexing, require a fair amount of creativity on the part of an intermediary. on the other hand, when a term is “good” for searching, i.e., it is a single-meaning term and it can be mapped to a descrip- tor, intermediaries can look at several options, as pre- sented in the check-list. in other words, only after termi- nological difficulties or peculiarities in representing request terms have been overcome for searching, can an intermediary consider additional factors that are essen- tial to the success of the search. therefore, intermediary expert systems must be able to resolve terminological dif- ficulties before they can be equipped to deal with addi- tional factors. second, using the selection routine, one can identify flaws in existing intermediary expert systems and at the same time propose methods to overcome these failings. while some flaws can be readily identified and possible remedies suggested already at this time, other issues re- quire additional research before their nature can be clearly defined. to demonstrate the ability of the selec- tion routine to illuminate such issues, a few examples are discussed below. one of the flaws in existing intermediary expert sys- tems that readily stands out is their inability to distin- guish between common and single-meaning terms. this distinction is important because if a term is a common term, experienced searchers almost always select it as a descriptor even if they have to change databases (unless they use it in a trial [b]). yet, to my knowledge, no exist- ing system, provides safeguards to advise end-users against the use of common terms as free-text search keys. for example, when asked about “information retrieval, ” cite suggested information systems, information ser- vices, information theory, and information centers, as descriptors, and retrieval, retrieving, and information as free-text search keys [ . experienced searchers normally will avoid using these common terms as free-text keys, though end-users may prefer them because none of the descriptors exactly matches the original concept. the idea that some terms are not suitable for auto- mated processing is not new. in linguistics, homonymy and polysemy are specific cases of common terms which might be better described as ambiguous terms or terms that belong to more than one semantic domain. these concepts are essential to thesaurus construction. from the information science field, fugmann, for example, differentiates between “individual” and “general” con- cepts, the latter being non-lexical concepts which are bet- ter searched with controlled vocabulary [ ]. in addition, the assumption that some terms carry more information than others is a fundamental premise in automated in- dexing and abstracting. control over common terms should require relatively modest modifications in intermediary expert systems. first, we have to devise a working definition of what con- stitutes a common term. for this purpose, it would be useful to test the hypothesis that terms which occur with high frequency in a database are also common terms. suppose, however, that a system selects to define a com- mon term by, say, a consensus among three knowledge- able searchers who are highly experienced with a data- base. these searchers can then check each term in a thesaurus, including those that represent an entry or part of an entry and those that occur in a descriptor or in a lead-in entry, and determine which terms are common. common terms can then be coded so they are not dis- played or used as free-text search keys. this method requires additional effort to identify com- mon terms that do not occur in a thesaurus or in other semantic networks. some shortcuts can be devised, how- ever, such as the use of a number of thesauri. on the other hand, other methods to identify common terms may apply to all terms in a database whether or not they are listed in a thesaurus. for instance, a test can be con- ducted to discover *whether a correlation exists between the frequency in which a word occurs in a database and its adequacy for free-text searching. if a well-selected journal of the american society for information science-january sample of databases proves that common terms occur with high frequency in those databases and vice versa, then common terms can be singled out by frequency counts. once a common term is defined, an intermediary sys- tem can interact with users. suppose a common term cannot be mapped to a descriptor. by giving messages to users, a system can interrogate them to determine whether to select another database or whether to enter the term as a free-text search key to probe indexing. this interaction may reveal request-related requirements that are not reflected in a request statement, e.g., that the term is central to the request or that it always should be associated with another term. a second example of a flaw in existing intermediary expert systems is their failure to suggest the use of free- text terms for inclusive search [g]. as explained earlier, if the descriptor students does not exist and a term is then mapped to descriptors such as university students or gifted students through partial match, the free-text term students can be entered to search for any kind of student. this function can be easily automated. how- ever, one should be careful because some terms are not suitable to be entered as free-text keys for inclusive searching. the term attitudes is a case in point. if an exact match does not exist, the term attitudes can be mapped through a partial match to descriptors such as employee atti- tudes, mother attitudes, or negative attitudes. in searching for material about attitudes of students toward themselves, a user may find the last descriptor relevant, but the first two are a source for unwanted retrieval. to search the term attitude as a free-text key would magnify the problem. in this case, the user is better advised to scan all the descriptors to which the term is mapped and to select the relevant ones. thus, we can designate for each word in a multi-words descriptor whether it can be automatically searched as a free-text key or whether it should be displayed to users when a partial match occurs. at this time, we do not have any scale based on systematic investigations that can de- termine which terms are suitable for inclusive searching in a free-text mode. we can, however, adapt a pragmatic approach and determine the status of each term by con- sulting with experienced online searchers. in the future, research in terminology may provide more rigorous crite- ria. a third example is the inadequacy of intermediary ex- pert systems for term analysis processes. various statisti- cal approaches could be used to determine which attrib- utes of terms are significant for searching. for instance, the “trustworthiness” of a descriptor can be measured by the degree of consistency with which it is assigned. an extrapolation of the measurement of term consistency suggested by king & bryant [ ] could be used to measure trustworthiness of descriptors. a test database could be indexed by several indexers. a measurement for trust- worthiness could then be determined by the ratio between the number of documents to which a descriptor has been assigned by all indexers and the total number of docu- ments to which it has been assigned by any number of indexers (possibly weighted by the number of indexers se- lecting to assign a descriptor for each document). here again, each descriptor that is not trustworthy can be flagged. during the search process, a system can then use free-text terms whenever it encounters a single-meaning term which is mapped to such a descriptor. the system may also convey its action to users. as these examples show, parts of the selection routine can be automated quite easily and with existing tech- niques. other parts of the selection routine, such as, when searchers use a descriptor for a single-meaning term because they do not fully understand the concept it represents, may prove to be unsuitable for the design of intermediary expert systems. there are yet further decisions which can usefully be implemented after additional research is performed. consider a situation in which a single-meaning term is mapped to a broader descriptor. there are three main options as shown in the decision tree (figure .) at the points [hi, [i], and [j]. now, suppose the term anorexia nervosa is mapped to the descriptor appetite disorders. an intermediary system may decide to enter anorexia nervosa as a free-text search key and possibly retrieve documents in which the subject is only mentioned, rather then discussed. or, it can combine this free-text key with the descriptor appetite disorders, using the and opera- tor, to retrieve articles that indeed discuss the disorder but may miss relevant documents that were not indexed under the broader descriptor. or, the system can select to enter the descriptor, in which case relevant documents might still be missed while documents discussing appetite disorders other then anorexia nervosa probably will be re- trieved. no option is better than any other; it all depends on the nature of a request. in other words, one option is re- quired for one type of request and another is required for another type. we do not have yet a general typology of requests that we can use to support the selection of the best option. further research in online searching behav- ior, however, can provide criteria to be used in automated systems. statistical approaches may suggest some help. for ex- ample, one may statistically analyze user satisfaction rate with each of the options. thus, even though we do not know explicitly which type of request requires which op- tion, we can implicitly detect which type is most common among a particular group of users by the option they find to be most satisfactory. we can then design an intermedi- ary expert system that first will always try the most com- monly satisfactory option, then ask for user’s reaction, and then utilize the next option if the first failed to pro- duce acceptable results. a much more promising approach suggests that online journal of the american society for information science-january searching behavior be investigated to reveal under which conditions each of the options is selected. we may find, for example, that when a term that is not central to a re- quest is mapped to a broader descriptor-and a user is primarily concerned with precision-searchers decide to combine a free-text search key with a descriptor. trans- ferring this condition to automated systems will enable a system to decide when and how to interrogate users about request requirements that are relevant to the search pro- cess. more specifically, a system may proceed searching independently until it encounters a problematic term, such as one that is mapped to a broader descriptor, and then ask users for a specific kind of feedback. in sum- mary, it is not difficult to envision an intermediary expert system which would: ( ) identify situations in which a re- quest statement is sufficient, and conversely those in which additional request criteria are needed for the search process to succeed, ( ) list the relevant criteria so that users can provide the pertinent data, and ( ) act upon data received to improve search results. quite a powerful system! these few examples clearly demonstrate the benefits that could be gained from the study of searching behavior of human intermediaries, and from utilizing the experi- ence of human intermediaries in the design of intermedi- ary expert systems. the selection routine presented here is not sufficient to develop adaptive algorithms. many issues, such as the nature of single-meaning and common terms or the con- ditions for the selection of a broader descriptor, need to be further investigated and rigorously defined. this rou- tine identifies problematic points in the search process and provides guidelines for research into searching be- havior that is relevant for automated systems. on the one hand, systems can interrogate users on request parame- ters-a subject users know best. on the other hand, they can select the most appropriate search keys-a decision casual end-users are not well enough informed to make. based on searchers’ experience, intermediary expert sys- tems can become experts indeed. acknowledgment the author wishes to thank dagobert soergel and irene l. travis for their helpful and insightful review of this article. references i. . . . . . . . doszkocs, t. e. “automatic vocabulary mapping in online search- ing.” international classification. ( ): - ; . marcus, r. s. “an experimental comparison of the effectiveness of computers and humans as search intermediaries.” journal of the american societyfor information science. ( ): - ; . pollitt, a. s. “a ‘front-end’ system: an expert system as an online search intermediary.” aslib proceedings. ( ): - ; . brenner, e. h.; lucey, j. h.; martinez, c. l.; meleka, a. . “api’s machine aided indexing project.” science technology li- braries. ( ): - ; . fidel, r. “online searching styles: a case-study-based model of searching behavior.” journal of the american societyfor informa- tion science. ( ): - ; . fidel, r. “the case study method: a case study.” library and in- formation science research. ( ): - ; . fugmann, r. “the complementarity of natural language and in- dexing languages.” international classification. ( ): - ; . king, d. w.; bryant, e. c. the evaluation of information services and products. washington, d.c.: information resources press; , p. . journal of the american society for information science-january introduction modeling the selection of search keys fig. . table . the selection routine discussion acknowledgment references doi: . /j.eswa. . . available online at www.sciencedirect.com www.elsevier.com/locate/eswa expert systems with applications ( ) – expert systems with applications structure clustering for chinese patent documents su-hsien huang a,d,*, hao-ren ke b, wei-pang yang c a institute of computer science and engineering, national chiao tung university, ta hsueh road, hsinchu, taiwan, roc b library and institute of information management, national chiao tung university, ta hsueh road, hsinchu, taiwan, roc c department of information management, national don hwa university, , section , da hsueh road, shou-feng, hualien, taiwan, roc d department of information management, minghsin university of science and technology, , hsin hsin road, hsin feng, hsinchu, taiwan, roc abstract this paper aims to cluster chinese patent documents with the structures. both the explicit and implicit structures are analyzed to represent by the proposed structure expression. accordingly, an unsupervised clustering algorithm called structured self-organizing map (som) is adopted to cluster chinese patent documents with both similar content and structure. structured som clusters the similar content of each sub-part structure, and then propagates the similarity to upper level ones. experimental result showed the maps size and number of patents are proportional to the computing time, which implies the width and depth of structure affects the performance of structured som. structured clustering of patents is helpful in many applications. in the lawsuit of copyright, companies are easy to find claim conflict in the existent patents to contradict the accusation. moreover, decision-maker of a company can be advised to avoid hot- spot aspects of patents, which can save a lot of r&d effort. � elsevier ltd. all rights reserved. keywords: structure clustering; chinese patent; structure expression; metadata . introduction . . background digital libraries provide various integrated services to coordinate information over internet. however, the dis- tributed and heterogeneous data complicate the integration when it involves several issues, such as format, content, semantics, etc. the discrepancy and redundancy confuses users in readability. to provide a unified view, data cluster- ing with an overview by integrating similar data has received considerable attention in recent researches. conventional data clustering adopts feature vector to represent data. it lacks some aspects of consideration to identify similarity. for example, chemical compounds with the same molecules but different structures are not the same. moreover, two trademarks with the same compo- - /$ - see front matter � elsevier ltd. all rights reserved. doi: . /j.eswa. . . * corresponding author. e-mail address: sshuang@cis.nctu.edu.tw (s.-h. huang). nents but different placements cannot be identified similar, too. these two examples account for the structure is a key-factor to influence the clustering in particular domains. although conventional structure clustering focuses on visual pattern query, chemical structure identification and syntactic parsing tree, rare mentioned in documents. this is owing to the structure in documents is hard to obtain. interestingly, the development in rhetorical structure the- ory (rst) facilitates the extraction of structure. rst was proposed in the s and has been successfully applied to documents summarization, automatic layout, and so on. rst categorizes the characteristics of phrases and con- structs a closure tree to represent structure. therefore, applying rst to represent documents structure becomes possible and practical. several kinds of documents, like patent and electronic thesis, contain both structure and content information. to cluster these documents, only the same content in the same structure can be identified similar. the tort of copy- right is a good illustration to require conflict in both claim mailto:sshuang@cis.nctu.edu.tw s.-h. huang et al. / expert systems with applications ( ) – and claim structure. the structure of documents can be categorized into two types. explicit structure represents the existent attributes of documents, like ‘‘subject’’, ‘‘abstract’’, ‘‘publisher’’ and ‘‘classification’’ in patent doc- uments. on the other hands, the structure analyzed from content is implicit structure. the implicit structure can be extracted by analyzing the writing style and the well- established convention. for example, in the claim field of patents, ‘‘the . . . of claim ’’, ‘‘comprising’’, ‘‘wherein’’, ‘‘having’’, ‘‘consist of’’, etc. construct the implicit structure. with both explicit and implicit structures, document struc- ture can be constructed and applied in structure clustering. this study tries to apply chinese patent documents in structure clustering. patent structure has been mentioned in recent literatures. shinmori, okumura, marukawa, and iwayama ( ) provides a rhetorical method to cate- gorize japanese patent structure into three styles – process sequence style, element enumeration style and jepson-like style. each claim in the patent is given a weight to evaluate the relationships among different claims. fujii, iwayama, and kando ( ) extends shinmori’s work and adds delimiter to punctuate claim into components. mase, matsubayashi, ogawa, iwayama, and oshio ( ) analyze the patent structure into premise, description and target parts. keywords in different part can be re-weighted to enhance query precision (iwayama, fujii, kando, & marukawa, ). clustering in patent is a real-time application. real-time means the time constraint of clustering algorithm is tight when users require the result on-line. unfortunately, struc- ture clustering is more complicate than conventional one, where the computing time is subject to the structure. hence, the first requirement of structure clustering is effi- ciency. additionally, real-time also implies there is no training data in advance. therefore, unsupervised algo- rithm is suggested when there is unnecessary to use training data in the process. unsupervised clustering integrates sim- ilar data in definite circles, and adjusts the parameters to obtain result. kohonen proposes an unsupervised neural network self-organizing map (som) and receives excellent performance (kohonen, ). som provides unsuper- vised neural network clustering and maps high dimension data (usually two) into a low-dimension map. in som, clo- ser nodes in the map imply shorter distance in real data. som applies in many domains, like bio-structure cluster- ing, graph structure clustering and audio-pattern cluster- ing, etc. the multi-dimension representation of som facilitates the readability which is suitable for patent clustering. . . literatures of early studies it is noteworthy that conventional som cannot deal with structure data. in , a general framework is pre- sented to deal with structure by neural network (frasconi, gori, & sperduti, ). this framework constructs direc- ted ordered acyclic graphs (doag) for structure data and recurrently proceeds data by following doag sequence. this framework further motivates the upcoming researches. somsd (sperduti, ) applies self-organiz- ing map (som) to manipulate structured data. each neu- ron in somsd is equipped with two weights: one for previous structure and one for two-time previous structure. structure similarity is estimated by concatenating current similarity and these two structure similarity. rsom (voegtlin, ) recursively calculates som by adding cur- rent similarity and previous structure. hammer, micheli, and sperduti ( ) propose a general framework to further extend somsd. vesanto and alhoniemi ( ) cluster each sub-structure part in single som and use hier- archical k-means cluster to investigate similar structure. smith and ng ( ) and roussinov and chen ( ) derive user navigation patterns as structure to cluster web page navigation pattern. similar navigation patterns (trea- ted as structure graph) are clustered together by som. xu, chang, and paplinski ( ) use on-line expansion to enlarge som size into multiple layers. other of structure clustering can be found in literatures (hagenbuchner, sperduti, & tsoi, ; hammer & jain, ; rossi, conan-guez, & golli, ; sperduti & starita, ; strickert & hammer, ). . . purpose of this study this study is intended to apply chinese patent docu- ments in structure clustering. a patent structure is analyzed first and provides structure expression to represent it. hav- ing the structure established, self-organizing map (som) is adopted. several modification of som is undertaken to process structure, including input, output and training algorithm. in brief, three major objectives involved in this study are: ( ) construct expression model to represent the patent structure. ( ) develop structure clustering algorithm to process structure patent. ( ) evaluate the structure clustering algorithm. this paper is organized as follows. chapter analyzes the structure of chinese patent documents. chapter pro- poses structured som to cluster patents. chapter imple- ments structured som and a series of experiments are conducted in chapter . chapter draws the conclusions. . structure analysis as shown in fig. , the preprocessing of structure clus- tering is to analyze the structure. the received documents are first represented by structure expression, and then determine the input sequence and maximum branch num- ber (mbn). to formally describe the structure, two formal constructors are defined. given a structure s, s can be represented by the following two constructors: s.-h. huang et al. / expert systems with applications ( ) – ( ) tuple constructor (tc): a tuple constructor tc of structure s is an ordered list constructed by the union of single-value attributes. for example, the set of attributes ci, c Æ occurence = , c Æ occurence = , . . .,cn Æ occurence = , are represented as tcs = {c , c ,. . .,cn}s. the subscript s is the name of tc. ( ) set constructor (sc): a set constructor sc of struc- ture s is a multi-value type with the same occurrence larger than one. for example, the set of attributes ci, c Æ occurence = i, c Æ occurence = i,. . .,cn Æ occu- rence = i, are represented as hc ; c ; . . . ; cni i s. the sub- script s represents the name of sc and i represents the maximum occurrence. . . explicit structure a structure document contains two types of structure. the first one is explicit structure, which is obtained from the schema of structure documents (like data metadata). explicit structure is applicable in documents with regular schema. to take a simple example of electronic thesis, the attributes ‘‘subject’’, ‘‘abstract’’, ‘‘chapter’’, ‘‘section’’, and ‘‘paragraph’’ contain explicit structure. the structure expressions above example can be expressed as follows: subject; abstract; contenth in paragraph d en section � �n chapter ( ) book where n , n and n stand for the maximum occurrence of the attributes. applying explicit structure in structure clustering is meaningful when the documents contain regular schema and massive content in each attribute. for example, elec- tronic thesis and xml-based news are suitable for explicit fig. . structure of structure clustering. in this study, chinese patent docu- ments contain most information in ‘‘claim’’ field. there- fore, implicit structure is introduced in the next section. . . implicit structure the structure analyzed from content is implicit structure. the implicit structure can be extracted by analyzing the writing style and the well-established convention. in the ‘‘claim’’ field of chinese patent documents, two types are categorized: • composition style: as in ‘‘ ’’ (comprising), ‘‘ ’’ (includes), the set of element is described. these key- words are used in method to imply the composition of the inventions. this type of style represents a set of ele- ments contains in the claim structure. • pre-condition style: as in ‘‘ ’’ (as claimed in . . .), ‘‘ ’’ (wherein), these descriptions imply the state- ments has followed a list of composition. this type of style represents a list of compositions construct the claim structure. comparing with other-linguistic patents, there are sub- stantial difficulties to obtain several relationships in chi- nese patent. for example, in shinmori’s method, the process sequence style (means the sequence of processes, like ‘‘does’’, ‘‘and does’’) is ambiguous in chinese gram- mar. it’s always followed by ‘‘ ’’ where is the same meaning with the article ‘‘one’’. these two types of style can derive the implicit structure of patents. for the presence of these keywords, a hierarchy relationship can be constructed. for example, fig. illus- trates an example of chinese patent and represents the implicit structure analyzed by these two styles. to give each chinese patent. s.-h. huang et al. / expert systems with applications ( ) – segmented claim an id, naming mechanism is defined for each claim as follows: patent id-claim idlevel - � � �claim idleveli the structure can be formally described by the structure expression, which can be more comprehensive for structure clustering. the composition style means the occurrence happens in the following statements. by following the structure expression in previous section, it can be expressed as hii, where i stands for the structure level. the pre-condi- tion style implies an attribute contains in the following statement. by following the structure expression, it can be expressed as {}i, where i stands for the structure level. for the example of fig. , the structure expression can be expressed as follows: fclaima;fclaimbglevel ; claimcglevel � � patent . . determine input sequence after the structure is analyzed, the input of each sub- part structure is determined by following directed acyclic graph (dag). a depth-first search is adapted in the naviga- tion of structure tree. for the formal expression of both explicit and implicit structure, the input of structure follows two principles: ( ) by order of the attributes in tuple constructor. ( ) iteratively expand set constructor to i times. for example, the input sequence of explicit structure subject; abstract; contenth in paragraph d en section � �n chapter ( ) book fig. . structu is subject ! abstract ! chapter ! section ! paragraph !���! paragraphn ! section !���! paragraphn !���! chaptern ! sectionn ! paragraphn for the example of fig. , the input sequence to structure clustering is: p ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! . . determine maximum branch number the next step is to determine the maximum branch num- ber (mbn) of tuple and set constructor. the maximum branch number (mbn) represents the maximum number of branches in patent structure. in the structure expression of both explicit and implicit, mbn is the maximum num- ber of tc attributes and sc occurrence. for example in fig. , the mbn is when the largest branch and occur- rence is less than . . structured som self-organizing map (som) (kohonen, ) provides unsupervised neural network clustering and maps high- dimension data into a low-dimension map (usually two). there are five steps to train som (in fig. ): ( ) initialize weight vectors of output map as the same number features with input document vector. ( ) present input documents in order. red som. s.-h. huang et al. / expert systems with applications ( ) – ( ) compute the distance between the input document and all nodes in the map and select the closest node as the winner. ( ) update the weights of the winner node and its neighbors. ( ) repeat steps ( )–( ) to other documents and iterate all inputs until convergence. label the regions of the final map to represent the clustering result. som with structure receives considerable attention in recent years (rossi et al., ; strickert & hammer, ; voegtlin, ). the structured som in this study refers to hagenbuchner’s self-organizing map clustering algorithm in structured data (hagenbuchner et al., ). hagenbuchner’s approach applied structured som on the images identification. the main concept is to calculate each sub-structure respectively, and concatenate the result to the upper structure. notably, the images in hagenbuch- ner’s method have simple structure and regular compo- nents. experiment result has shown that structured som can cluster similar images together and distinguish the structure difference in the map. this paper applies struc- tured som in chinese patents and tries to identify similar patent with different structure. the process of structured som is described in fig. . . . training for explicit structure applying som in chinese patent documents requires modification of input/output vectors and algorithm. the explicit structure can be represented by structure expres- sion shown in previous section. the next step is to deter- mine the maximum branch number (mbn) of tuple and set constructor in explicit structure. the shortage of nodes less than mbn requires additional null nodes. for the sim- plicity, assume the mbn = in the example of explicit structure (n = , n = , n = ), the structure expression can be rewrite as: subject;abstract; content;null;nullh in paragraph d en section ; �� null;nullin chapter � book the output nodes of som also need to be modified as the same number fields as input vector. for example, the output node of structured som in previous example is d x;y ¼ðv d x;y ;ðx ; y Þ;ðx ; y Þ;ðx ; y ÞÞ the structure expression is assigned into som input vector by directed acyclic graph (dag) sequence, for example, given a three node input vector, the input vector are rewrite as following: d book ¼ðv book;ðxsubject;ysubjectÞ;ðxabstract;yabstractÞ;ðxchapter;ychapterÞÞ additionally, the distance calculation is updated as following: d ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðv node � v bookÞ q þjv node � v subjectjþ jv node � v abstractjþ jv node � v chapterj where |vnodei � vnodej| represents the distance between these two vectors. the modification of distance formula means the cascaded calculation to all connected nodes. the adapting of structured som is updated as following: wd x;yðt þ Þ¼ wd x;yðtÞþ gðtÞ� jv node � v subjectj wd x ;y ðt þ Þ¼ wd x;yðtÞþ gðtÞ� jv node � v abstractj wd x ;y ðt þ Þ¼ wd x;yðtÞþ gðtÞ� jv node � v chapterj . . training for implicit structure for the implicit structure, the dimension of the vector is also determined by the mbn in all input patents. assume the mbn = , the representation in previous example is: ðv patent; claim ; claim ; claim Þ ðv node ; claim ; claim ; claim Þ ðv node ; claim ; claim ; claim Þ ðv node ; claim ; claim ; claim Þ the output nodes of som also need to be modified as the same number fields as input vector. for example, the output node of structured som in previous example is d x;y ¼ðv d x;y ;ðx ; y Þ;ðx ; y Þ;ðx ; y ÞÞ additionally, the distance calculation is updated as follow- ing formula: d ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðv node � v patentÞ q þjv node � v claim jþ jv node � v claim jþ jv node � v claim j where |vnodei � vnodej| represents the distance between these two vectors. the modification of distance formula means the cascaded calculation to all connected nodes. the adapting of structured som is also updated as following: wd x;yðt þ Þ¼ wd x;yðtÞþ gðtÞ� jv node � v claim j wd x ;y ðt þ Þ¼ wd x;yðtÞþ gðtÞ� jv node � v claim j wd x ;y ðt þ Þ¼ wd x;yðtÞþ gðtÞ� jv node � v claim j . implementation fig. displays the structured som system. the system contains six major parts: ( ) som system: choose standard som and structured som. ( ) parameter setting: set the map dimension, learning rate, adaptation neighborhood and iteration to train. ( ) function selection: select open file, close result, train- ing, save result, resize map and quit training. parameter setting function selection map information system message som map som system fig. . implementation of structured som. s.-h. huang et al. / expert systems with applications ( ) – ( ) map information: display the containing patents in the som grid. ( ) system message: show the current iteration and pro- cessing message. ( ) som map: illustrate the clustering result. fig. demonstrates an example to apply structured som in chinese patent documents. initially, chinese patent documents are selected from database. these docu- ments are sent to standard som to obtain preliminary clustering. in the clustering, totally four clusters are labeled, e-commerce, binary-tree, identification and data- e-commerse ) standard som ) structured som ) claim clustering claim conflict structure clustering fig. . patent clustering. base. the interesting is paid in the five patents of e-com- merce cluster. there are three observations: ( ) commerce system is subjected. ( ) client–server architecture is addressed. ( ) encrypt system is applied. hereafter, these patents are sent to structured som. in the structured som, these five patents are separated into three clusters, which represent they are different in struc- ture. for cluster , the patent primarily describes the con- table time vs. documents (s) document number standard som structured som ( · map, iterations, mbn = ). table times vs. mbn (s) mbn structured som ( · map, documents, iterations). s.-h. huang et al. / expert systems with applications ( ) – struction of commerce system and is written in a very flat style. on the contrary, cluster mainly focuses on encrypt system and is written in a very deep style. these two clus- ters are distinguished with cluster not only the different subjects but the different patent structure. notably, three documents marked as cluster clusters together and might have high risk to conflict in the claims. therefore, these claims are segmented by following the naming mechanism and cluster by som again to obtain conflict clusters. in the third step of fig. , claim conflicts have observed in three areas: ( ) wire and wireless. ( ) client-side and server-side service. ( ) digital authorization and public key encrypt. these three patent conflicts provide good advisement for patent inventors to avoid involvement in this area. struc- tured som is also helpful in many patent applications. in the lawsuit of copyright, companies are easy to find claim conflict in the existent patents, which is easy to con- tradict the accusation. moreover, decision-maker of a com- pany can be advised to avoid hot-spot aspects of patents from structured som, which can save a lot of r&d effort. . experiments a series of experiments are conducted to examine the performance of structured som. the data set comes from the claim attribute of chinese patent documents, to exam- ine the implicit structure of patents. the experimental plat- form is intel celeron . ghz cpu, mb ram, microsoft os. the structured som was implemented by java. in table , the experiment was conducted to examine the performance in different map size. the parameters were set to documents in iterations with mbn = . the computing time of both standard and structured som is positive proportional to the map size. however, structured som is more time-consuming than standard one. in table , the experiment was conducted to examine the performance in different document numbers. the parameters were set to · map in iterations with mbn = . the computing times of both standard and structured som are positive proportional to the document numbers. the experiment corresponding to table implies the map size and documents are important factors to influ- ence the computing time. however, the time complexity between standard and structured som is extremely large, table time vs. map size (s) map size · · · · standard som structured som , ( documents, iterations, mbn = ). too. one explanation for the phenomenon is the document structure complicates the clustering. the deeper the struc- ture is, the longer the computing time. the experimental result of table further introduced another factor to influence computing time. with docu- ments in five iteration of · map, structured som is nearly exponential proportional to the mbn. this is the most important factor account for the complexity of struc- tured som because both input vector and output map are modified to fit the dimension of mbn. these results lead to the conclusion that the width (mbn) and the depth of the structure are two key-factors for structured som performance. . conclusions and future work increasing chinese patent documents require the clus- tering in the application. to cluster precisely, patents with similar content but different structure should be distin- guished. however, conventional clustering lacks of struc- ture consideration cannot identify similar patents with different structure, and rare research mentioned in text domain. in this paper, a clustering algorithm called struc- tured som is proposed to clusters structure patents by considering the content and structure information simulta- neously. structured som requires the analysis of patent structure in advance. two types of structure expression, explicit and implicit structure, are provided. structured som modifies the input and output vectors for structure documents and iteratively training by adapting each sub- part of structure. an example to cluster chinese patent documents has successfully implemented. claim conflict appeared in the analysis provides advisements for patent inventors and decision-makers of business. in the lawsuit of copyright, companies are easy to find claim conflict in the existent pat- ents, which is easy to contradict the accusation. moreover, decision-maker of a company can be advised to avoid hot- spot aspects of patents from structured som, which can save a lot of r&d effort. experiments are conducted to examine the performance of structured som. conclusions are given that both map s.-h. huang et al. / expert systems with applications ( ) – size and documents are positive proportional to the computing time. this implies two factors are important: the width (mbn) and depth of structure. mbn is nearly exponential drop-off the performance of structured som. in some practical application, prediction is adapted to esti- mate node distance to reduce computing time but sacrifice neglectful accuracy. structure clustering is rarely mentioned in the text domain because the text with structure (like data metadata) has small amount of content in the attributes (explicit structure). with the development of digitalization, some structure documents, like electronic thesis and on-line news, can be applied in structured som. the future direc- tion for this study might be to discover more applications for structure clustering. it is also lots of work to improve the efficiency of structured som. acknowledgements this research was supported by the software technol- ogy for advanced network application project of institute for information industry in and sponsored by moea, roc. references frasconi, p., gori, m., & sperduti, a. ( ). a general framework for adaptive processing of data structures. ieee transactions on neural networks, ( ), – . fujii, a., iwayama, a., & kando, n. ( ). test collections for patent retrieval and patent classification in the fifth ntcir workshop. in proceedings of the fifth international conference on language resources and evaluation (pp. – ). hagenbuchner, m., sperduti, a., & tsoi, a. c. ( ). a self-organizing map for adaptive processing of structured data. ieee transactions on neural networks, ( ), – . hammer, b., & jain, b. j. ( ). neural methods for non-standard data. in european symposium at artificial neural networks (pp. – ). hammer, b., micheli, a., & sperduti, a. ( ). a general framework for unsupervised processing of structured data. in european symposium on artificial neural networks (esann’ ) (pp. – ). iwayama, m., fujii, a., kando, n., & marukawa, y. ( ). evaluating patent retrieval in the third ntcir workshop. information processing and management, ( ), – . kohonen, t. ( ). the self-organizing map. neurocomputing, , – . mase, h., matsubayashi, t., ogawa, y., iwayama, m., & oshio, t. ( ). two-stage patent retrieval method considering claim structure. in ntcir workshop meeting. rossi, f., conan-guez, b., & golli, a. e. ( ). clustering functional data with the som algorithm. in proceedings of european symposium on artificial neural networks (esann’ ) (pp. – ). roussinov, d. g., & chen, h. ( ). information navigation on the web clustering and summarizing query results. information processing and management, , – . shinmori, a., okumura, m., marukawa, y., & iwayama, m. ( ). patent claim processing for readability – structure analysis and term explanation. in acl- workshop on patent corpus processing. sapporo: association for computational linguistics. smith, a., & ng, a. ( ). web page clustering using a self-organizing map of user navigation patterns. decision support systems, , – . sperduti, a. ( ). neural networks for adaptive processing of structured data. lecture notes in computer science, , – . sperduti, a., & starita, a. ( ). supervised neural networks for classification of structures. ieee transaction on neural networks, ( ), – . strickert, m., & hammer, b. ( ). unsupervised recursive sequence processing. in proceedings of european symposium on artificial neural networks (esann ) (pp. – ). vesanto, j., & alhoniemi, e. ( ). clustering of the self-organizing map. ieee transactions on neural networks, ( ), – . voegtlin, t. ( ). recursive self-organizing maps. neural networks, , – . xu, p., chang, c. h., & paplinski, a. ( ). self-organizing topological tree for online quantization and data clustering. ieee transactions on systems, man, and cybernetics, ( ), – . structure clustering for chinese patent documents introduction background literatures of early studies purpose of this study structure analysis explicit structure implicit structure determine input sequence determine maximum branch number structured som training for explicit structure training for implicit structure implementation experiments conclusions and future work acknowledgements references kinship algebra expert system (kaes): a software implementation of a cultural theory ucla ucla previously published works title kinship algebra expert system (kaes): a software implementation of a cultural theory permalink https://escholarship.org/uc/item/ dz q sr journal social science computer review, ( ) author read, dwight w publication date doi . / peer reviewed escholarship.org powered by the california digital library university of california https://escholarship.org/uc/item/ dz q sr https://escholarship.org http://www.cdlib.org/ . / social science computer reviewread / kinship algebra expert system kinship algebra expert system (kaes) a software implementation of a cultural theory dwight w. read university of california, los angeles the computer program kinship algebra expert system (kaes) provides a graphically based framework for constructing, if possible, a generative algebraic model for the structure of a kin- ship terminology (the terms used to refer to one’s kin). the algebraic modeling is based on a the- ory of kinship terminologies elaborated through writing the software program. the theory relates the properties and structure of kinship terminologies to an underlying logic that the kaes program helps uncover and model as a generative structure. the program then relates the structural logic of a kinship terminology modeled by the kaes program to a genealogical space based on genealogical tracing of kin relations. the kaes program demonstrates the surpris- ingly logical character of kinship terminologies and challenges the received view of the primacy of genealogical relations in defining cultural kinship through showing how genealogical definitions of kin terms can be accurately predicted in the terminologies considered to date. keywords: kinship; cultural theory; terminology; algebraic modeling the computer program, kinship algebra expert system (kaes), is a graphical user inter-face application for investigating and modeling the structure of kinship terminologies and the ways in which these are instantiated (for further information and for downloading a copy of the application program, see read & fischer, ). it is based on a theory of kinship terminology structures elaborated through writing the kaes program. the program is an example of how a theory can be approached and evidence evaluated with respect to that the- ory to produce results that identify a future research trajectory rather than simply accounting for relations in a body of data. the results of research based on kaes will have far-reaching consequences for many other domains in the social sciences. the focus on kinship terminologies may strike many readers as a rather specialized and narrow application, especially following more than a decade in which the study of kinship has diminished in the curricula of major anthropology programs, only to be replaced by courses on gender and new reproductive technologies. however, kaes successfully imple- ments a theory of kinship terminology structures that is fundamental to our understanding of the kinship basis through which a newborn person initially obtains social identity crucial throughout his or her life. in addition, by providing a fully worked out theory and model of one culturally constructed symbolic system (symbolic in the linguistic or mathematical and social science computer review volume number february - © sage publications . / http://ssc.sagepub.com hosted at http://online.sagepub.com not the interpretive sense), we gain insights for the critical task of understanding the relation- ship between culturally constructed symbolic systems and their cultural instantiation. the program owes its genesis in the s to a problem with communicating the results of algebraic representation of the structure of kinship terminologies to an anthropological audi- ence with little or no mathematical background, especially in the area of algebraic structures. i had just published a paper (read, ) showing how the american kinship terminology (akt) could be structurally described in the form of what has been called a kin term map (leaf, ) and, more importantly, how the resulting structure could be generated in the form of an algebraic structure known as a semigroup. it became evident that the number of readers in anthropology who could follow the algebraic argument was very limited, and one of my (then) graduate students, dr. clifford behrens, suggested that we develop a software program for doing the structural analysis that would not require a mathematical background on the part of the user. the algebraic part, we decided, could be implemented using prolog, and we could develop an expert system as a way to provide assistance to a user without back- ground in modeling using algebraic structures. we made the assumption that the algebraic part of the analysis would require, even with the algebraic machinery presented through a user-friendly interface, considerable guidance because the degree to which kinship terminol- ogies had a common, underlying logic was unknown at the time. i had demonstrated that the akt did have a logic that allowed for the terminology structure displayed in the form of a kin term map to be isomorphically generated from first principles. additional work on other ter- minologies made it evident that the same was likely to be true for terminologies very different from the akt, such as the trobriand kinship terminology used by the inhabitants of the trobriand islands. the trobriand terminology has been the basis for considerable theoretical work on kinship terminologies using ideas such as rewrite rules borrowed from linguistic analysis (see lounsbury, ) and provided a good test case for determining those features of a kinship terminology that could better be explicated through an algebraic or generative approach grounded in ethnographic observations about the way kin relations are calculated by individuals rather than linguistic methods. consequently, the program was not simply a translation of an existing theory and application of that theory into a computer idiom. instead the software program became the idiom through which the theory was developed (see read & behrens, ). the advantage of using a computer program as the basis for developing a theory of kinship terminology structures became evident as work progressed. much of the algebraic analysis related to working out the structure of an algebra once its generating set and structural equa- tions have been specified. this part of the analysis could be programmed in a relatively straightforward manner using prolog and obviated the need for a theorem-proof written for- mat. with the algebraic machinery developed in prolog, the analytical and programming effort could then be focused on what precisely was meant by a terminology having a structure that could be generated from first principles and on what were those first principles. many of the issues that arose during the writing of the software program had to do with developing the formal, structural equivalent of concepts that were used more intuitively in anthropological theorizing about kinship terminologies, such as reciprocals of kin terms, a key aspect of any kinship terminology. briefly, the reciprocal of a kin term is the kin term that another person would use to refer to ego when ego refers to a person by that kin term. in the akt, for example, if i refer to a person as uncle, then that person would refer to me as nephew (because i am a male person). the kin terms uncle and nephew are said to be reciprocal kin social science computer review terms. although this is a satisfactory notion from an ethnographic viewpoint, a structural analysis requires that the concept of reciprocity be given a structural definition (to be dis- cussed below) for the concept to be implemented in a software program that provides the machinery for modeling the structure of a kinship terminology as a generative, algebraic structure. an unexpected consequence of simultaneously working out a theory of kinship terminol- ogy structures while implementing that theory in the form of an algorithmically based com- puter program was that the expert system part of the original design for the computer program began to take on less importance as it was discovered that kinship terminology structures are highly logical and have few degrees of freedom in terms of how the structure is generated. structural properties that are necessary for consistency in calculations with kin terms—such as consistency in sex marking of kin terms used in kin term calculations—and properties of kin terms such as “parent of sibling is parent” in the akt can be automatically introduced by the algebraic machinery. as the theory became more completely developed, the algebraic analysis performed by kaes almost became a single, large composite algorithm. it cannot truly be a single algorithm as there are different terminology structures that begin with the same generating elements and initial structural equations. what distinguishes different kinship terminology structures are not differences in the algebraic methods of analysis (which would require expert system assistance for the naïve user) but structural differences in terminologies relevant to the ethnographic level of cultural differences in the specification of what constitutes kinship in a particular society. for exam- ple, the trobriand terminology and the tongan terminology (used by the inhabitants of the kingdom of tonga in the pacific ocean) have the same initial starting conditions (set of gen- erating elements, structural equations satisfied by these generating elements) but differ at the point where ascending terms (terms used for persons in a generation older than the speaker) are combined with descending terms (terms used for persons in a younger generation) in accordance with structurally alternative ways these two sets of kin terms may be combined together (see bennardo & read, n.d.). the structural implications for alternative ways the ascending and descending terms are combined together to form a single structure relate directly to differences in the usages between the two terminologies for terms that otherwise are in similar structural positions in the two terminologies. the unexpected discovery that terminologies may be modeled almost as a single, compos- ite algorithm has major implications for our understanding of cultural constructs such as kin- ship terminologies and their relationship to kinship systems in human societies. rather than providing support for the received view in anthropology that considered terminologies to be a labeling system for categories of kin distinguished for reasons external to the kinship termi- nology, the kaes program makes it evident that the structural properties of kinship termi- nologies arise from logical constraints on the generation of structure and differences among terminologies relate to places in the generation of a structure where alternative ways to develop the full terminology structure are logically possible. the current version of kaes has been implemented in java with michael d. fischer (read & fischer, ). in the java version of kaes, the underlying theory has been elabo- rated further, especially with regard to the relationship between a kinship terminology and genealogy. no attempt was made to translate the pascal and prolog code into java as that code suffered, from a programming viewpoint, from the fact that it was written while developing the theory on which the current computer program is based and was written without having read / kinship algebra expert system access to a fully object-oriented programming language. the latter has made possible a much more succinct and more precise representation of the underlying theory. overview of the kaes program the kaes program has three major components, with menu options including map oper- ations, algebra operations, and operations. the first component (map operations) provides the machinery for entering of kinship terms in the form of a kin term map (see figure ) that displays the manner in which kin terms are linked to one another through a kin term product based on ethnographic observations about the ways individuals calculate kinship relations. the kin term map can be represented in the form a table that shows the kin terms that result from taking the product of a kin term with one of the primary kin terms such as mother, father, son, or daughter in the akt. the kin term map is also presented in the form of a directed graph linked to the table form for displaying the product of kin terms with the primary kin terms. when a kin term is entered by the user, its attributes such as sex marking and whether or not it is a generator kin term are entered by the user. the kin term map displays native knowledge about a kinship system expressed through a kinship terminology and provides the target structure for the algebraic modeling done in the second component (algebra operations) of the program. the algebraic modeling begins by structurally simplifying the kin term map through user- selected options for simplification of a kin term map. the simplification is run in reverse dur- ing the algebraic modeling as one elaborates on an initial algebraic model constructed in accordance with the simplified kin term map. a modified kin term map is defined to be sim- plified when the modified kin term map has a single, ancestral generating kin term and only consists of ancestral kin terms. the kaes program allows for the user to model a kin term map using user-selected options that activate an ensemble of individual steps composing a stage in the algebraic modeling, such as constructing the descendant structure from the ini- tial, ascendant structure (see figure ). alternatively, the user may sequentially activate the individual steps making a stage in the algebraic modeling. as the modeling proceeds, the kaes program automatically introduces structural equations that are part of the distinctions made regarding kin terms in the kinship terminology. each stage in the algebraic modeling of a kinship terminology structure is successful if the algebraic modeling ends up with an algebraic structure isomorphic to the kin term map from the initial simplification of the kin term map corresponding to the current stage in the alge- braic modeling. the algebraic construction stops when an algebraic structure that is isomor- phic with the kin term map for the complete kinship terminology has been generated. the algebraic structure may be locally modified by user-selected rules that implement the logic by which a portion of the generated structure may be culturally modified. for example, the generated algebraic model for the akt has an undifferentiated “cousin space.” the cousin space is then modified by a logic applied consistently to kin term products to arrive at the cousin space with the features found in the kinship terminology and expressed through the kin terms of the form “ith cousin j-times removed.” rules are largely terminology spe- cific. rules for the terminologies considered so far are included in the kaes program. addi- tional rules will be added to the kaes program as the analysis expands beyond the kinship terminologies considered so far (see figure for the algebraic model of the akt generated by the kaes program, including the rule for the cousin terms). social science computer review the third component (operations) constructs a genealogical instantiation of the algebraic structure. the instantiation is constructed by first linking the generating elements in the alge- braic structure to the genealogical space. in the akt, for example, the kin terms parent and child are the primary, generating kin terms. these two kin terms are mapped to the sets of genealogical positions (genealogical mother, genealogical father) and (genealogical son, genealogical daughter), respectively. this mapping, along with the mapping of the identity element to the set (ego), is next extended to other kin terms through set products of the map- ping of the generating elements in corresponding to the primary elements in a kin term prod- uct. the result of carrying out the genealogical instantiation of kin terms can be displayed in tabular form or in the kind of genealogical diagrams used by anthropologists to show the mapping of kin terms onto a genealogical structure (see figure ). anthropological problem: what constitutes kinship? unraveling what constitutes kinship relations and kinship behavior in human societies has been a central theme in anthropological theorizing ever since the publication by lewis henry morgan of systems of consanguinity and affinity of the human family in . the central- ity of kinship in anthropological theorizing stems from the fact that all human societies are constructed, to one degree or another, around kinship and kinship relations. although larger scale societies may have developed institutions for societal integration that are not directly based on kinship, such as the notion of citizenship and elected government in state democra- cies, no society has removed kinship as a fundamental basis for the social identity of individ- uals. regardless of other aspects of societal organization, we are born into families that are formed on the basis of kin relations, and from birth onward we are provided, through kinship, with an identity through which we are linked to other individuals, and through these linkages we are provided with a social identity fundamental to our existence as social beings. our kin- based social identity may coexist with other identities we also hold. in american culture, we are given individual identity at birth through a supposedly unique name—or at least a name unique within the domain of persons known to the parents of the newborn child. the notion of individual identity in american culture is in accord with the concept of a governmental system that also provides citizens with individual rights that recognize the uniqueness of each person as a biological entity. a child also takes on a kinship identity in american culture even though the latter has the opposite effect of constructing a social identity that overrides the biological uniqueness of each person. a newborn child may be a brother or a sister to any other offspring of her or his parents, a nephew or a niece to the siblings of her or his parents, a cousin to the children of the siblings of her or his parents, and so on, simply by virtue of birth and regardless of individual qualities. this kinship identity provides a social identity expressed (in american culture) through terms such as brother, sister, son, daughter, and so on and carries with it expected patterns of behavior by virtue of the kind of relationship involved. americans have shared (though not identical) notions of how children should behave in their status as children or how siblings should interact in their status of siblings, regardless of their biological uniqueness. in small scale societies, kinship is the basis on which societal boundaries are formed and is a prerequisite to social interaction among individuals. often the term used to identify the group to which one belongs translates into something like the meaning of the term ju/ ‘hoansi, used by a hunter-gatherer group in botswana to refer to themselves, namely “we, the read / kinship algebra expert system real people” (lee, ). social interaction between individuals does not take place unless they first recognize each other as kin. nonkin, or strangers, are feared or perhaps considered potentially harmful because those who are not your kin may not share the same set of moral values and modes of behavior as do your kin. nonkin are not predictable in terms of behavior in contrast to kin, whose behavior is constrained by patterns of behavior associated with the kinds of kin relations. stranger, in the context of kin-based societies, refers not to those with whom one has not had prior interaction but to individuals outside of the domain of individu- als who share the same moral and social values. the world of individuals with a shared set of moral and social values is one’s kin. kin, by virtue of being kin, share a common framework of behavior that is appropriate even if the persons in question have not previously met. with larger scale societies, a domain of kin-like relations that extends many of the properties of shared, expected patterns of behavior among kin to a larger domain of individuals has been added. tribes, for example, include individuals who may not know their kin relationships but do know their connection to one another through the social groups to which they belong such as lineages and how these social groups are linked via connections derived from kin connec- tions. in state societies, patterns of expected behavior have been extended outside of the domain of kin and kin-like relations through institutions such as citizenship that define for a person the set of rules and laws that are to be obeyed and where infraction of those rules and laws may lead to punishment. but even citizenship has a quasikin basis through the notion that citizenship is passed on from parent to offspring. we may reasonably assert, then, that kinship is the fundamental concept on which human societies have been formed and constructed. without understanding what constitutes kinship and how kinship articulates with, and provides the basis for, interaction in all other domains, we only have an incomplete picture of the nature of human societies and social systems. try- ing to understand human societies without understanding the underlying kinship system is analogous to trying to understand the chemical properties of matter without understanding the underlying atomic and molecular properties from which those chemical properties derive. the kin relations that connect individuals provide the underlying framework for the interaction of individuals, and without understanding the properties and structural form that kinship relations may take on in a given society, we cannot fully understand how the individ- uals making up that society become social and not just individual beings and how the society is structured as a social system. although the patterns of expected kin behaviors may vary from one society to another, all societies have a means for identifying or labeling some set of relationships between an indi- vidual and other individuals that is expressed linguistically through a set of terms known to anthropologists as a kinship terminology. for americans, the kinship terminology is com- posed of the terms such as son, daughter, brother, sister, mother, father, wife, husband, and so on that are used by one individual as a way of referring to the kin relation one has with another individual. it is important to make a distinction between terms of reference (the terms one uses when referring to another individual, such as in the expression from an advertisement for boys town in the s: “he ain’t heavy—he’s my brother.”) and terms of address that are used when talking to another individual. the latter is based on the former but may include terms or expressions that relate not simply to the kin relationship between the individuals in question but to patterns of speech that bring into play other aspects of the relationship between the individuals that is being conveyed by the way one person addresses another per- son, such as the use of honorific forms of speech. in american culture, we have a variety of social science computer review expressions that a child may use when addressing her or his male parent, such as “dad,” “pop,” “father,” “pa,” or even the person’s first name, with each mode of address having a dif- ferent connotation. for our purposes here, we will focus on terms of reference only. even when we limit ourselves to terms of reference, a complication enters through the fact that we have two distinct ways by which we may refer to someone’s kin relationship to speaker. one way is through kin terms, such as the expression, “this is my aunt sally.” the other is through specification of the genealogical relationship of the person in question to the speaker, as in the expression, “she is my mother’s sister.” sometimes both ways of referring to a kin person may occur sequentially, as in the expression, “this is my aunt sally—she is my mother’s sister,” depending on the information that the speaker wants to convey about her or his relationship to the person in question. more generally, corresponding to each kin term, there is one or more genealogical relation for which the kin term may be substituted. for some kin terms, the kin term and the genealogical relations may not be linguistically distin- guished. in american culture the expression, “she is my mother,” is ambiguous in that the term mother can either be the kin term mother or the genealogical relation mother. similar comments apply to the american or english kin terms father, brother, sister, son, and daugh- ter. this correspondence between kin terms and genealogical relations is not true of all termi- nologies. for many terminologies their term that we would transliterate as mother, for exam- ple, does not simply correspond to the genealogical relation mother but may also be properly used for other genealogical relations such as mother’s sister (mother’s daughter), mother’s mother’s daughter’s daughter, and so on. (a similar comment applies to the term we would transliterate as father.) for terminologies of this kind, the kin term mother is similar to kin terms such as uncle and aunt in the american or english kinship terminology in that the kin term can properly be used to refer to several, different genealogical relations. kinship terminologies generally cannot be faithfully translated from one language into another as the relationships that are recognized in one society may not have an exact counter- part among the relationships recognized in another society. terminologies also vary with respect to the specificity of a kin term with regard to genealogical relations. whereas the terms aunt and uncle in american culture do not distinguish whether the relation is through the siblings or spouse of siblings of father versus mother, other terminologies use different terms depending on whether or not the relationship is through one’s (genealogical) mother or father as can be seen by comparing the kin term map for the akt with the kin term map for the kinship terminology used by the shipibo indians of peru (see figure ). the basis for the differences in the kin term relationships that are recognized by different societies has been a source of contention in anthropological theorizing. equally problematic, though not as systematically addressed, has been identification of the basis for the particular relationships recognized within a given society. at one extreme is the presumption that the relationships linguistically identified through a kinship term are essentially biological in ori- gin and are to be understood as distinctions arising through the process of darwinian evolu- tion. the biological argument relates the distinctions made within the kinship terminology to reproductive success, taking into account kin selection, inclusive fitness, and so on. a typical kind of argument would be the one advanced for why, in many lineage-based societies, the brother of a sister has a close and authoritative role vis-à-vis his sister’s children and lines of inheritance may be from him to his sister’s son. the emphasis on sister’s brother rather than (putative) genetic father stems, it is argued, arises when paternity is uncertain because of the fact that extramarital affairs are not severely sanctioned. uncertainty of paternity does not read / kinship algebra expert system extend to his sister’s children (assuming he and his sister are genetic siblings), and so direct- ing his parenting toward his sister’s children when paternity uncertainty is sufficiently high increases his reproductive success. more precisely, because his genetic correlation with his genetic children is one half and his genetic correlation with his genetic sister’s children is one fourth, then if the certainty of paternity is less than one half, his genetic correlation with his sister’s children is greater than his likely correlation with his wife’s children (½ � probability of being the genetic father), and so parenting directed toward his sister’s children gives him greater reproductive success than does parenting directed toward his wife’s children. a more middle-of-the-ground viewpoint, and one that characterizes most anthropological theorizing about kinship systems, has assumed that the kin terms are based on the biological relationship of reproduction but given cultural expression that may deviate from their biolog- ical grounding (keesing, , p. ). a distinction is made between genetic father (the man who provides the sperm) and genetic mother (the woman who provides the ovum) and geni- tor (the man identified as the physical father but who need not be the man who provided the sperm) and genetrix (the woman identified as the physical mother but who need not be the woman who provided the ovum) because knowledge of genetics and its role in reproduction was not known until the work of gregor mendel in the mid- s. because of the obvious link between pregnancy and birth, in most societies the woman recognized locally as the genetrix (assuming such an identification is made) is the genetic mother; however, adoption social science computer review figure comparison of the american kinship terminology with the shipibo kinship terminology allows for a woman to be identified as mother even though she is not the genetic mother or the birth mother. in societies such as the inuit, a woman adopting a child is as real a mother as a woman giving birth to a child, and no distinction is made between a child to whom she has given birth and a child she has adopted. in american culture, we give the adopting woman the kin designation mother but recognize that she is not the physical or biological (or birth) mother. the assumption in this perspective is that reproduction and marriage are the basis for kin relations and that the former may have local, cultural constructions that need not match actual genetic linkages. it is assumed that all societies have a notion of genealogy grounded in reproduction and marriage—however, these may be locally constructed—and the genea- logical relations are the basis for asserting a kinship connection. kin terms are viewed as lin- guistic labels identifying the classification that the members of a society make of a genealogical space grounded in marriage and procreation. a third perspective, and the one underlying the kaes program, considers kin terms to constitute a system of interconnected concepts expressed through the conceptual connec- tions among the kin terms (read, ). the terminology has a structure in its own right that may be identified without first referring to another domain such as a genealogical space. the structure is based on the manner by which we calculate kin relations using kin terms without first translating kin terms to genealogical representations for the kin terms. in this perspec- tive, the set of kin terms is not just a set of linguistic symbols but also a set of linked symbols that forms the kinship terminology into a structured set of symbols. the structure expressed through the linkages among the kin terms is generative in that the structure can be generated from a few primitive kin terms and structural equations that express the logic underlying the structural form for the terminology. structural properties arise, in this perspective, through the logic by which the structure is generated from these primitive kin terms. thus, a kinship terminology is a structured set of symbols that cannot be fully understood without reference to the way the system is generated as a conceptual construct. in this perspective, genealogical tracing, which is the conceptual basis underlying a genealogical space, is also recognized as generating a conceptual system—the genealogical space—distinct from the kin term concep- tual system but connected to it through a mapping of the primitive kin terms onto the genea- logical relations used in genealogical tracing. although the previous perspective that posits kin terms to be semantic labels for the classes making up a classification of the genealogical space presumes a conceptual hierarchy going from a genealogical space to a kinship termi- nology, this third perspective views the genealogical space and the kinship terminology space to be two coexisting conceptual systems, each with a structure determined by the logic of how the space in question—genealogical or terminological—is generated and with a map- ping that connects the two spaces but that is necessarily from the kin terminology space onto the genealogical space. the mapping demonstrates that the classification made of the genea- logical space and the definition of kin terms using genealogical relations are both the consequence of a more fundamental set of concepts. hence the definitions of kin terms using genealogical relations and assumed to be primitive concepts in the second perspective are not primitive but instead are derived concepts. the differences among these three perspectives—biological, genealogical, and cultural— are subject to verification and falsification as each perspective has different implications for what constitutes a kinship terminology. first of all, the biological perspective implies that the kin relations identified through the kin terms making up a kinship terminology should bear a close resemblance to the actual biological relatedness of individuals and differences in termi- read / kinship algebra expert system nologies should relate to the pattern of reproductive success in the societies in question in accordance with terminological differences between kin terms and their referents. in this per- spective, there should be a close connection between what are effective patterns of social behavior with respect to reproductive success and normative patterns of interaction among kin distinguished through the kinship terminology. in brief, the biological perspective implies that the structure of a kinship terminology and its usage cannot be understood with- out reference to the biological relatedness of pairs of individuals, their behaviors, and the consequences of those behaviors for reproductive success. second, the genealogical perspective implies that a terminology can primarily be under- stood through the criteria used to classify the possible genealogical relations between pairs of individuals and how different classifications of genealogical relations arise in accordance with linguistic, psychological, and/or ecological constraints. kin relations in this perspective should be demonstrable as first being conceived genealogically and then given terminologi- cal expression according to the definition of kin terms using genealogical relations. the genealogical perspective implies that the structure of a terminology should reflect the criteria and constraints through which a classification of a genealogical space is formulated. in contrast to both of these perspectives, the cultural perspective implies that the terminol- ogy will have a structure whose features can be constructed or generated from a few funda- mental concepts along with a general process for the production of the structure of a kinship terminology. a key aspect of this perspective is the logic underlying the structure of the ter- minology. that logic implies that the kinship terminology should be a logically consistent structure, hence many of its features will be predictable from the logic of how the structure may be generated. in contrast, neither the biological perspective nor the genealogical per- spective imply that kinship terminologies should be highly logical in terms of how they con- stitute a symbolic system. hence analysis of a kinship terminology with respect to it being constituted in accordance with an underlying logic provides a basis for distinguishing among these three perspectives on what constitutes kinship in human societies. kinship terminology logic the kaes program is designed to explore the logic underlying the way the kinship terms form a structured system of symbols (with symbols taken in a linguistic and mathematical sense). but more than this, the program is simultaneously an idiom through which a theoreti- cal framework has been developed, a research tool aimed at resolving a fundamental question about a central feature of human societies, and an application program for the analysis of the structural properties of a kinship terminology and its relationship to the genealogical domain. the theoretical framework centers around working out the logic of kinship terminologies based on properties that are based on ethnographic observation. the key ethnographic obser- vation lies in the means by which individuals work out kin relationships. kin term product the structural modeling portion of the kaes program is based on the common ethnographic observation that individuals usually do not determine kin relations by identify- ing their biological pedigree or their genealogical connectedness but by doing computations with kin terms. as the anthropologist marshall sahlins ( ) has noted with regard to moala kinship: social science computer review [kin] terms permit comparative strangers to fix kinship rapidly without the necessity of elabo- rate genealogical reckoning—reckoning that typically would be impossible. with mutual rela- tionship terms all that is required is the discovery of one common relative. thus, if a is related to b as child to mother, veitanani, while c is related to b as veitacini, sibling of the same sex, then it follows that a is related to c as child to mother although they never before met or knew it. kin terms are predictable. if two people are each related to a third, then they are related to each other. (p. , italics added) based on ethnographic observations similar to sahlins’ comment about moala kinship, we may define a kin term product informally as follows. suppose one person (ego) properly refers to a second person (alter ) by the kin term k, and alter properly refers to a third person (alter ) by the kin term l. the product of the kin terms k and l will be the kin term, m, if any, that ego properly uses for alter . for example and with respect to the american or english kin- ship terminology, when ego refers to alter by the kin term aunt and alter refers to alter by the kin term daughter, then ego (properly) uses the kin term cousin to refer to alter . so the product of the kin terms daughter and mother is the kin term cousin in the akt (see figure ). we need to make a clear distinction here between kin types and kin type products, which have to do with genealogical relations, and kin terms and kin term products, which have to do with the conceptual system we refer to as a kinship terminology and the structure formed from read / kinship algebra expert system figure illustration of a kin term product for the kin terms aunt and daughter from the american kinship terminology note: the product of these kin terms yields the kin term cousin. relations linking kin terms via kin term products. kin term products express the ways in which the kin terms (viewed as linguistic symbols) form a structured set of symbols. in contrast, kin type products refer to the ways in which a genealogical relation can be constructed through concatenation of pairs of genealogical relations, such as the expression, “mother’s brother’s daughter.” in the context of genealogy, the words mother, brother, and daughter refer to gene- alogical tracing based on genitors and genitrixes. genealogical mother in american culture is the woman who begat ego, and if her identity is unknown, then genealogical tracing through a female parent cannot proceed. this contrasts (in american culture) from kin relations through adoption that are excluded for purposes of genealogical tracing. in both cases, the same kin term, mother, may be used for purposes of reference. thus, if ego has been adopted, then ego has a mother from the perspective of the akt but not a genealogical mother from the perspective of kin types. anthropologists have used a technical distinction between social mother, or mater, and physical mother, or genetrix, to distinguish these two situations. the adopted mother, in american culture, is a social mother but not a genetrix. however, this dis- tinction is culturally salient only to the extent that it is a culturally recognized distinction. for some groups such as the inui mentioned above, no distinction is made between genetrix and mater (or genitor and pater), and so a social mother is also a genetrix for the inuit. from the perspective of kin terms, the matter is simpler because either the person in ques- tion is one for whom a kin term of reference is appropriate or not according to cultural rules regarding the usage of kin terms. in american culture, the kin term mother applies equally to a genetrix or to a social mother. mother (in american culture) is thus a polysemic term that can have the kin type, genealogical mother, as its reference, that can have the social mother as its reference in the case of adoption, or that can be taken as a kin term with linkages to other kin terms through kin term products. when necessary, a distinction between a kin type versus a kin term will be made by capitalizing the kin term and using lower case when referring to a kin type. for kin terms, the kind of calculation discussed by sahlins ( ) provides the conceptual basis for viewing kin terms as forming a structured set of symbols. more formally, we define a kin term product as follows: definition: let k and l be kin terms in a given kinship terminology, t. let ego, alter , and alter refer to three arbitrary persons, each of whose cultural repertoire includes the kinship terminol- ogy t. the kin term product of k and l, denoted k o l, is a kin term, m, if any, that ego may (prop- erly) use to refer to alter when ego (properly) uses the kin term l to refer to alter and alter (prop- erly) uses the kin term k to refer to alter . right-to-left product notation will be used as it permits reading the expressing k o l as “k of l” for a kin term product, and this contrasts with the reading of kl as “k’s l” for the genea- logical (kin type) product kl of kin types. for example, the symbolic expression s o m = a, where m stands for mother, s for sister, and a for aunt, would be read “sister of mother is aunt,” whereas the kin type product ms would be read “mother’s sister” and aunt would be the kin term that can be applied to a person who is one’s ms. kin term map kin term products are the bases on which the analysis of a kinship terminology structure proceeds. one part of the kaes program deals with representing the structure of a kinship social science computer review terminology based on kin term products. a diagram displaying the structure for kin terms based on kin term products is called a kin term map and is the beginning point for the analysis of the structure of a kinship terminology. figure shows a kin term map for the akt based on taking products with the kin terms parent, child, and spouse. a variant of this kin term map can be constructed using kin term products with the kin terms mother, father, son, daughter, wife, and husband. in general, more than one kin term map may be possible for the same terminology. generating terms although the kin term product is defined for any pair of kin terms, we quickly discover that certain kin terms play central roles in constructing a kin term map. for the akt, these are terms such as mother, father, and parent and their reciprocals daughter, son, and child. other kin terms in the akt can be constructed by taking products of these kin terms. for example, the kin term grandfather = father o parent (read, “grandfather is father of parent,” or more completely, “the kin term grandfather results from taking the kin term product of the kin term father with the kin term parent”), where this is a statement about the three symbols, grandfather, father, and parent. the symbol o between father and parent indicates that we have a binary product (namely the kin term product) that maps a pair of symbols (in this instance, the father, parent pair) to another symbol (in this instance, the symbol grandfather). read / kinship algebra expert system figure kin term map for the american kinship terminology based on taking products with the kin term parent and its reciprocal term child and the kin term spouse though this equation may look like a genealogical statement, nothing could be further from the meaning of this equation. the equation states that when ego properly refers to a per- son as parent (and regardless of the actual genealogical connection, if any, between the two persons) and that person properly refers to a third person as father (again, regardless of the actual genealogical connection between them, if any), then ego would properly refer to that third person by the kin term grandfather (again, regardless of the actual genealogical rela- tionship between them). some of the persons in question might be adopted and so the actual genealogical relations that are involved are unknown; nonetheless the users of the akt still know the proper usage of the terminology in cases of adoption by virtue of the conceptual relations that hold among the kin terms. criteria for generating terms the kin term map is constructed by first deciding on the kin terms that will serve as the generating terms through which kin terms are linked to one another. the generating terms must account for the closest kin term linkages between a person and another person, and so they will likely be some variant on the kin terms that can be transliterated as mother, father (and possibly parent), and their reciprocals, daughter, son (and possibly child). in addition, terms that can be transliterated as brother or sister (or possibly older brother, younger brother, etc.) will be generating terms in some terminologies but not in other terminologies. in the akt, the analysis of the logic underlying the structural form of the kin term map dis- played in figure establishes that the kin term brother is the product of the kin terms parent and son: brother = son o parent. in other terminologies, a kin term that can be transliterated as brother (or sister) or possibly as older brother (or older sister) is a generating term and not expressible as a kin term product. the distinction between a sibling term taken as a product versus a sibling term taken as a generating term provides the conceptual basis for the distinc- tion between so-called descriptive terminologies versus classificatory terminologies (see below). kin term map as a generated structure the kin term map, even without any additional analysis, serves as an effective graphical device for displaying the structural differences among kinship terminologies. each terminol- ogy has its own structural form (or variant on a structural form for closely related terminolo- gies) that expresses differences in the ways in which the kin terms of a terminology form a structure. the kin term map is more than just an effective descriptive means to display the conceptual structure formed by the kin terms in a kinship terminology. an obvious question is whether the map can be generated from a few simple structural properties, and if so, does this account for the features of the terminology (such as the odd way in which the aunt and uncle terms in the akt are used both for individuals linked genealogically through parent and child links to ego and for individuals who are linked to ego only through marriage). if the kin term map is the graph of a generative structure, then it follows that a kinship terminology is neither simply a representation of biological kin relations nor a classification of a genealogical space by cri- teria that lie outside of the kinship terminology. instead the kinship terminology is a cultural construct, and the meaning of kin terms cannot be fully understood without reference to that construct and the logic underlying how it is generated. social science computer review the kaes program is designed to have a number of features aimed at helping construct a kin term map. kin terms may be entered in any order and at any time. any kin term that has been entered may be designated as a generating term, and the pattern of arrows that may be drawn showing the kin term product of a kin term with a generating term will be constrained by the features of the generating term such as its sex marking and whether it is an ascending term (i.e., refers to alters in an older generation with respect to ego). generating terms: descriptive versus classificatory terminologies one of the key discoveries about terminology structures that has been obtained to date relates to the distinction morgan ( ) made between what he called descriptive terminolo- gies versus classificatory terminologies. the choice of labels was unfortunate as it led to sub- stantial debate about the validity of the distinction within anthropological theorizing about kinship terminological systems based on the labels descriptive versus classificatory rather than the structural differences that distinguish the two kinds of terminologies (e.g., lowie, ). at the level of kin terms, some kin terms in a terminology may be descriptive in the sense that the kin term can be given a genealogical definition using a single kin type (such as mother is the term used in reference to the kin type mother), and other terms in the same ter- minology may be classificatory in the sense that a genealogical definition of the kin term will refer to more than a single kin type, such as uncle with a genealogical definition: mother’s brother, father’s brother, mother’s sister’s husband or father’s sister’s husband. the distinc- tion morgan was attempting to identify had to do with structural properties such as whether or not lineal relations traced using either parent kin types only or child kin types only were kept distinct from collateral relations traced using first parent kin types and then child kin types. thus in the akt, parent is a lineal kin term, whereas aunt or uncle are collateral kin terms, and the collateral kin types to which the later two kin terms refer are kept distinct from the lineal kin type to which the kin term parent refers. in classificatory kinship terminologies that are common in polynesia and other groups inhabiting islands in the pacific ocean, parent and parent’s siblings are not kept distinct, and a single kin term, mother, refers to mother and mother’s sister (and other females on the same generation as mother), and a single kin term, father, refers to father and father’s brother (and other males on the same generation as father). morgan’s ( ) structural distinction based on lineal versus collateral only partially identifies the basis on which the lineal-collateral distinction arises. by working out the gener- ative implications of taking a sibling term as a primitive during the process of writing the kaes program, it became evident that the structural distinction morgan was attempting to identify relates to whether or not the sibling kin terms are generating kin terms or are products of other kin terms. as noted above, in the akt, the kin term brother is constructed via brother = son of parent (which can also be expressed via the two equations brother = son of father or brother = son of mother). the kin term map for the akt is not consistent with a claim that the sibling terms are generating terms for the terminology, and so the akt is a descriptive termi- nology. in contrast, classificatory terminologies such as the trobriand terminology or the ton- gan terminology in the pacific ocean area are found to be terminologies for which sibling terms are generating terms. the analysis of the structure of these terminologies demonstrates that the kind of genealogical equations that have been used to define classificatory terminolo- gies, namely f = fb and m = mz, are a consequence of the structural implications that arise from including the sibling kin terms among the generating terms for the terminology. read / kinship algebra expert system algebraic model for a kin term map the algebra part of kaes refers to the use of algebraic concepts to model the structure of a kinship terminology as displayed through a kin term map. by algebraic structure is meant the structure that is formed by a set of symbols (e.g., the kin terms making up a terminology), a binary product defined for all pairs of symbols in that set (e.g., the kin term product for kin terms), including whether or not the binary product is associative—for example, for the binary operation, +, the sum a + (b + c) is the same as the sum (a + b) + c; for kin terms, the product k o (l o m) = (k o l) o m—or commutative (e.g., for the binary number operation, +, a + b = b + a; however, commutativity does not hold for most kin term products; e.g., son of father is not the same as father of son), and certain equations that the binary product must sat- isfy. the equations give an algebra its particular structural form. for kinship terminologies, we have the following basic algebraic machinery: . a set of symbols—the terms making up the kinship terminology. . a binary product defined over those symbols—the kin term product, including as a symbol indicating that the product does not yield a kin term; e.g., father o father-in-law = in the akt because there is no kin term representing the product father o father-in-law. . the binary product is associative—the kin term product is associative, except in certain special cases that are dealt with in the analysis on an individual basis. an algebraic system satisfying all three of the above is known in the mathematical litera- ture as a semigroup. the equations that are used when modeling a kinship terminology are used to introduce structural properties that characterize kin terms. the structural equations provide a way to express the structural properties of a kind of kin term with respect to the kin term map and thereby provide a formal definition for what otherwise are informal concepts associated with terminologies. these properties provide definitions for kinds of kin terms such as: (a) a sib- ling kin term, s, is a kin term for which s o s = s (e.g., in american culture, brother of brother is brother as a kin term product); (b) a spouse kin term, sp, is a kin term for which sp o sp = self (with self a symbol representing the concept of self, a concept fundamental to any kin- ship system), and so on. the equation s o s = s, for example, follows from the fact that as users of our own terminology, we know what the kin term brother means. among the various meanings associated with the term is the structural property that if ego properly refers to alter as brother and if alter properly refers to alter as brother, then ego properly refers to alter as brother, hence the structural equation s o s = s becomes the way to structurally define a kin term as a symbol that can be given instantiation with the concept of brother (or equiva- lently, with the concept of sister because sister of sister is sister). the structural equation does not provide an exhaustive definition of the meaning of, say, the term brother, as the instantiation (see below) of an (abstract) symbol may include properties that are not part of the kin terminology structure and bring to bear other cultural aspects associated with cultural constructs relevant to the cultural notion of what constitutes (in this example) brotherhood. the structural equations also express relations between kin terms such as reciprocity of kin terms. by reciprocity of kin terms is meant that if ego properly uses the kin term k to refer to alter, then the kin term l that alter properly uses to refer to ego is the reciprocal kin term for k. because of sex marking of kin terms, a kin term need not have a unique reciprocal term. the akt kin term mother, for example, has both son and daughter as reciprocal kin terms. social science computer review the structural property that relates two kin terms k and l as reciprocal kin terms is either the structural equation, k o l = self or l o k = self, depending on whether k or l is the ascending kin term. if k is the ascending kin term, then k o l = self is the reciprocal equation that defines l as a reciprocal kin term for k. the notion of reciprocity is almost equivalent to the notion of an inverse element for algebraic structures in that in one direction—first using a descending term and then an ascending term—reciprocity refers back to ego, and self is the concept used by ego to refer to herself or himself. in the other direction, the product does not refer back to ego; otherwise the only kin terms, expressed genealogically, would be lineal kin terms. thus, for the reciprocal kin terms father and son in the akt, father of son is self as a kin term product, but son of father is not self and instead identifies a new kin term, brother. the kaes program generates these structural equations that express the structural properties of kin terms automatically as the analysis progresses. general model for the structure of a kinship terminology the following is an analytical sequence for generating a kinship terminology that provides the basic framework for the algebraic analysis of a kin term map. the algebraic portion of the kaes program is structured around this analytical sequence. . a structure isomorphic to a sequence of ascending kin terms, all with the same sex marking, is constructed based on an ascending generating term such as mother if the terms are to be marked female, father if the terms are to be marked male, and parent if the terms are to be neutral (see figure ). included in this ascending structure is an identity element that can be designated male self, female self, or self, depending on whether the ascending terms are marked male or female or are neutral. by an identity element is meant an algebraic element i such that for any element k, k o i = i o k = k. the identity element will be designated by the symbol self (with or without a sex marker) on the grounds that a concept of self is both universal and necessary for there to be a kinship terminology; that is, all terminologies must make a distinction between self and other, whether or not there is a linguist symbol representing the concept self in the kinship terminology. . a structure of descending kin terms is constructed from the structure of ascending kin terms by making an isomorphic copy of the ascending kin term structure and then a larger algebraic structure made up of both ascending kin terms and descending kin terms, and cross products between ascending and descending kin terms are formed (see figure ). included here is the structural equation that makes an ascending kin term—call it p—and its isomorphic kin term— call it c—into reciprocal terms, namely p o c = i. . the next step introduces kin terms with the other sex marking or, alternatively, bifurcation of neutral kin terms into male-marked and female-marked kin terms. first, when the kin terms constructed in steps and are already sex marked, then terms of the opposite sex are con- structed by taking an isomorphic copy of the structure generated in steps and . the isomor- phic copy will be a structure of kin terms with the opposite sex. the original structure and this isomorphic structure are combined together, along with appropriate structural equations, to make a new, larger structure consisting of terms marked with both sexes. second, when the kin terms constructed in steps and are all marked neutral, each kin term is bifurcated into two kin terms marked as male and female terms. appropriate structural equations that relate male kin term and female kin term products are introduced as appropriate. read / kinship algebra expert system . affinal kin terms are introduced next, along with appropriate equations that express the proper- ties of affinal kin terms. (see the double-headed, gray arrows in figure that represent taking products with the spouse generator. figure is the structure produced from steps to .) . rules are identified that refer to ways that portions of the overall structure are locally modified. for example, in the akt, not all kin terms are marked as either male terms or female terms according to the sex marking rule: “a kin term, k, is marked neutral when the product of k with a spouse term is not a kin term.” another rule in the akt identifies the way in which cousin kin terms may be further distinguished according to properties such as “ith cousin” or “ith cousin j- times removed.” these rules relate to ways in which parts of the overall structure are modified for cultural reasons and are not modifications necessary for a terminology to have the structure of a kinship terminology (see the shape of the cousin space in the upper right portion of figure ). criterion for the validity of a structural model an algebraic model for a kinship terminology is a valid model for that terminology if and only if the algebraic model is isomorphic to the kin term map representation of the kinship terminology. not all conceivable kin term maps can be validly represented by an algebraic model. success in formulating a valid algebraic model for a kinship terminology map identi- fies the means through which the terminology structure can be logically generated in accor- dance with the general model for the construction of a kinship terminology. the kin term map social science computer review figure an ascending structure based on the generator, p, and the identity element, i. note: the structure will be structurally defined as an ascending structure by the structural properties that the gen- erating term, p, will satisfy when an isomorphic copy of this structure becomes the descending structure (see figure ). shown in figure is structurally isomorphic to the algebraic structure shown in figure (read & behrens, ), and so the algebraic construction has generated a valid algebraic model for the kin term map of the akt. properties of the algebraic model are properties that arise from the logic of how a kinship terminology can be generated and do not require reference to properties outside of the termi- nology. for example, the algebraic model that validly represents the akt has the property, spouse o aunt = uncle and spouse o uncle = aunt, hence the lack of terms marked with the suf- fix -in-law such as aunt-in-law or uncle-in-law for these relations by marriage is a conse- quence of the logic underlying the generation of the akt as a kinship terminology structure. in figure , the double-headed, gray spouse arrow maps the uncle node onto the aunt node as a consequence of the logic of how the structure displayed in figure is generated. rules that are introduced in step identify structural properties that are not necessary for the structure to have the structural form of a kinship terminology; hence they are properties that may have been introduced for reasons extrinsic to the terminology structure per se. in other words, the analysis distinguishes between properties of the terminological structure that are necessary for the structure to be complete and consistent as a kinship terminology structure and properties that are parts of the terminology but are not necessary for its com- pleteness and consistency and hence are likely to have origin for reasons extrinsic to the generative logic of the kinship terminology structure. instantiation of the symbolic structure the algebraic model of a kinship terminology structure consists of abstract symbols and the relations among these symbols as determined by structural equations and other properties guiding the generation of the symbolic structure. these abstract symbols are culturally instantiated (see figure ; see read, , for a discussion of cultural instantiation) with read / kinship algebra expert system figure an isomorphic copy of the ascending structure, with generating terms c and i, is constructed and joined to the ascending structure, including products between terms in the ascending structure and the descending structure note: the structural equation, p o c = i, is introduced to structurally define p and c as reciprocals. the structural equation also structurally identifies p as an ascending generating term and c as a descending generating term because an equation defining one term as the reciprocal of another term is of the form, ascending term � descend- ing term = self. respect to the individuals for whom the symbols may be properly used as terms of reference when using the kinship terminology to identify kin relationships between individuals. a pri- mary cultural instantiation is through genealogical concepts and is obtained by identifying the instantiation of the generation terms in the algebraic structure with kin types underlying the construction of a genealogical space. a male-marked kin term used in the construction of the ascending structure (e.g., it is instantiated with the kin type, father; more precisely, with genealogical father). the reciprocal male-marked kin term in the descending structure is instantiated with the kin type genealogical son and so on. predicted genealogical definitions of kin terms once the generating terms have been instantiated with genealogical kin types, then the instantiation of all other terms can be worked out from the instantiation of the generating kin social science computer review figure the algebraic structure generated from the structure in figure after steps to are implemented note: the solid, single-headed arrows show the result of taking a product with the generator, p. the dashed, single- headed arrows show the result of taking a product with the reciprocal generator, c. the gray, double-headed arrows show the result of taking a product with the spouse generator, s. the oval around a pair of nodes indicates a node from the structure produced in step (see figure ) that was bifurcated into a pair of sex-marked nodes in step . the gray nodes are the affinal (relations by marriage) nodes introduced in step when a spouse-generating ele- ment was added to the structure. the gray nodes are labeled either with kin terms having an in-law suffix or are labeled with the spouse terms, husband and wife, when the algebraic structure is isomorphically mapped onto the kin term map shown in figure . terms and the algebraic structure by identifying algebraic products with products of sets of kin types. for example, grandfather has the instantiation (genealogical father) � (genealogi- cal father) = (genealogical father’s father) because grandfather = father o father and father has the instantiation (genealogical father). this process of instantiation of the generating terms and construction of the instantiation of all other kin terms based on the algebraic structure leads to a predicted set of genealogical definitions of kin terms. for all terminologies considered to date, the predicted genealogical definitions match with % accuracy the genealogical definitions elicited by anthropolo- gists (see figure ). these genealogical definitions have been assumed to be the primary and irreducible data for the analysis of kinship terminologies. the kaes analysis demonstrates to the contrary that these definitions are predictable from the way in which the terminology is generated as a structure. instantiation is not limited to genealogical instantiation. instead, instantiation depends on cultural rules for mapping abstract symbols to individuals and can be based on other criteria read / kinship algebra expert system figure diagram showing the general process of cultural instantiation of the algebraic structure note: the nodes in the algebraic structure have semantic meaning with respect to the structure in which they are embedded and are then given semantic meaning with regard to individuals to whom the kin terms are applicable through cultural rules of instantiation. one set of rules relates the algebraic structure to a genealogical space by mapping the generating terms for the algebraic structure onto the genealogical space and using the logic of the algebraic structure to determine the mapping of all other nodes in the algebraic structure onto the genealogical space. such as adoption or other means by which individuals are incorporated within the conceptual structure of a kinship terminology. metaphoric instantiation of kin terms the constraint imposed by the algebraic structure on instantiation lies in the structural properties of a kinship terminology that are mapped onto individuals through the process of instantiation, such as instantiation must be consistent with the reciprocity of kin terms and the generative properties of the kinship terminology structure. the latter implies that kin term products must also map over to individuals under the cultural rules of instantiation. thus an adopted child is not only a child vis-à-vis the adopting parents, but the adopted child has the appropriate kin term relationship to other individuals in the kinship domain of the adopting parents. this contrasts with instantiation of individual kin terms that does not carry over to other kin terms, such as the use of the kin terms aunt and uncle by users of the akt in an hon- orific sense (e.g., the use of the uncle and aunt terms for the friends of one’s parents). the dis- tinction between the more extended and the more limited forms of instantiation provides a social science computer review figure the mapping of the american or english kinship terminology kin terms onto the genealogical space as predicted from the algebraic analysis performed by the kinship algebra expert system program note: all of the predictions match the genealogical definitions of kin terms for the american or english kinship terminology. more formal way to express what has sometimes been referred to as metaphoric usages of kin terms (scheffler, ) and distinguishes when instantiation is metaphoric and not primarily for the usage of the kinship terminology. the genealogical perspective on kin terms assumed, without ethnographic justification, that the primary meaning of kin terms is with respect to kin types, and so any application of kin terms outside of the genealogical domain or where the genealogical definition included noncomparable kin types became metaphorical extension. classificatory terminologies, in this viewpoint, had kin terms with translation mother and father whose primary meanings were genealogical mother and genealogical father, respec- tively, and the usage of the same kin terms for mother’s sister and father’s brother (and other females and males) was taken to be metaphoric extension of the primary, genealogical mean- ing of the kin terms. the analysis developed through writing the kaes program construc- tively shows that the metaphoric extension hypothesis is unnecessary and was devised as a way to accommodate an unwarranted assumption about the nature of kinship terminologies to ethnographic facts that were not in accord with that assumption. there is no need to posit metaphoric extension as there is no need to posit extension from a primary, genealogical meaning of kin terms. the analysis produced by the kaes program constructively demon- strates how the range of kin types included under a kin term arises from the mapping of the generating kin terms to the genealogical domain and the structural equations that determine the structural form for the kinship terminology structure. there is no metaphoric extension as the full range of kin types included under a kin term is generated as part of the logic of the kinship terminology structure and how that structure relates to the genealogical domain. conclusion the kaes program has successfully implemented a theory of kinship terminology struc- tures that makes evident the underlying logic giving a terminology its particular form and accounts for the way the kin terms relate to genealogical relations. writing the kaes pro- gram has been central to working out the logic of kinship terminology structures and, through that logic, the broader implications for what we understand to constitute cultural constructs. the program clarifies the way in which we have two conceptual systems regarding kin rela- tions—the manner in which individuals are connected to one another through genealogical connections and the kin terms that identify the kin relation one individual has to another indi- vidual. the biological argument cannot account for the logic modeled through the kaes program; hence it is insufficient to consider kinship, as it occurs in human societies, through biological criterion alone. human societies are based on what we can call cultural kinship in contrast to biological kinship. this is not to say that biological kinship is irrelevant; cultural kinship has its evolutionary origins in biological kinship but has transformed those origins. similarly, the received view in anthropology that has viewed kinship as arising from mar- riage and reproduction is not so much incorrect as inadequate. historically, some kind of pro- cess of genealogical tracing of connections must have preceded the “invention” of the con- ceptual structures we refer to as kinship terminologies. one possibility for the evolutionary origin of kinship terminology systems is that they provided a computational device for calcu- lating relations among individuals that was cognitively less demanding than memorizing extensive genealogical linkages. but even if it arose as a computational system, the computa- tional facility enabled the terminology to become the central means by which individuals determine if they have a kinship connection or not as indicated in the quote from sahlins read / kinship algebra expert system ( ) given above. one can calculate kin relations with the kinship terminological system without necessarily referring back to genealogical linkages, and hence the relative impor- tance of genealogical connections versus kin terminology connections has become society specific. some australian groups, for example, find it difficult to compute genealogical linkages as it is not a familiar way for determining one’s kinship status vis-à-vis another individual (shapiro, ). the kaes program is a work in progress and will be modified to accommodate features of kinship terminologies not encompassed by the terminologies considered to date as additional terminologies are analyzed. the structure used to implement the algebraic modeling is not likely to be changed. however, if there is a terminology with structure that cannot be accom- modated within the structural sequence implemented in the kaes program, then that fact alone will be a striking discovery. the kaes program will also be expanded beyond its cur- rent architecture aimed at analyzing and modeling the structure of kin term maps. the instantiation of the algebraic model with respect to genealogical tracing provides the inter- face needed to embed the structural logic into an agent-based demographic model so as to follow out the logic of kinship terminology structures and their instantiation in terms of a group of individuals undergoing demographic change. initial work in this direction has high- lighted the way in which implementation of an agent-based demographic model raises ques- tions about the constraints on the transmittal of information of kinship relations within a group to progeny. for example, if progeny learned in totality the set of parental kin relations, then the amount of information about kin being transmitted across generations expands at least exponentially and will outstrip memory capacity. hence the amount of information transmitted must be limited, thereby raising the issue of how kin relations can be calculated for more distant potential kin and answered through the logic modeled by the kaes pro- gram. coupling the kaes program with agent-based demographic simulation should make it possible to examine in more detail the consequences of marriage rules in conjunction with terminologies and how marriage rules and kin terminology structure relate to social organization in kin-based societies. the logic of terminologies raises questions about human cognition and why terminologies are highly logic. in part it relates to calculation of kin relations in a manner consistent across all individuals sharing the same terminology. in another part, it raises questions about the evolutionary trajectory that leads from behavior centered around biological kin to behavior centered around cultural kin and whether that trajectory can be accommodated in a darwin- ian evolutionary framework or whether it signals a change to social organization based on culturally constructed conceptual systems for which the idiom of descent with modification driven by reproductive success is no longer the primary driver for evolutionary change (cf. read, ). references bennardo, g., & read, d. (n.d.). the tongan kinship terminology: insights from an algebraic analysis. unpub- lished manuscript. keesing, r. ( ). kin groups and social structure. new york: holt, rinehart & winston. leaf, m. ( ). the punjabi kinship terminology as a semantic system. american anthropologist, , - . lee, r. ( ). dobe ju/‘hoansi. fort worth, tx: harcourt brace. lounsbury, f. ( ). the structural analysis of kinship semantics. in h. hunt (ed.), proceedings of the ninth international congress of linguists (pp. - ). the hague, the netherlands: mouton. social science computer review lowie, r. h. ( ). a note on relationship terminologies. american anthropologist, , - . morgan, l. h. ( ). systems of consanguinity and affinity of the human family. washington, dc: smithsonian institution. read, d. w. ( ). an algebraic account of the american kinship terminology. current anthropology, , - . read, d. w. ( ). what is kinship? in r. feinberg & m. ottenheimer (eds.), the cultural analysis of kinship: the legacy of david schneider and its implications for anthropological relativism (pp. - ). urbana: uni- versity of illinois press. read, d. w. ( ). cultural construct + instantiation = constructed reality. human complex systems (paper dwr ). retrieved month, day, year?, from http://repositories.cdlib.org/hcs/ dwr read, d. w. ( ). from behavior to culture: an assessment of cultural evolution and a new synthesis. complex- ity, ( ), - . read, d. w., & behrens, c. ( ). kaes: an expert system for the algebraic analysis of kinship terminologies. journal of quantitative anthropology, , - . read, d. w., & fischer, m. d. ( ). kinship algebra expert system. retrieved august , , from http:// kaes.anthrosciences.net sahlins, m. ( ). moala: culture and nature on a fijian island. englewood cliffs, nj: prentice hall. scheffler, h. w. ( ). kinship semantics. annual review of anthropology, , - . shapiro, w. ( ). the place of cognitive extensionism in the history of anthropological thought. journal of the polynesian society, , - . dwight w. read, ph.d., is a professor in the department of anthropology and the department of statistics at the university of california, los angeles. his research interests relate to the use of mathematical modeling as a way to extend and elaborate anthropological theorizing about the structure of human societies and the evolutionary origins of culture. read / kinship algebra expert system [pdf] ergonomic risk assessment based on a bayesian-optimised expert system | semantic scholar skip to search formskip to main content> semantic scholar's logo search sign increate free account you are currently offline. some features of the site may not work correctly. doi: . /ijhfms. . corpus id: ergonomic risk assessment based on a bayesian-optimised expert system @article{dansie ergonomicra, title={ergonomic risk assessment based on a bayesian-optimised expert system}, author={chris dansie and r. sesek and d. bloswick}, journal={international journal of human factors modelling and simulation}, year={ }, volume={ }, pages={ } } chris dansie, r. sesek, d. bloswick published engineering international journal of human factors modelling and simulation improper application of an ergonomic analysis tool increases the likelihood of high-risk jobs not being detected, thus jeopardising worker’s health. likewise, significant time and cost may be incurred by redesigning jobs improperly identified as high risk. utah intelligent data driven ergonomic assessment system (iddeas) is an expert system that aggregates the outputs of multiple analysis tools to create a more predictive ergonomic analysis tool. rules in the expert system were optimised by… expand view via publisher inderscience.com save to library create alert cite launch research feed share this paper citations view all figures and tables from this paper table figure figure table table table table table table view all figures & tables one citation citation type citation type all types cites results cites methods cites background has pdf publication type author more filters more filters filters sort by relevance sort by most influenced papers sort by citation count sort by recency an application of bayesian regression in ergonomics ashutosh kumar singh, r. dalpatadu computer science save alert research feed references showing - of references rula: a survey method for the investigation of work-related upper limb disorders. l. mcatamney, e. nigel corlett engineering, medicine applied ergonomics , pdf view excerpt, references methods save alert research feed the strain index: a proposed method to analyze jobs for risk of distal upper extremity disorders. j. s. moore, a. garg medicine american industrial hygiene association journal view excerpt, references methods save alert research feed a functional job evaluation technique, in ergonomics occupational medicine: state of the art reviews, nonfatal occupational injuries and illnesses requiring days away from work , – table [ online ] evaluation and refinement of ergonomic survey tools to evaluate worker risk of cumulative trauma disorders [dissertation], university of utah, salt lake city, ut related papers abstract figures and tables citations references related papers stay connected with semantic scholar sign up about semantic scholar semantic scholar is a free, ai-powered research tool for scientific literature, based at the allen institute for ai. learn more → resources datasetssupp.aiapiopen corpus organization about usresearchpublishing partnersdata partners   faqcontact proudly built by ai with the help of our collaborators terms of service•privacy policy the allen institute for ai by clicking accept or continuing to use the site, you agree to the terms outlined in our privacy policy, terms of service, and dataset license accept & continue recommender system based on pairwise association rules expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications journal homepage: www.elsevier.com/locate/eswa recommender system based on pairwise association rules timur osadchiy a , ∗, ivan poliakov a , patrick olivier a , maisie rowland b , emma foster b a open lab, school of computing, newcastle university, newcastle upon tyne, united kingdom b institute of health and society, newcastle university, newcastle upon tyne, united kingdom a r t i c l e i n f o article history: received april revised july accepted july available online august keywords: association rules cold-start problem data mining ontologies recommender systems a b s t r a c t recommender systems based on methods such as collaborative and content-based filtering rely on ex- tensive user profiles and item descriptors as well as on an extensive history of user preferences. such methods face a number of challenges; including the cold-start problem in systems characterized by ir- regular usage, privacy concerns, and contexts where the range of indicators representing user interests is limited. we describe a recommender algorithm that builds a model of collective preferences indepen- dently of personal user interests and does not require a complex system of ratings. the performance of the algorithm is analyzed on a large transactional data set generated by a real-world dietary intake recall system. © the authors. published by elsevier ltd. this is an open access article under the cc by license. ( http://creativecommons.org/licenses/by/ . / ) a t l d b u p e f s u e a c r t s u n k m f i s p p m a i u fi i c w n m i b m ( c & h . introduction recommender systems aim to identify consumer preferences nd accurately suggest relevant items (e.g. products, services, con- ent). they are used in various application domains, including on- ine retail, tourism and entertainment ( covington, adams, & sargin, ; linden, smith, & york, ). widely adopted recommen- ation techniques often utilize collaborative filtering or content- ased recommendation methods ( pazzani & billsus, ). collaborative filtering produces recommendations based on ser preference models that are generated from explicit and/or im- licit characteristics and metrics corresponding to user interests. xplicit indicators normally imply users assigning ratings to items; or example, to products viewed in, or purchased from, an online tore. examples of implicit indicators include the amount of time sers spend interacting with content (e.g. watching a video) or lev- ls of interaction (e.g. scroll offset of a web page containing an rticle they are reading). items that are positively rated or pur- hased by consumers with similar preference models are used as ecommendations for target users. user similarity can be expressed hrough correlations in purchasing history or ratings given to the ame products, which can be further amplified with demograph- ∗ corresponding author at: open lab, school of computing, newcastle university, rban sciences building, science square, science central, newcastle upon tyne, e tg, united kingdom. e-mail addresses: t.osadchiy@newcastle.ac.uk (t. osadchiy), ivan.polia ov@newcastle.ac.uk (i. poliakov), patrick.olivier@newcastle.ac.uk (p. olivier), aisie.rowland@newcastle.ac.uk (m. rowland), emma.foster@newcastle.ac.uk (e. oster). u u c m o a t ttps://doi.org/ . /j.eswa. . . - /© the authors. published by elsevier ltd. this is an open access article u cs (e.g. age, gender, occupation). content-based filtering identifies imilarities between items based on a set of their descriptors (e.g. urpose of an item, author, artist, keywords). items similar to those ositively rated or purchased by the target user, are used as recom- endations. collaborative and content-based filtering recommender systems re therefore heavily dependent on extensive user- or item- profile nformation and are most effective when there is a rich history of ser preferences or behavior. sparse data sets and lean user pro- les typically result in low-quality recommendations or an inabil- ty to produce recommendations at all. this is referred to as the old-start problem, where new users are added into the system ith empty behavior profiles or new items are added that have ot been reviewed or rated by anyone ( shaw, xu, & geva, ). any solutions to the cold-start problem have been considered, ncluding hybrid methods that combine collaborative and content- ased filtering ( schein, popescul, ungar, & pennock, ), and ethods that aim to predict user preferences from demographics lika, kolomvatsos, & hadjiefthymiades, ), or knowledge of so- ial relationships ( carrer-neto, hernández-alcaraz, valencia-garcía, garcía-sánchez, ). there are, however, a number of application contexts where sers interact anonymously; for example, online shops where an nregistered user browses and adds products to their basket to heck out later. in other contexts, such as email recipient recom- endation ( roth et al., ), applications requiring high levels f privacy, or those where individual interactions with a system re necessarily infrequent (e.g. dietary surveys bradley et al., ) here are no features and ratings to exploit, and the construction nder the cc by license. ( http://creativecommons.org/licenses/by/ . / ) https://doi.org/ . /j.eswa. . . http://www.sciencedirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi= . /j.eswa. . . &domain=pdf http://creativecommons.org/licenses/by/ . / mailto:t.osadchiy@newcastle.ac.uk mailto:ivan.poliakov@newcastle.ac.uk mailto:patrick.olivier@newcastle.ac.uk mailto:maisie.rowland@newcastle.ac.uk mailto:emma.foster@newcastle.ac.uk https://doi.org/ . /j.eswa. . . http://creativecommons.org/licenses/by/ . / t. osadchiy et al. / expert systems with applications ( ) – o g g c l t t i a t a m fi a b t t u t d c fi l a b c o w t t a t q p e o v d b p t w s i d i d d l m t c t d f t of personal behavior and preference models is not possible. to ad- dress these challenges we describe a recommender algorithm that is independent of any personal user model and does not require a complex system of ratings. based on a set of observed items se- lected by a user, the algorithm produces a set of items ranked by confidence of their being observed next. in designing the under- lying algorithm, we review existing methods that aim to address similar tasks, adapt them to meet the constraints of the application context that is our primary concern (dietary surveys), and propose a novel alternative. the performance of three methods is compared through the task of recommending omitted foods in a real world dietary recall system. . related work while various approaches have been proposed to address the cold-start problem in recommender systems, the majority of these rely on knowledge of content ( popescul, pennock, & lawrence, ) and users ( lika et al., ), including social relationships between users ( carrer-neto et al., ), whereas our concern is with contexts where such information is not available. shaw et al. addressed the cold-start problem in recommender systems by us- ing a data analysis technique, which is applied to large data sets for discovering items that frequently appear together in a single transaction. this technique is known as association rules ( agrawal, imielinski, & swami, ; shaw et al., ). each association rule normally consists of a set of antecedent items that lead to a con- sequent item with a certain confidence. pazzani and billsus con- sider the list of topics of books users voted for as transactions, which allowed them to extract association rules for topics that fre- quently appeared together as part of a user’s interests ( pazzani & billsus, ). to expand preferences for each user, the algo- rithm then generates all possible combinations of topics for every book the user voted for and filters association rules, where the an- tecedent part of the rule matches one of the combinations. the consequent list of topics is added to the preferences of that user. in combination with a domain ontology association rules can be effectively employed for extracting, understanding and for- malizing new knowledge ( ruiz, foguem, & grabot, ; sene, kamsu-foguem, & rumeau, ). however, association rules have to be adapted for recommendation tasks since they are primar- ily designed to be used as exploratory tools ( rudin, letham, salleb-aouissi, kogan, & madigan, ) to discover previously un- known relations that need to be analyzed for their interesting- ness ( atkinson, figueroa, & pérez, ). as we will probably want to provide more specificity, and recommend the exact titles of the books instead of generic categories, this potentially leads to a vast number of mined association rules and matching all pos- sible combinations of the observed items may not result in rules being found. furthermore, a consequent item may appear in mul- tiple matching rules, meaning that a function must be introduced that aggregates the confidences of found rules into a single score for the consequent item. finally, only the associations with a sup- port (i.e. how often a rule holds as true across the data set) higher than a defined threshold are normally extracted ( li & deng, ). the produced list of rules is supposed to be of a reasonable size, to allow manual examination. in a recommender system, even as- sociations with low frequencies could still be relevant, if other rel- evant rules with higher confidence are not found. this requires the extraction of as many rules as possible, making the mining process a computationally expensive task ( zheng, kohavi, & mason, ). roth et al. ( ) introduced a method for building implicit so- cial graphs based on histories of interaction between users and es- timations of their affinity and applied it to the problem of email recipient recommendation. based on a set of email addresses se- lected by a user (the seed group) the algorithm extracts all groups f contacts with whom the user has ever exchanged emails. here a roup of contacts is a set of email contacts that were observed to- ether in a recipient list. for each of the contacts in each group, ex- luding members of the seed group, an interaction score is calcu- ated based on the volume of messages exchanged with that group, he recency of those messages, and the number of intersections of he seed group with the group of contacts that is being considered. nteraction scores of contacts that are present in multiple groups re aggregated into a single interaction score, which is then used o rank the set of email recommendations. the implicit social graph is a promising alternative to mining ssociation rules. it instead measures the confidence of recom- ended items based solely on observed transactions that are pre- ltered by intersections with given items. however, the method lso estimates the relevance of a group of recommended emails ased on the strength of social interaction of the target user with hat group, which is not a meaningful metric for applications hat do not assume communication (or other interaction) between sers. association rules and implicit social graphs are data en- ities that represent item-to-group relationships. however, umouchel and pregibon ( ) suggested that a more effi- ient approach to discovering “interesting” associations is to first nd pairs of items that frequently appear together and then ana- yze larger sized item sets that contain those pairs. for example, if bc appears in a data set with a certain frequency, then pairs ab, c and ac would be at least as frequent as the triplet. raeder and hawla ( ) effectively analyzed associations through a graph f individual items connected to each other with edges that are eighted by the frequency of two items appearing together. items hat have stronger relationships with each other are compared o other items form clusters, which are then targeted for further nalysis. similar to the implicit social graph this method avoids he need to mine all possible association rules, but without re- uiring any additional indicators of relevance except for the item airs frequencies. the relevance of produced recommendations is ffectively inf erred from the likelihood of their appearing with the bserved items. . associated food recommender algorithm . . intake we introduce a new recommendation algorithm that was de- eloped for intake , a system for conducting -h multiple-pass ietary recall surveys ( bradley et al., ). intake is designed to e a cost-effective alternative to interviewer-led -h recalls and rovides respondents with a web-based interface through which hey enter their dietary intake for the previous day. respondents ill likely only ever use the system if they are a part of a dietary tudy and only for a short period of time. within a survey, a respondent typically records their dietary ntake for the previous day on three separate occasions. a single ay normally consists of four to seven meals (e.g. breakfast, morn- ng snack, lunch etc.) which include a selection of foods, drinks, esserts, condiments, and such (referred to generically as foods). uring the first step of a recall session, a respondent reports a ist of names of foods consumed during each of the previous day’s eals in a free-text format. for each text entry, the system re- urns a list of relevant foods selected from a taxonomy of around foods, organized in a tree-based, multi-level structure. spe- ific foods are terminal nodes of this taxonomy and are linked to heir nutrient values and portion size estimation methods. respon- ents select one food from the returned list to add to their meal; or example: coca-cola (not diet); beef, stir fried (meat only); toma- oes, tinned; basmati rice; onions, fried; chilli powder; kidney beans . t. osadchiy et al. / expert systems with applications ( ) – t a o i t m s o d a t c i w t d m r s t n f t d p a p a s o a r o a a o e t p t s s s o o r s e q w s i r m t e l h d i n r s d d s t p r s c p f t i s s i p m t ( ( i s t b i c d t c completing an accurate recall requires respondents to be able o identify foods they ate from a database that covers more than foods; for example, there are more than types of bread lone. thus, one of the key features in determining the usability f a dietary recall system is its presentation of food search results. f respondents are not able to readily identify items from a list re- urned in response to their textual description of the food, they are ore likely to select foods perceived as the closest match or even kip reporting the intake of that food. in other words, the relevance f search results, in terms of prioritizing them appropriately, may irectly affect the accuracy of dietary recall through level of effort nd time required to select the correct foods and report intake. the main application for the recommender algorithm is to au- omate the extraction of questions about foods that are commonly onsumed together (associated foods). in intake , this feature is mplemented as a link between an antecedent food (e.g. toast, hite bread ) and the consequent associated food category (e.g. but- er/margarine/oils ) along with a question that is asked if a respon- ent selects the antecedent food (e.g. did you have any butter or argarine on your toast? ). such food associations prompts are cur- ently hand-crafted by trained nutritional experts, which for thou- ands of foods is inevitably a time consuming process that is prone o omissions. eating habits depend on region, culture, diet, and a umber of other factors, which requires defining new associations or every context in which a system is deployed. furthermore, over ime new foods and recipes emerge and dietary trends change. in- eed, existing rules are often curated based only on personal ex- erience or previous research, and no published study has evalu- ted their appropriateness or explored alternative data driven ap- roaches to extracting such associations. . . generic procedure our approach assumes that the patterns in eating behaviors of n observed population; that is, the respondents who took part in urveys conducted in a given country, has some relevance to those f an individual in that population. the recommender algorithm ssumes no prior knowledge about an individual except their cur- ently selected food items. the algorithm is trained on a large set f observed meals and produces a model of the eating behavior of given population, where a meal is a group of uniquely identifi- ble foods (e.g. vanilla ice cream, pear juice) reported to be eaten n a single occasion. each individual food can be recorded as being aten only once during a meal. during the recommendation step, he resulting model accepts a set of foods, which we refer to as in- ut foods if , and returns a set of recommended foods rf mapped o likelihoods of being reported along with if (recommendation cores). if are excluded from recommendations. in the following ection we discuss three possible implementations that were con- idered for the recommender algorithm. along with the description f our methods we include examples of generated models and rec- mmendations for a sample transaction data set. . . association rules we introduce a recommender algorithm based on association ules (ar) that generates a model of eating behavior from a data et of meals (in the training step) in the form of association rules. ach rule consists of a set of antecedent foods and a single conse- uent food, together with the confidence that the consequent food ill be present in a meal given the antecedent foods that were ob- erved. the procedure for retrieving association rules is described n agrawal et al. ( ) . the ar algorithm makes predictions from stored association ules with antecedent part antc similar to if , and produces recom- endations from the consequent parts of the rules. to do so, ar akes association rules that have a consequent food that is differ- nt from any of if and the antecedent foods antc that include at east one of if ( algorithm ). the algorithm calculates the likeli- algorithm : recommendations based on association rules. function recommend input : am, association rules based model if , foods selected by a respondent returns : rf , list of food recommendations rf ← ∅ ; foreach rule r l ∈ am & r l.consequent / ∈ if : f ← rl.consequent; if ∃ a f : a f ∈ rl.antecedent & a f ∈ if : if f / ∈ rf : rf [ f ] ← ant c ← rl.antecedent ; c ← rl.con f idence ; intr ← size ({ a f : a f ∈ antc & a f ∈ if } ) ; ms ← intr / (size (antc) ∗ size (if )) ; rf [ f ] ← rf [ f ] + c ∗ ms return rf ood of a recommended food f to be selected next as the confi- ence of the rule c multiplied by the similarity between antc and f (i.e. match score ms ). the match score ms is calculated as the umber of foods that appear both in if and antc (i.e. intersections) aised to the power of two and divided by the size of if and the ize of antc . we introduce the match score so that the recommen- ations from the rules with antc that are more similar to if pro- uce recommendations that will appear higher. we then sum the cores for every f as its single recommendation score rf [ f ]. recommendations produced by ar applied to the example ransaction data set { abcd, ade, de, ab } and given items { ab } are rovided in table . . . transactional item confidence we adapted the implicit social graph method described in oth et al. ( ) for our food recommendation task, which re- ulted in a recommender algorithm based on transactional item onfidence (tic). one key difference to our food recommendation roblem is that the original email recipient recommendation task or which the implicit social graph was developed assumed two ypes of relationships between items in a data set (outgoing and ncoming emails). our data set assumes only one type of relation- hip, which is the co-occurrence of foods in a meal. for that rea- on, tic produces recommendations based on similarity of histor- cally observed transactions to if and the frequency of foods ap- earing in those transactions. during the training step, tic converts all reported meals to a ap of unique meals (or transactions) tm , so that there are no wo transactions of the same length containing the same foods algorithm ). for every food f in a transaction m , the confidence conditional probability) is calculated as tm [ m, f ] of f being present n m given that the rest of the foods from m were observed. to do o, the algorithm counts the number cf of reported meals that con- ain all the foods from m , excluding f , and divides it by the num- er cm of reported meals containing all of the foods from m . this s similar to the confidence measured in ar, but in this case we alculate it only for the full-sized meals that were observed in the ata set m , and not for all possible combinations of foods within hose meals. at the recommend step, the algorithm retrieves all transactions ontaining any of the input foods if ( algorithm ). within each of t. osadchiy et al. / expert systems with applications ( ) – table recommender algorithm based on association rules applied to the example data set. model based on ar filtered rules recommendations . a ⇒ b . , d . , c . , e . . a ⇒ d . , c . , e . d: . . b ⇒ a . , c . , d . . b ⇒ c . , d . c: . . c ⇒ a . , b . , d . . a, b ⇒ c . , d . e: . . d ⇒ a . , e . , b . , c . . a , c ⇒ d . . e ⇒ d . , a . . a , d ⇒ c . , e . . a, b ⇒ c . , d . . a , e ⇒ d . . a, c ⇒ b . , d . . b , c ⇒ d . . a, d ⇒ b . , c . , e . . b , d ⇒ c . . a, e ⇒ d . . a, b , c ⇒ d . . b, c ⇒ a . , d . . a, b , d ⇒ c . . b, d ⇒ a . , c . . c, d ⇒ a . , b . . d, e ⇒ a . . a, b, c ⇒ d . . a, b, d ⇒ c . . a, c, d ⇒ b . . b, c, d ⇒ a . algorithm : training the model based on transactional item confidence. function train input : m, data set of all meals returns : t m, map of unique meals with confidence for every food t m ← ∅ ; foreach meal m ∈ m : if m / ∈ t m : t m[ m ] ← ∅ cm ← size ({ m : m ∈ m & m ∈ m } ) ; foreach food f ∈ m : m ← { f : f ∈ m & f � = f } ; c f ← size ({ m : m ∈ m & m ∈ m } ) ; t m[ m, f ] ← c f /cm return tm algorithm : recommendations based on transactional item confidence. function recommend input : t m, map of unique meals with confidence for every food if , foods selected by a respondent returns : rf , list of food recommendations rf ← ∅ ; foreach meal m ∈ t m : if ∃ f : f ∈ m & f ∈ if : foreach food f ∈ m & f / ∈ if : if f / ∈ rf : rf [ f ] ← con f ← m [ f ] ; inter ← size ({ f : f ∈ m & f ∈ if } ) ; rf [ f ] ← rf [ f ] + inter ∗ con f return rf t v d o a w p t a t c t a c t t c i the retrieved transactions m , foods f that are not included in if are mapped to a score that is calculated as the number of intersections of m with if (i.e. similarity) multiplied by the food’s confidence tm [ m, f ]. multiple scores for f measured from different transactions are summed into a final recommendation score rf [ f ]. recommendations produced by the tic applied to an example ransaction data set { abcd, ade, de, ab } given items { ab } are pro- ided below ( table ). . . pairwise association rules unlike the previous two algorithms, which produce recommen- ations from association rules and transactions similar to currently bserved if , the recommender algorithm based on pairwise associ- tion rules (par) recommends foods that are likely to be observed ith any of if in pairs. during the training stage ( algorithm ), ar for every observed food f counts the number od [ f ] of meals algorithm : training the model based on pairwise associa- tion rules. function train input : m, data set of all meals returns : p m, pairwise association rules od ← ∅ , food occurrences; cd ← ∅ , food co-occurrences; foreach meal m ∈ m : foreach food f ∈ m : if f / ∈ od & f / ∈ cd : od [ f ] ← ; cd [ f ] ← ∅ ; od [ f ] ← od [ f ] + ; foreach food f ∈ m & f � = f : if f / ∈ cd [ f ] : cd [ f, f ] ← cd [ f, f ] ← cd [ f, f ] + p m ← [ od, cd ] ; return pm hat contain that food. for every observed pair of foods { f, f }, it lso counts the number cd [ f, f ] of reported meals that contain hat pair. at the recommend step ( algorithm ), par retrieves pairs d [ inf ], where one food inf is observed in if . for every pair { inf, f }, he algorithm calculates the conditional probability p , of f being in meal, given that inf was observed as the number of meals that ontain that pair cd [ inf, f ], divided by the number of meals od [ inf ] hat contain only inf . for example, if item a appeared times in he data set and co-occurred with item b only times, then the onditional probability that item b will occur the next time the a s present is . . multiple probabilities retrieved for f from differ- t. osadchiy et al. / expert systems with applications ( ) – table recommender algorithm based on transactional confidence applied to the example data set. model based on tic filtered rules recommendations . a . , b . , c . , d . . a . , b . , c . , d . d: . . d . , a . , e . . d . , a . , e . c: . . d . , e . e: . . a . , b . algorithm : recommendations based on pairwise association rules. function recommend input : p m, pairwise association rules if , foods selected by a respondent returns : rf , list of food recommendations rf ← ∅ ; p ← ∅ , conditional probabilities of foods; w ← ∅ , conditional probability weights; od ← p m[ od ] , food occurrences ; cd ← p m[ cd ] , food co-occurrences ; foreach input food in f ∈ if : foreach food f ∈ cd [ in f ] & f / ∈ if : if f / ∈ p & f / ∈ w : p [ f ] ← ∅ ; w [ f ] ← ∅ p ← cd [ in f, f ] / od [ in f ] ; p [ f ] ← p [ f ] + { p} ; w [ f ] ← w [ f ] + { od [ in f ] } foreach food f ∈ p : rf [ f ] ← sum (p [ f ]) ∗ sum (w [ f ]) return rf e p i d t m g s a f c p r a c w s t v p p w f i s t r p d m a c o r r t w e p w e a l ( d t a ( a s m fi t ( e b a c a n i z p i a s fi o t ( d nt associations are summed into its single recommendation score [ f ]. as demonstrated in roth et al. ( ) the number of times tems are observed together is an important relevance metric. in- eed, if we simply aggregate the probabilities derived from mul- iple associations, we lose information as to whether a recom- ended food has ever been observed with all if . for example, iven two input items c and d , the aggregation may produce two cores r cd (a ) = . , where item a appeared with both c and d , nd r c (b ) = . , where item b appeared only with item c . there- ore a should receive a higher score. likewise we take into ac- ount the frequency of an input food inf that matched a retrieved air. for example, we may have two equal scores, r c (a ) = . and d (b ) = . , where a and b historically appeared only with items c nd d respectively; but c appeared times and item d appeared times, which implies that the recommendation produced by should have a higher score. for these reasons, the algorithm eights aggregated probabilities p [ f ] by multiplying them by the ummed frequency of inf . recommendations produced by par applied to the example ransaction data set { abcd, ade, de, ab } given items { ab } are pro- ided in table . . methodology we compare the three algorithms for randomly sam- led meals, each containing no fewer than two foods, reported by articipants of various ages in the uk between and . e also randomize the order of foods in each meal. we use k- old ( k = ) cross validation to segment the data set into train- ng and testing sets ( salzberg, ). on each step we use nine ubsets for training a model, leaving out one subset for testing. he testing procedure is similar to the procedure described in oth et al. ( ) : we sample a few foods from each meal (in- ut foods), leaving the rest (at least one food) to simulate respon- ents’ omitted foods to be guessed by the algorithm. every trained odel makes predictions, starting from an input size of one food nd gradually incrementing it to five. in the course of the evaluation, we plot the precision-recall (pr) urves for every algorithm on every increment. for the purposes f the evaluation, we measure the recall as the percentage of cor- ect predictions out of the total number of foods selected by the espondent, and the precision as the percentage of correct predic- ions out of the total number of predictions made by the algorithm. e count predictions that were present in the set of foods actually ntered by the respondent (excluding input foods) as correct (true ositives). we analyze the quality of the top recommendations, hich is a slightly larger size than viewed by most users ( burges t al., ; van deursen & van dijk, ). as the measure of lgorithm ranking quality for every size of input foods we calcu- ate the mean value of normalized discounted cumulative gain ndcg) at rank ( burges et al., ) as ndc g = dc g /idc g . iscounted cumulative gain is measured as dcg = ∑ i = ( r(i ) − ) / log (i + ) , where r ( i ) is the relevance score of the i th food. as he relevance score, we use for a wrong prediction and for correct prediction. thus, the ideal discounted cumulative gain idcg) in our case is always , which is a single correct prediction s the first result. we then select the implementation that demon- trates the highest performance and apply it to the task of recom- ending foods omitted by respondents with a lower level of speci- city, and for ranking search results returned in response to their ext queries. for the implementation of ar we use the fp-growth algorithm frequent patterns algorithm) ( li & deng, ). fp-growth is an fficient and scalable association rules mining algorithm that is ased on building frequent-pattern tree structure. in contrast to priori-like algorithms that serve the same purpose, the fp-growth ompresses a large database into a much smaller data structure voiding costly repeated database scans and generation of a large umber of candidate sets. we use a parallel version of fp-growth mplemented in the apache spark framework ( li, wang, zhang, hang, & chang, ; meng et al., ). as a parameter, this im- lementation accepts the minimum support for an item set to be dentified as frequent and the minimum confidence for the gener- ted association rules. to gather as many association rules as pos- ible we set both the minimum support and the minimum con- dence to the lowest value ( × ) that allows the completion f the mining process of our data set on our machine within a ime limit of minutes. the evaluation is conducted on mac pro . ghz intel core i , gb). . results . . general performance as can be observed from the pr curves ( figs. and ), par pro- uces the largest area under the curve, which increases with the t. osadchiy et al. / expert systems with applications ( ) – table recommender algorithm based on pairwise association rules applied to the example data set. model based on par filtered rules recommendations . a . ⇒ b . , d . , c . , e . . a . ⇒ d . , c . , e . d: . . b . ⇒ a . , c . , d . . b . ⇒ c . , d . c: . . c . ⇒ a . , b . , d . e: . d . ⇒ a . , e . , b . , c . . e . ⇒ d . , a . fig. . precision-recall curves for an input size of foods. fig. . precision-recall curves for an input size of foods. table mean training and recommendation times in mil- liseconds. model training mean recommendation ar . . par . . tic . . fig. . the ratio of mean ndcg for the top results to the number of input foods. fig. . the ratio of recall for the results to the number of input foods for pair- wise association rules with the first and the second levels of specificities and man- ually entered associated food prompts. r f a f a i t i e r c c c size of input foods. par also demonstrates higher ndcg than tic and ar for all input sizes ( fig. ). par is the second fastest algo- rithm to produce a model (after ar) but the fastest to produce a single set of recommendations ( table ). based on this comparison, par is selected to be used for the implementation of the associ- ated foods recommender algorithm. at the same time, these results demonstrate that the quality of predictions produced by par is still relatively low. in the following experiments we aim to improve the performance of the algorithm by exploiting the context of the task it is used for. . . associated food questions to compare the efficacy of recommendations produced by the ecommender algorithm to the existing hand-coded associated ood questions we go through the same evaluation procedure as bove, except that on the recommend step a trained model returns ood categories instead of exact foods. in this case, true positives re considered to be foods selected by the respondent (excluding nput foods) that belong to one of the food categories predicted by he recommender algorithm. the taxonomy of foods implemented n intake allows control of the specificity of the returned cat- gories. so, we demonstrate the performance of the algorithm in eturning the direct parent category of a food (first level, e.g. flake ereals is the parent category for choco flakes ) and a more generic ategory (second level, e.g. breakfast cereals ) that is a parent of the ategory with the first-level specificity ( fig. ). since the existing t. osadchiy et al. / expert systems with applications ( ) – table omitted foods captured with pairwise association rules but not with manually entered associated food prompts. input foods first-level specificity second-level specificity chicken breast; fanta; instant potato gravy sauces, condiments, gravy and stock bananas; fruit and yoghurt smoothie; semi skimmed milk sugar sugar, jams, marmalades, spreads and pates blackcurrant squash (juice), e.g ribena; heinz beans and sausages brown bread toasted brown, wholemeal and : bread porridge, made with skimmed milk; tea; white sugar butter butter/margarine/oils tuna mayo sandwich; volvic mineral water, still or fizzy chocolate covered biscuits sweet biscuits bread sticks; coffee dips pickles, olives, dips and dressings cheese and tomato pizza (includes homemade); raspberries ice cream ice cream & ice lollies cheese sandwich; tea crisps and snacks crisps, snacks and nuts green olives; water wine alcohol bottled mineral water; chicken breast fillet; chips, fried; hot sauce fizzy drinks drinks still energy drink, eg lucozade hydroactive, gatorade, powerade; tuna in brine, tinned; white bread sliced mayonnaise sauces/condiments/gravy/stock a t r d f a w l r o s i t l d o c d e f s fi t t m e e v w t t n t b a fig. . the ratio of mean ndcg for the top results to the number of input foods for the search results ranked based on pairwise association rules and frc. u q p i s t a w o g e w m i f a u p r p ssociated food questions do not store any relevance scores plot- ing their pr curves or assessing their ndcg is impossible. for that eason we compare the recall of the top recommendations pro- uced by the algorithm to the recall of all hand-coded associated oods rules extracted for given input foods. in the simulation of respondents omitting foods hand-coded ssociation food rules recognize . % of omitted foods at most, hereas the recommender algorithm’s peak recall is at . % and . % for the first and the second levels of specificity respectively. table includes examples of commonly forgotten foods estab- ished in the validation of intake ( bradley et al., ) but cor- ectly predicted by the recommender algorithm with two levels f specificity. at the time of writing this paper, none of these as- ociations were covered by hand-coded associated foods rules in ntake . in addition to that, controlling the specificity of the re- urned recommendations allows us to address the cold-start prob- em, so that new foods that have not been reported by any respon- ents can still be captured by their categories. however, the names f some food categories predicted with the second-level specificity ould be perceived as too generic (e.g. ”pickles, olives, dips and ressings”) and may require being assigned names that would be asier to understand by respondents when displayed in associated ood prompts. . . search ranking in response to a respondent’s text query, the existing intake earch algorithm ranks foods based on two types of scores. the rst is the matching cost of the known food description against he query. the matching cost is based on several metrics, including he edit distance between matched words (the approximate string atching is performed using levenshtein automata burges et al., ); phonetic similarity of words (using a pluggable phonetic ncoding algorithm that depends on the localization language, .g. soundex or metaphone for english elmagarmid, ipeirotis, & erykios, ); the relative ordering of words; the number of ords not matched; and so forth. the lower the matching cost, he better the food name matches the query. the second score is he likelihood of the food being selected, which is measured by the umber of times the food was previously reported. the results are hen sorted, first by decreasing food report count (frc) and then y increasing matching cost. the evaluation of the associated foods recommender algorithm pplied to the task of ranking search results, follows the same eval- ation procedure, with some variations. in response to each text uery that was recorded into the intake database for each re- orted food (excluding input foods), we retrieve a list of foods us- ng the existing search algorithm. we count foods selected by a re- pondent as true positives and the rest of the results as false nega- ives. we compare the mean ndcg produced by the existing search lgorithm and by the new search algorithm, where frc is replaced ith par. as we can see from the figure below ( fig. ), par slightly utperforms frc starting from an input size of two foods, with the ap gradually widening as the number of input foods increases. . conclusions we aimed to address one of the key issues in automated di- tary assessment, which is unintentional under-reporting. to do so, e developed an associated foods recommender algorithm to re- ind respondents of omitted foods and improve the ranking qual- ty of search results returned in response to respondents’ free-text ood name queries. the algorithm, in contrast with collaborative nd content-based filtering approaches, is independent of personal ser profiles and does not require an extensive history of users’ references or a multitude of item descriptors. instead, the algo- ithm uses transactions performed by respondents from a given opulation to build a collective model of preferences. t. osadchiy et al. / expert systems with applications ( ) – d e l l l l m p p r r r r s v z we considered three approaches to the implementation of the recommender algorithm, based on an implicit social graph ( roth et al., ), association rules ( agrawal et al., ), and an- alyzing pairwise association rules ( dumouchel & pregibon, ). the evaluation, performed on a large data set of real dietary re- calls, has demonstrated that the implementation based on pair- wise association rules performs better for the defined task. by con- trolling the specificity of the produced recommendations within a reasonable level we achieved a recall of . %. that is signifi- cantly higher than food associations hand-coded by trained nutri- tionists, the recall for which reached only . %. where a respon- dent filled in at least one food, the recommender algorithm im- proves the ranking of search results. the algorithm was evaluated on dietary recalls of respondents from the uk. as a future work we are planning to analyze how dietary specificities of different regions affect the accuracy of the recommender algorithm. although the evaluation results described in this paper were produced by analyzing food contents of meals reported by respondents in intake , the described methods apply to any recommender tasks where selection of items by the target user can be observed (e.g. email recipients or tags recommenda- tions on community platforms). supplementary material supplementary material associated with this article can be found, in the online version, at doi: . /j.eswa. . . . references agrawal, r., imielinski, t., & swami, a. ( ). mining association rules between sets of items in large databases. in proc. acm sigmod international conference on management of data, sigmod’ : (pp. – ). acm. doi: . / . . atkinson, j. , figueroa, a. , & pérez, c. ( ). a semantically-based lattice approach for assessing patterns in text mining tasks. computación y sistemas, ( ) . bradley, j., simpson, e., poliakov, i., matthews, j. n. s., olivier, p., adamson, a. j., & foster, e. ( ). comparison of intake (an online -h dietary recall tool) with interviewer-led h recall in – year-old. nutrients, ( ), . doi: . /nu . burges, c., shaked, t., renshaw, e., lazier, a., deeds, m., hamilton, n., & hullen- der, g. ( ). learning to rank using gradient descent. in proceedings of the nd international conference on machine learning (pp. – ). acm. doi: . / . . carrer-neto, w., hernández-alcaraz, m. l., valencia-garcía, r., & garcía- sánchez, f. ( ). social knowledge-based recommender system. application to the movies domain. expert systems with applications, ( ), – . doi: . /j.eswa. . . . covington, p., adams, j., & sargin, e. ( ). deep neural networks for youtube rec- ommendations. in proceedings of the th acm conference on recommender sys- tems, recsys ’ (pp. – ). acm. doi: . / . . umouchel, w., & pregibon, d. ( ). empirical bayes screening for multi-item associations. in proceedings of the th acm sigkdd international conference on knowledge discovery and data mining, kdd’ (pp. – ). acm. doi: . / . . lmagarmid, a. k., ipeirotis, p. g., & verykios, v. s. ( ). duplicate record detec- tion: a survey. ieee transactions on knowledge and data engineering, ( ), – . doi: . /tkde. . . i, c. r. j., & deng, z. h. ( ). mining frequent ordered patterns without candi- date generation. in proceedings th international conference on fuzzy systems and knowledge discovery, fskd : vol. (pp. – ). ieee. doi: . /fskd. . . i, h., wang, y., zhang, d., zhang, m., & chang, e. ( ). pfp: parallel fp-growth for query recommendation. in proceedings of the acm conference on recom- mender systems (pp. – ). acm. doi: . / . . ika, b., kolomvatsos, k., & hadjiefthymiades, s. ( ). facing the cold start prob- lem in recommender systems. expert systems with applications, ( part ), – . doi: . /j.eswa. . . . inden, g., smith, b., & york, j. ( ). amazon.com recommendations: item-to- item collaborative filtering. ieee internet computing, ( ), – . doi: . / mic. . . eng, x., bradley, j., yavuz, b., sparks, e., venkataraman, s., liu, d., et al. ( ). ml- lib: machine learning in apache spark. the journal of machine learning research, ( ), – . doi: . / . . azzani, m. j., & billsus, d. ( ). content-based recommendation sys- tems. in the adaptive web: vol. (pp. – ). springer. doi: . / - - - - _ . opescul, a. , pennock, d. m. , & lawrence, s. ( ). probabilistic models for uni- fied collaborative and content-based recommendation in sparse-data environ- ments. in proceedings of th conference on uncertainty in artificial intelligence (pp. – ). morgan kaufmann publishers inc . aeder, t., & chawla, n. v. ( ). market basket analysis with networks. social net- work analysis and mining, ( ), – . doi: . /s - - - . oth, m., ben-david, a., deutscher, d., flysher, g., horn, i., leichtberg, a., et al. ( ). suggesting friends using the implicit social graph. in proceedings of th acm sigkdd international conference on knowledge discovery and data mining (pp. – ). acm. doi: . / . . udin, c. , letham, b. , salleb-aouissi, a. , kogan, e. , & madigan, d. ( ). sequential event prediction with association rules. in proceedings of th annunal confer- ence on learning theory (pp. – ) . uiz, p. p. , foguem, b. k. , & grabot, b. ( ). generating knowledge in maintenance from experience feedback. knowledge-based systems, , – . salzberg, s. l. ( ). on comparing classifiers: pitfalls to avoid and a recommended approach. data mining and knowledge discovery, ( ), – . schein, a. i., popescul, a., ungar, l. h., & pennock, d. m. ( ). methods and met- rics for cold-start recommendations. in proceedings of the th annual interna- tional acm sigir conference on research and development in information retrieval, sigir ’ (pp. – ). acm. doi: . / . . ene, a. , kamsu-foguem, b. , & rumeau, p. ( ). discovering frequent patterns for in-flight incidents. cognitive systems research, , – . shaw, g., xu, y., & geva, s. ( ). using association rules to solve the cold- start problem in recommender systems. in pacific-asia conference on knowl- edge discovery and data mining: (pp. – ). springer. doi: . / - - - - _ . an deursen, a. j., & van dijk, j. a. ( ). using the internet: skill related problems in users’ online behavior. interacting with computers, ( – ), – . doi: . /j.intcom. . . . heng, z., kohavi, r., & mason, l. ( ). real world performance of association rule algorithms. in proceedings of th acm sigkdd international conference on knowledge discovery and data mining, kdd’ (pp. – ). acm. doi: . / . . https://doi.org/ . /j.eswa. . . https://doi.org/ . / . http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref https://doi.org/ . /nu https://doi.org/ . / . https://doi.org/ . /j.eswa. . . https://doi.org/ . / . https://doi.org/ . / . https://doi.org/ . /tkde. . https://doi.org/ . /fskd. . https://doi.org/ . / . https://doi.org/ . /j.eswa. . . https://doi.org/ . /mic. . https://doi.org/ . / . https://doi.org/ . / - - - - _ http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref https://doi.org/ . /s - - - https://doi.org/ . / . http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref https://doi.org/ . / . http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref http://refhub.elsevier.com/s - ( ) -x/sbref https://doi.org/ . / - - - - _ https://doi.org/ . /j.intcom. . . https://doi.org/ . / . recommender system based on pairwise association rules introduction related work associated food recommender algorithm . intake . generic procedure . association rules . transactional item confidence . pairwise association rules methodology results . general performance . associated food questions . search ranking conclusions supplementary material references can machine learning offer anything to expert systems? machine learning, , - ( ) © kluwer academic publishers, boston. manufactured in the netherlands. can machine learning offer anything to expert systems? bruce g. buchanan professor of computer science, medicine, and philosophy, university of pittsburgh, pittsburgh, pa today's expert systems have no ability to learn from experience. this commonly heard crit- icism, unfortunately, is largely true. except for simple classification systems, expert systems do not employ a learning component to construct parts of their knowledge bases from libraries of previously solved cases. and none that i know of couples learning into closed- loop modification based on experience, although the soar architecture [rosenbloom and newell ] comes the closest to being the sort of integrated system needed for continuous learning. learning capabilities are needed for intelligent systems that can remain useful in the face of changing environments or changing standards of expertise. why are the learn- ing methods we know how to implement not being used to build or maintain expert systems in the commercial world? part of the answer lies in the syntactic view we have traditionally taken toward learning. statistical techniques such as linear regression, for example, are "knowledge-poor" proce- dures that are unable to use knowledge we may bring to the learning task. however, learning is a problem-solving activity. as such, we can and should analyze the requirements and functional specifications, the input and output, and the assumptions and limitations. if there is one thing we have learned in ai it is that complex, i.e., combinatorially explosive and ill-structured, problems often can be tamed by introducing knowledge into an otherwise syntactic procedure. there is not a single model for learning effectively, as is confirmed by the proliferation of methods, just as there is not a single model of diagnosis or of scheduling. if the machine learning community is going to make a difference in the efficiency of building commercial expert systems, we need not only more methods but a better understanding of the criteria for deciding which method best suits which learning problem. no matter what model one adopts, there is a significant chance of success in coupling a learning program with an expert system. and there will be significant benefits to the machine learning community in terms of research problems that crystallize out of the solution of applications. the workhorse of knowledge acquisition for commercial expert systems is still knowledge engineering. it is labor intensive and prone to error because it involves discussions among persons with very different backgrounds. we are slightly better at teaching knowledge engi- neering, and practicing it, than we were a decade ago. it has the outstanding virtue that it works. partly because it is slow, it also allows more deliberate examination of conceptual frameworks, of assumptions, and of mistakes. but it is too slow: it stands in the same rela- tion to knowledge acquisition as do paper and pencil to calculation. e.g. buchanan speeding up the knowledge engineering process has led to the development of customized editors of the salt variety [marcus and mcdermott ]. these are being used success- fully and provide some assistance to knowledge base developers. but they are not learning programs; they are editors. as an aside, though, they allow us to meet mccarthy's challenge [mccarthy ] of building programs that can accept advice before we build systems that learn on their own. in a customized editor we see elements of expert systems. they use knowledge of a generic task such as diagnosis to guide the acquisition and debug- ging of a knowledge base. with our increased experience in building systems over the last decade, we should now be able to integrate the techniques of teiresias [davis ] and roget [bennett ] for even more powerful editors. as we move closer to automated knowledge acquisition and away from knowledge engi- neering, we must understand better how to convey the built-in assumptions of the programs that assist in constructing knowledge bases. with expert systems, there is a potential danger that the designers of programs will have different conceptual frameworks in mind than the users. in the case of learning programs, the same potential danger arises in the mismatch of frameworks between the designers of the learning programs and the designers of the expert systems. we in the machine learning community have not adequately addressed this question, nor would we think of it unless there were a pressing need to apply machine learning techniques to problems of interest to another community. inductive learning has received considerable attention since the s, with several ap- proaches now in our growing toolkit of programs that can assist in knowledge acquisition. some programs assume the data are correct, that the descriptions and classifications of examples are error-free; some of the same ones assume that the examples are all available at once, and hereafter remain available; others assume that classification rules are categorical; most assume that the rules to be learned are one-step conjunctive rules. successful applica- tions of id and c [quinlan, et al. ] and their cousins have provided knowledge bases for working expert systems whose task is to classify. these, together with a few other isolated cases of an expert system being built with the assistance of a learning program, have kept alive the promise of payoff from the -year investment in learning research. explanation-based learning (ebl) is now receiving even more attention than induction or similarity-based learning. here, more than with induction, the problem-solving nature of acquiring new knowledge is apparent. (see, for example, [mitchell ].) its main assumption is that enough theory exists to provide a rationalization of why one instance is or is not a prototypical member of a class. what else? does it have to be a formal, deduc- tive rationalization, or will an empirical argument suffice? (see [smith, et al. ] for an argument that a weaker, empirical model will suffice.) unfortunately, ebl has not found any applications in building expert systems, thus it is still demonstrated only in laboratory prototypes. case-based and analogical reasoning are still pipe dreams when matched against the harsh standards of robustness of commercial applications. some of the problems stem from the dilemma that we can find some mappings from any previous case or previously studied domain to the present one: we almost have to know the answer in order to find the case or the analogy that lets us program a machine to find the answer. analogical reasoning involves at least two problem-solving steps: finding a useful analogous domain (or object or process) and finding relevant mappings between the analog and the original thing. it can machine learning offer anything to expert systems? is not a robust method in the sense that it still produces too many false positive results. moreover, we don't see many suggestions for making these methods more selective. their power seems to lie in offering suggestions when we have run out of other ideas. what about neural nets? they certainly are popular because the idea of getting something for nothing has always held great appeal. they have shown some success in knowledge-poor tasks, such as character recognition and other perceptual tasks. expert systems, by their very nature, are knowledge-intensive, however, and thus are less amenable to learning with syntactic reinforcement methods. once a neural net is tuned, there is no way to understand why it succeeds on some cases and fails on others except to say that the weights on some nodes are higher than on others. moreover, their construction is not free; considerable effort must be invested in laying out the structure of the network before examples can be presented. only a few of the techniques in the literature are immediately useful, and these have their limits. part of our research charge needs to include understanding the scope and limits of different methods and determining when they are applicable. expert systems provide a strong rationale for continued funding of research on machine learning, but they also serve to sharpen our understanding of problems. expert systems offer a focus for development of new machine learning methods and better understanding of old ones. of course, we need basic research as well as demand-driven research. at the moment, however, there is an imbalance in the amount of work on learn- ing in domains where we do not need to learn and on techniques with crippling assumptions. let us attempt to understand the context of the commercial, military, and medical needs. in research on machine learning, as on other problem-solving methods, new wrinkles— perhaps new opportunities—will arise in experimenting with real, complex knowledge bases and applications. using an expert system as a testbed offers a tough test of success. the commercial world of expert systems at large seems unconvinced that machine learning has anything to offer yet. i strongly disagree. inductive learning, at least, is already in limited use, and present methods can be extended to make them more useful. some of the issues that need to be resolved in order to make inductive methods a mainstay of commercial knowledge acquisi- tion are already modestly well understood: learning in the context of noisy and uncertain data, exploiting an existing partial theory, representing the objects of learning in a form other than classification rules, and tailoring the learning to the specific context in which the learned information will be used. these are partly issues of utility, but they are impor- tant research problems as well. machine learning is ready for development now, with atten- dant benefits to us in crystallizing current and new research problems. references bennett, j.s. . roget: a knowledge-based system for acquiring the conceptual structure of a diagnostic expert system. j. automated reasoning , - . davis, r. . teiresias: applications of meta-level knowledge. in r. davis and d. lenat (eds.). . knowledge-based systems in artificial intelligence. new york: mcgraw-hill. b.g. buchanan marcus, s. and mcdermott, j. . salt: a knowledge acquisition language for propose-and-revise systems. artificial intelligence , - . mccarthy, j. . programs with common sense. proc. symposium on the mechanisation of thought processes. reprinted in m. minsky (ed.), semantic information processing. cambridge, ma: mit press, . mitchell, t.m., keller, r., and kedar-cabelli, s. . explanation-based generalization: a unifying view. machine learning , - . quinlan, j.r., compton, p.j., horn, k.a., and lazarus, l. . inductive knowledge acquisiton: a case study. proceedings second australian conference on applications of expert systems. in j.r. quinlan (ed.), applications of expert systems. maidenhead: academic press (in press). rosenbloom, p.s., and newell, a. . the chunking of goal hierarchies: a generalized model of practice. in r. michalski, j. carbonell, and t. mitchell (eds.), machine learning: an artificial intelligence approach (vol. ). los altos, ca: morgan-kaufmann. smith, r.g., winston, h., mitchell, t., and buchanan, b.g. . representation and use of explicit justification for knowledge base refinement. in proceedings of ijca . los altos, ca: morgan-kaufmann. a semantic framework for textual data enrichment accepted manuscript a semantic framework for textual data enrichment yoan gutiérrez , sonia vázquez , andrés montoyo pii: s - ( ) - doi: . /j.eswa. . . reference: eswa to appear in: expert systems with applications received date: december revised date: february accepted date: march please cite this article as: yoan gutiérrez , sonia vázquez , andrés montoyo , a semantic framework for textual data enrichment, expert systems with applications ( ), doi: . /j.eswa. . . this is a pdf file of an unedited manuscript that has been accepted for publication. as a service to our customers we are providing this early version of the manuscript. the manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. http://dx.doi.org/ . /j.eswa. . . http://dx.doi.org/ . /j.eswa. . . accepted manuscript a cc ep te d m a n u sc ri pt highlights  a semantic framework for recommender systems is presented  an in-depth analysis of different natural language processing resources is showed  a description of different natural language processing approaches is addressed  related research works are described  a case of study to evaluate our proposal with real data is presented accepted manuscript a cc ep te d m a n u sc ri pt a semantic framework for textual data enrichment yoan gutiérrez, sonia vázquez * and andrés montoyo department of software and computing systems. university of alicante, spain. ygutierrez, svazquez, montoyo@dlsi.ua.es abstract: in this work we present a semantic framework suitable of being used as support tool for recommender systems. our purpose is to use the semantic information provided by a set of integrated resources to enrich texts by conducting different nlp tasks: wsd, domain classification, semantic similarities and sentiment analysis. after obtaining the textual semantic enrichment we would be able to recommend similar content or even to rate texts according to different dimensions. first of all, we describe the main characteristics of the semantic integrated resources with an exhaustive evaluation. next, we demonstrate the usefulness of our resource in different nlp tasks and campaigns. moreover, we present a combination of different nlp approaches that provide enough knowledge for being used as support tool for recommender systems. finally, we illustrate a case of study with information related to movies and tv series to demonstrate that our framework works properly. keywords: recommender systems, framework, integrated semantic resources, sentiment analysis, word sense disambiguation, content categorization . introduction recent advances in modern technologies have motivated the development of different techniques to improve human-machine communication. internet and new communication tendencies such as: short messages, forum participations, social networks, etc, have led to a revolution in the way in which people work, communicate and manage their free time. as a consequence of this technological revolution, a huge quantity of information is generated in different social contexts via diverse sources such as: forums, blogs, microblogs, social networks, etc. as a result, people are able to share their knowledge, expectations and emotions through internet and they may also influence political, economic or social behaviour. at this point, governments, enterprises or even celebrities need to manage this information in order to extract relevant knowledge, social tendencies, etc. because of this new context, research community in natural language processing (nlp) have developed different tools with which to analyse news and opinions in order to discover what people think or how they perceive past, present and future. at present, personalization and recommender systems have gained popularity. in fact, recommender systems began to appear in the market in (udi et al., ). since then, several approaches have been developed (gediminas and alexander, ):  content-based: these systems try to find products, services or contents that are similar to those already evaluated by the user. in this kind of systems, user‟s feedback (that can be collected in many ways) are essential to support and accomplish recommendations (marco de et al., ).  knowledge-based: these systems model the user profile in order to, through inference algorithms, identify the correlation between their preferences and existing products, services or content. (walter et al., )  collaborative filtering: these systems create/classify groups of users that share similar profiles/behaviours in order to recommend products, services or content that has been well evaluated by the group to which a user belongs (perner et al., ).  hybrid: these systems combine two or more techniques previously mentioned to improve the „„quality‟‟ of recommendations (shinde and kulkarni, ). dealing with textual information and obtaining valuable knowledge require advanced natural language techniques to solve different kinds of problems: document correction, automatic translation, accepted manuscript a cc ep te d m a n u sc ri pt summary elaboration, opinion extraction, word sense disambiguation, etc. solving all of these problems requires a considerable linguistic knowledge and, even more importantly, a high computational cost. in the vast majority of tasks in nlp it is necessary to use external resources such as: machine- readable dictionaries , thesaurus , ontologies and others. these resources have different internal structures, interfaces, concept relations and other characteristics. one of the most frequently used resources in its different versions is wordnet (wn) (miller et al., ). various semantic resources related to wordnet have consequently been developed in different domains or by using semantic integration. but it is still difficult to find resources that provide semantic integration in different domains and which are useful for specific nlp tasks. in this work we present a new semantic resource (isr-wn) and a set of different methods to take advantage of it with the aim of enriching texts with semantic information. as a result, we provide a semantic framework suitable of being used as support tool for content-based recommender systems by annotating texts with different features such as: sentiments, polarities or domain labels. in order to analyse the results of the semantic enrichment process, we have carried out a comprehensive case of study using texts from movies and tv series reviews obtained from imdb . finally, we have evaluated how our proposed framework works comparing our results with real ratings. to summarize, we point out the main contributions of this work:  taking advantage of a semantic resource with different dimensions previously developed (isr-wn)  the use of a set of nlp methods based on isr-wn to take advantage of each one of its semantic dimensions  providing a new semantic framework that is able to enrich texts in several dimensions with the aim of obtaining a support tool for content-based recommender systems  an exhaustive evaluation with real datasets to demonstrate how it works the document is structured as follows. after this introduction, each semantic resource used in isr- wn is described, and an in-depth analysis of the different approaches for semantic integration resources in nlp is also presented. having evaluated previous proposals, in section we go on to show how isr- wn was developed. an evaluation according to its integration effectiveness is then provided in section . in section we provide a brief description of the different nlp tasks selected to enrich texts. section describes the characteristics of a case of study to illustrate how our framework works with real data obtained from imdb. in section we show some examples of how the semantic enrichment approaches are used to annotate texts. section provides the experiment results of the case of study and section presents a discussion about the results obtained. finally, the conclusions and further works are presented in section . . related work this section presents the different semantic resources that are integrated into isr-wn and a comparison with other semantic integration resources. . wordnet as mentioned in the previous section, wn is one of the most frequently used semantic resources in computational linguistics (navigli, ). wn is a lexical database for the english language also considered as ontology. it was created at the university of princeton and it represents a semantic conceptual and structured network of nouns, verbs, adjectives and adverbs. the basic unit of knowledge is the synset (synonym sets), which represents a lexical concept (Ševčenko, ). a synset is associated with a unique eight-digit number called an offset (this number is the position in the data file). each synset * corresponding author. tel. + ; fax: + . e-mail address: svazquez@dlsi.ua.es dictionaries of words available in electronic format. provides relationships among words (i.e., synonyms, antonyms and others). conceptualization of a domain in order to share information among different agents. http://wordnet.princeton.edu/ http://www.imdb.com http://wordnet.princeton.edu/wordnet/ accepted manuscript a cc ep te d m a n u sc ri pt represents different senses which are related through the use of semantic, conceptual or lexical connections. the result of this set of connections is a wide navigable network with a high number of interrelations among different word senses. the semantic relations among synsets are:  synonymy  antonymy  hyponymy / hyperonymy  meronymy / holonymy  entailment and cause,  and others…(more details in ) wn establishes the frequency of usage of each word sense (synset)in its internal relationships . for example, the word image has eight senses in wn . (see table ). as we can observe, one word has different senses, there is a sentence (gloss) which describes each one and each sense has a set of synonyms that are ordered by their frequency of usage. pos offset lemmas gloss noun effigy, image, simulacrum a representation of a person (especially in the form of sculpture); "the coin bears an effigy of lincoln"; "the emperor's tomb had his image carved in stone" noun picture ,image, icon, ikon a visual representation (of an object or scene or person or abstraction) produced on a surface; "they showed us the pictures of their wedding"; "a movie is a series of images projected so rapidly that the eye integrates them" noun persona, image (jungian psychology) a personal facade that one presents to the world; "a public image is as fragile as humpty dumpty" noun image, mental_image an iconic mental representation; "her imagination forced images upon her too awful to contemplate" noun prototype, paradigm, epitome, image a standard or typical example; "he is the prototype of good breeding"; "he provided america with an image of the good father" noun trope, figure_of_speech, figure, image language used in a figurative or non-literal sense noun double, image, look-alike someone who closely resembles a famous person (especially an actor); "he could be gingrich's double"; "she's the very image of her mother verb visualize, visualize, envision, project, fancy, see, figure, picture, image imagine; conceive of; see in one's mind; "i can't see him on horseback!"; "i can see what will happen"; "i can see a risk in this strategy" table . word senses of “image” in wn . it is important to emphasize that wn has been adapted to different languages: english, spanish, dutch, italian, german, french, czech, estonian, swedish, norwegian, danish, greek, portuguese, basque, catalan, romanian, lithuanian, russian, bulgarian, slovenian and others that are under development. these versions have been developed under the supervision of the university of princeton and later under that of the global wordnet association . this research work is based on two versions of wn: wn . with , synsets, of which , are nouns, , are adjectives, , are adverbs and , are verbs, and wn . with , synsets, of which , are nouns, , are adjectives, , are adverbs and , are verbs. . semantic resources aligned to wordnet http://wordnet.princeton.edu/man/wninput. wn.html https://wordnet.princeton.edu/man/cntlist. wn.html http://www.globalwordnet.org/ accepted manuscript a cc ep te d m a n u sc ri pt owing to the fact that wn has been used in many nlp research works, a set of different semantic resources aligned to wn synsets has been developed with the aim of obtaining more knowledge. some of these resources were created from wn, such as: wordnet domains (magnini and cavaglia, ), wordnet affect (magnini and cavaglia, , sara and daniele, ) and semantic classes (izquierdo et al., ). others emerged from the association of pre-produced tags, i.e., sumo . the resources used in our proposed semantic integration resource (isr-wn) are described in detail below. . . wordnet domains this is a resource for the english language. wordnet domains (wnd) includes a set of subject field codes (sfc) (magnini and cavaglia, ) with which to enrich wn synsets. each sfc groups a set of words related to the same domain. on the one hand, these domains identify the context of the definition and on the other, they allow a quick search of concepts to take place. for example, if we are searching for the meaning of disc in the computer science context, we need only check the domain label preceding each definition (in this case, computer science) until we find the correct definition ([magnetic_disc, magnetic_disc, disck, disk]). this resource therefore provides the integration of domain labels in wn with the aim of reducing wn sense granularity by grouping different senses in the same domain or semantic category. in wnd, wn has been annotated by using a semi-automatic process that assigns one or more domain labels to each synset. domain labels are selected from a set of hierarchically organized labels. in this research we use of these labels. the main purposes of annotating wn with sfc are:  to create new relations among words. domain labels allow us to relate words that pertain to different grammatical categories.  semantic annotation. domain labels are associated with synsets, signifying that the annotation takes place at the semantic level rather than at the word level.  synsets pertaining to different syntactic categories can be included in the same domain label.  word senses pertaining to different sub-hierarchies of wn can be included in the same domain label.  to reduce word sense granularity. grouping different senses of the same word in the same domain label reduces the polysemy of words. sense domain gloss man# person an adult male person (as opposed to a woman); “there were two women and six men on the bus” man# military someone who serves in the armed forces; “two men stood sentry duty” man# person the generic use of the word to refer to any human being; “it was every man for himself” man# factotum all of the inhabitants of the earth; “all the world loves a lover” man# biology, person any living or extinct member of the family hominidae man# person a male subordinate; “the chief stationed two men outside the building”; “he awaited word from his man in havana” man# person an adult male person who has a manly character(virile and courageous competent); “the army will make a man of you” man# person (informal) a male person who plays a significant role (husband or lover or boyfriend) in the life of a particular woman; “she takes good care of her man” man# person a manservant who acts as a personal attendant to his employer; “jeeves was bertie wooster’s man” man# play a small object used in playing certain board games; “he taught me to set up the men on the chess board”; “he sacrificed a piece to get a strategic advantage” table . domain labels and senses for man http://wndomains.fbk.eu/ http://wndomains.fbk.eu/wnaffect.html http://rua.ua.es/dspace/bitstream/ / / /ranlp blc .pdf http://www.ontologyportal.org/ accepted manuscript a cc ep te d m a n u sc ri pt for example, the word man has ten senses in wn (see table ). however, in wnd we can group different senses according to their domain labels, and the number of senses is thus reduced from ten to four (magnini et al., july ). moreover, there are different specification levels in the domain hierarchy. the deeper the level, the greater the specialisation is. fig shows a brief excerpt of the wnd hierarchy from (luisa bentivogli, ). fig . wordnet domains hierarchy among all the sfc domain labels there is a special one called factotum. this domain label was created in order to group two types of synsets (magnini and cavaglia, ):  generic synsets. those that it is difficult to classify into a particular domain label.  stop senses. those that frequently appear in different contexts, such as: numbers, days of the week, colours, etc. . . wordnet affect wordnet affect (wna) is an extension of wnd (magnini and cavaglia, , sara and daniele, ). it contains different subsets of affective concepts that group together synsets which denote emotional states. this resource was labelled by following a similar process to that of wnd. some of the represented concepts are: moods, situations eliciting emotions or emotional responses. this resource was extended with a set of additional labels called emotional categories. it has a hierarchical structure in which hyperonymy is used to relate the affective concepts of wn (valitutti et al., ). in a second revision, some modifications were made in order to differentiate those senses that are closer to emotional labels, and new labels were also included: positive, negative, ambiguous and neutral:  the first pertains to positive emotions. for example, it includes synsets such as: joy# or enthusiasm# .  negative defines negative states such as: anger# o sadness# .  ambiguous represents synsets whose semantics depends on the contexts in which they appear: surprise#  neutral represents synsets that refer to mental states but which are not characterised by valence. one important property of wna labels is that they associate the nouns and adjectives involved in emotional states. in this case, the adjective modifies the state of the noun, and may in some situations determine the modified noun state, i.e.: cheerful / happy boy (strapparava and valitutti, ). in other words, if the adjective pertains to an emotional state it could indicate how the noun is related. table shows a list of affective labels associated with synsets. labels are integrated into our research work. doctrines archaeology astrology history linguistic psychology art religion heraldry grammar psychoanalysis music dance drawing photography plastic_arts theatre mythology occultism roman_catholic theology painting philately jewelry numismatic sculpture accepted manuscript a cc ep te d m a n u sc ri pt affective labels examples emotion noun anger# , verb fear# mood noun animosisy# , adjective amiable# trait noun aggressiveness# , adjective competitive# cognitive state noun confusion# , adjective dazed# physical state noun illness# , adjective all in# hedonic signal noun hurt# , noun suffering# emotion-eliciting situation noun awkwardness# , adjective out of danger# emotional response noun cold sweat# , verb tremble# behaviour noun offense# , adjective inhibited# attitude noun intolerance# , noun defensive# sensation noun coldness# , verb feel# table . wordnet affects labels and their associated synsets. . . sumo sumo (suggested upper merged ontology) is considered to be an upper level ontology. it provides definitions for general terms and can be used as a basis for domain specific ontologies. it was created from the combination of different ontological contents in one single cohesive structure. it currently contains around , terms and , assertions (niles and pease, ). our research work only uses of the concepts that are aligned with wn. sumo was obtained from the information of: ontolingua and the developed ontologies of itbm- cnr (unrestricted-time, representation, anatomy, biologic-functions, and biologic-substances). it uses a standard language representation called suo-kif (pease, ) obtained from kif (knowledge interchange format) (genesereth and fikes, ). sumo was built by dividing the concepts into two groups: high level concepts and low level concepts. in the first group, the ontology of john sowa (sowa, ), and russell and norvig (russell and norvig, ) were considered, while the remaining concepts were included in the second group. finally, a unique conceptual structure combining the two high level ontologies was created. the remaining low level class contents were included after the combination. fig shows the high level concepts. fig . sumo high level concepts (Ševčenko, ). the higher level concept is the entity category, as occurs in most hierarchies. the entity concept groups the rest of the concepts, and the physical and abstract concepts are closest to it. an example of the word sense bank# is shown in fig , along with the sumo hierarchy. http://suo.ieee.org/suo/sumo/index.html http://www.ksl.stanford.edu/software/ontolingua/ http://www.ontologyportal.org/sumohistory/ entity abstract physical quantity attribute setorclass proposition relation processobject accepted manuscript a cc ep te d m a n u sc ri pt fig . sumo hierarchy for bank# . . . semantic classes a semantic class is a sense conceptualisation that can be manually or semi automatically created from different abstraction levels and in different domains. wn is composed of a set of related and connected synsets with different semantic relations. each of these synsets represents a concept and contains a set of words referring to the concept it describes (synonyms). synsets are classified into forty-five groups which include lexical categories (nouns, verbs, adjectives and adverbs) and semantic groupings (person, phenomenon, sentiment, place, etc.) (fellbaum, ). there are twenty six categories for nouns, fifteen for verbs, three for adjectives and one for adverbs in wn (izquierdo, ). the organisational design of wn helps lexicographers to obtain a structure with which to create and edit a set of words and senses in the same semantic class. these semantic categories are considered to be semantic classes that are more general than senses, in which different wn senses are grouped in a semantic class. it is also possible to use this sense conceptualisation in different languages because all the wn senses (in english) are linked to eurowordnet which contains wns in different languages. it is important to note that despite the fact that wn was generated by following a top-down process, it is difficult to distinguish between those synsets that were originally semantic classes and those that were not. if we wish to obtain the sense categorisation it is therefore necessary to apply a method with which to generate the semantic classes. the main goal of semantic class (sc) generation (izquierdo et al., ) is to reduce the polysemy, and various techniques with which to group senses have therefore been developed. in all cases, senses of the same word have been grouped together, thus reducing polysemy and improving wsd system results. the sc resource consists of a set of base level concepts (blc) obtained from wn using a bottom- up process with hyperonymy relations. for each synset of wn, its blc is obtained from the first local maxim according to its relative number of relations. as a result, semantic classes have a set of blcs that are linked to different synsets. the process follows an ascendant itinerary by using the hyperonymy relations in wn. in the case of one synset having several hyperonyms, the path with the maximum number of relations is selected. the process ends when a set of initial concepts with the synsets selected as blcs for other synsets is selected. in some cases, there are blcs that do not represent an adequate number of concepts or synsets. this situation is avoided by developing a final filtering process for these false blcs, in which those blcs that do not represent a minimum number of concepts are eliminated. each blc therefore has a minimum threshold associated with it, as is described in (izquierdo et al., ). a set of different blcs is obtained by combining different thresholds and types of relations (all or only hypo/hyperonyms). those synsets that do not have a blc associated with them (because their numbers of relations do not have the minimum threshold required) are processed again to select another blc from their hyperonymy relations. this eventually results in a new number of labels with which to categorise a set of senses. scs are applied with the use of different repositories. those currently being used are wn . and wn . . depository fina ncia l institution, ba nk, ba nking concern, ba nking compa ny object physical entity collection agent corporation organization group accepted manuscript a cc ep te d m a n u sc ri pt . . sentiwordnet sentiwordnet (swn) (esuli and sebastiani, , baccianella et al., ) is a lexical resource in which each synset is associated with three different sentiment categories: objectivity, positivity and negativity. each category has a score that moves from [ .. ] and the addition of the three scores is always . this means that one synset could have scores that are different from for each of the categories. for example, atrocious# in our proposal we use sentiwordnet . . this version was obtained in two steps: first a semi- supervised learning phase and then a random walk. the first step consists of four sub-steps, as in the first version of swn:  semi-supervised learning o first sub-step. two small sets of seeds are used, one in which all the synsets contain seven paradigmatically positive terms and the other in which all the synsets contain seven paradigmatically negative terms (turney and littman, ). both are expanded by using wn binary relations in order to connect synsets with the same polarity. the expansion is carried out with a specific k radius. o second sub-step. a previous set of synsets is used with another set of synsets from the objectivity category in order to build a set of training synsets with which to create a ternary classifier (one synset is classified as pos, neg or obj). the classifier uses synset glosses to conduct the process. in swn . , a bag of words model is used in which the bag of words is obtained from gloss words (those frequent words that are most important). in swn . , rather than using words from glosses (with the ambiguity problem) a bag of synsets is used. this sub-step is improved by varying the displacement radius. o third sub-step. all wn synsets are classified as being either pos, neg or obj via the classifier generated in the second sub-step. o fourth sub-step. the second sub-step can be performed by using different values of the k radius and different supervised learning technologies.  step (random walk). wn . is considered to be a graph and an iterative process is conducted. in this random walk pos(s) and neg(s) (and consequently obj(s)) are determined using the previous step. the process ends when the iterations have converged. there are , wn synset descriptions in our proposal. . . extendedwordnet extendedwordnet (xwn) (sanda m. harabagiu, ) is a lexical resource created at the university of texas. this resource was developed to improve semantic information in wn in its different versions. the goal is to add semantic information to glosses and establish new relations among words (now labelled with their senses) from glosses and synsets. the new annotated version of wn only uses information from gloss definitions, and gloss examples and other information are discarded. extendedwordnet was created by applying three processes:  syntactic analysis. a voting process with two syntactic analysers was applied using the outputs of (brill, ) as inputs. the content of the glosses was extended with: o adverbs. glosses of adverbs were extended by adding an adverb + is at the beginning of the gloss and a full stop at the end of the definition. for example, the gloss for automatically would be: automatically is in a reflex manner. a direct semantic annotation is thus obtained between word and gloss. o adjectives. glosses of adjectives were extended by adding an adjective + is something at the beginning of the gloss and a full stop at the end of the definition. for example, the gloss for pure would be: pure is something not mixed. o verbs. glosses of verbs were extended by adding to + verb + is to at the beginning of the gloss and a full stop at the end of the definition. for example, the gloss for shed would be: to shed is to cast off hair, skin, horn, or feathers. http://gandalf.aksis.uib.no/lrec /pdf/ _pdf.pdf http://xwn.hlt.utdallas.edu/ accepted manuscript a cc ep te d m a n u sc ri pt o nouns. glosses of nouns were extended by adding noun + is at the beginning of the gloss and a full stop at the end of the definition. for example, the gloss for play would be: play is the act using a sword (or other weapon) vigorously and skilfully.  logical analysis. after applying a syntactic analysis, a logical form transformation is then applied. syntactic relations, syntactic objects, prepositional links, complex nominalizations, etc.  semantic analysis. two versions were created: automatic and manual. automatic annotation was applied by using one specific system to disambiguate wn glosses (xwn wsd) and another system to disambiguate free text. the decision process was a voting system that selected those senses with a coincidence between both systems with a precision of %. there were three different categories for semantic annotation (the verbs to be and to have were managed in a special way): o gold: manual annotation. o silver: both wsd methods returned the same value. o normal: only uses xwn wsd value. this annotation obtained % coverage and % precision. words labelled with the same sense of both systems obtained % precision. in our approach , relations were integrated using xwn . and , using xwn . . . semantic integration resources additional information with which to solve different nlp problems was obtained by using different semantic integration resources (isr). however, one of the main problems is decentralisation. wn can solve this problem because it has been used as a basic resource to build others. different resources currently use wn as basis to build their structures. for example:  multiwordnet (mwn) (pianta et al., ) is a project that is being carried out by the itc-irst group in trento, italy in order to build an italian wn that is aligned with the english wn (princeton). its first version contains around , italian words that are organised in , synsets and their connections to english synsets. multiwordnet is different from eurowordnet because at least two methods can be used to build a multilingual wn. the first method was used in the eurowordnet project and consists of building the wns of specific languages independently, while correspondences are found in a second phase (vossen, ). the second method was used in multiwordnet and consists of building wns of specific languages from semantic relations in the english wn. new synsets are obtained from english synsets. in mwn, the domain information has been automatically transferred from english to italian, thus obtaining wnd (bentivogli et al., ). eurowordnet (ewn) (dorr and castellón, , vossen, ) was developed to align english, spanish, dutch, italian, german, french, czech and estonian. new versions included swedish, norwegian, danish, greek, portuguese, basque, catalan, romanian, lithuanian, russian, bulgarian and slovenian. the ili (inter-lingual-index) is used to connect each language (vossen et al., ). the use of ili allows closer senses to be obtained among each language, thus reducing polysemy and obtaining a large connection among different languages.  meaning: the multilingual central repository (meaning project mcr) (atserias et al., ) is integrated into the ewn framework with five local wns, including the english wn, and uses an improved version of superior concept ontology of ewn, mwn domains, sumo (suggested upper merged ontology) (zouaq et al., ) and new semantic relations from corpus. the first version of mcr includes only conceptual knowledge with semantic relations among synsets from local wns. the last version of mcr integrates: o ili based on wn . , including base concepts from ewn, superior concept ontology of ewn, multiwordnet domains and sumo; o local wns (basque, catalan, italian and spanish) related to ili, including wn english versions . , . , . and . . . o semantic preferences collections, from semcor and bnc, and nominal entities.  uby: this is a large-scale lexical-semantic resource for nlp based on the iso standard lexical markup framework (lmf). uby combines a wide range of information from expert and collaboratively constructed resources for english and german. it presents a web browser and an api http://multiwordnet.itc.it accepted manuscript a cc ep te d m a n u sc ri pt that provides facilities that allow this tool to be used. uby currently holds structurally and semantically interoperable versions of ten resources in two languages: o englishwordnet, wiktionary, wikipedia, framenet and verbnet, o german wikipedia, wiktionary, germanet and imslex-subcat, and multilingual omega wiki.  babelnet: this is a very large multilingual semantic network with millions of concepts obtained from: o an integration of wordnet and wikipedia based on an automatic mapping algorithm, and o translations of the concepts (i.e., english wikipedia pages and wordnet synsets) based on wikipedia cross-language links and the output of a machine translation system. as will be noted, these examples attempt to build semantic networks with a common interface. in most cases, isrs apply lexical integration with a few conceptual resources but in our proposal isr-wn provides semantic information and other types of relations. isrs are resources that help to improve the results of tasks such as: document classification, entity discrimination, author detection, etc. the improvement is owing to the fact that isrs provide the contexts analysed with additional information (for example, subjectivity detection, contextual domain, etc). in this work we present a resource that provides navigability from any word sense in wn, domain labels (and sentiments from wna), sumo categories or semantic classes through the use of semantic relations, as shown in fig . as a result, we can extract the multidimensionality of each sentence through the use of all the concepts and words related in a semantic network. we can also detect sentiment polarities in each sentence from swn and relate them with wnd, wna, sumo and sc. fig . sentence semantic characteristics extraction (only a few labels). after using the wn interface to study different lexical and conceptual resources, our goal is to develop a tool with which to align different wn-based resources and exploit all their relations: hyperonymy, meronymy, synonymy, etc. the result will be a graph with a set of nodes that represent wn synsets, wnd concepts, sumo categories, wna concepts, sc labels or swn sentiment polarities. trait social science quality pedagogyfactotum administration root domain free time wna wnd sumo subjective assessment attribute removing occupational role educational process organization normative attribute relational attribute attribute abstract entity group collection object physical social role organizational process international process process transfer motion use.v. be.v. exist.v. be.v. be.v. be.v. constitute.v. be.v. equal.v. comunicate.v. metalic_element.n. strike.v. travel.v. move.v. get_rid_of.v. sell.v. geographical_area.n. natural_process.n. content.n. teacher.n. change_of_state.n. structure.n. cognition.n. organization.n. gathering.n. sc positive swn negative swn objetive swn “but it is unfair to dump on teachers as distinct from the educationalestablishment” accepted manuscript a cc ep te d m a n u sc ri pt . new integrated resource: isr-wn our semantic integration proposal is denominated as isr-wn (integrated semantic resource aligned with wordnet). it consists of the integration of different resources that are isolated but which can be aligned with wn. considering that wordnet provides synsets (s) that contain a set of synonym words and wnd, wna, sumo, sc and swn provide a set of concepts (c) we can relate each concept with different synsets from wordnet. therefore, isr-wn is a lexical knowledge base (lkb) that is represented as a complete non directed graph . vertexes are represented with concepts or senses , that is, . relations between two vertexes i, j are represented with an edge . according to the different resources taken into account, we have developed the integration knowledge base of isr-wn based on the following resources: wn, wnd, sumo, wna, semantic classes (sc), swn and extended wordnet (xwn) . and . . note that xwn provides additional semantic relations among wordnet synsets. bellow, we describe each version and the integration process in each case. . integration process the integration process takes into account the above mentioned resources. the aim is to obtain a new environment from which to retrieve semantic information using a unified set of resources. it is necessary to mention that wn is one of the most frequently used resources during nlp, since it provides a sense inventory. its possibilities have motivated several researchers to develop taxonomies with new semantic information (magnini and cavaglia, , valitutti et al., , niles and pease, , niles, , sara and daniele, , forner, , strapparava and valitutti, ). with regard to the resulting resources, we can mention sumo (niles and pease, ) , wnd (sara and daniele, ), wna (strapparava and valitutti, ), sc (izquierdo et al., ), swn (esuli and sebastiani, ), extended wordnet (sanda m. harabagiu, ) and others. several authors such as (gliozzo et al., , magnini et al., , magnini et al., july , vázquez, , vázquez et al., ) have developed methods and systems based on these resources, and have also demonstrated improvements in several tasks: information extraction, automatic summarization, document indexing and lexical disambiguation. however, these authors have developed their approaches using one or two semantic resources, since there is no tool or resource with which to integrate all the semantic resources that are mapped onto wn. bearing in mind both this fact and researchers‟ current need to use a single tool or resource to obtain semantic information from different resources, this work is focused on integrating the greatest possible quantity of semantic resources mapped onto wn into a single tool. isr-wn includes wn as its lexical nucleus because its internal structure and relations provide relevant information for many nlp tasks. fig represents the conceptual model, in which each dimension (resource) is aligned with wn through the use of semantic interconnections (gutiérrez et al., a) (gutiérrez et al., a). the main challenge as regards the integration consists of dealing with different versions of wn and different versions of the rest of the resources involved in this proposal. it is therefore necessary to match the mappings to the wn versions used by each resource. see fig . this integration resource involves a set of semantic resources considering the following element distribution: , synsets for wn . and , synsets for wn . ; wnd (with labels), wna (with labels), sumo (with labels), swn (with , labels), sc (with , labels), and xwn ( , relations for xwn . and , new synset relations for xwn . ). this is known as “integration of semantic resources based on wordnet” (isr-wn) (gutiérrez et al., a) (gutiérrez et al., a). it is important to note that all the labels and relations included in both versions have been reused from the resources cited. the elements of which isr-wn is composed are shown in section . this is a non-directed graph because each relation connects two vertexes, and and connects and in both senses and is denoted with another semantic relation. the internal relations of wn are described in http://wordnet.princeton.edu/man/wninput. wn.html. , nouns, , adjectives, , adverbs and , verbs. , nouns, , adjectives, , adverbs and , verbs. accepted manuscript a cc ep te d m a n u sc ri pt fig . logic model of the integration of semantic resources (isr-wn). . integration architecture this section describes some of the process features used to integrate all the resources by using wn as an interlinking nucleus. this process takes into account each of their particularities and integrated resource annotations based on different wn versions, in this case by considering version . and . . despite the fact that this version only uses the english language, it can be extensible to other languages if the interlingua index (ili) of ewn (dorr and castellón, , vossen, ) or a similar technology is used to align it to them. fig shows wn synsets linked to several taxonomies (sumo, wnd and wna), and descriptions of sc and swn respectively. in many cases the links are established through the use of mapping files, thus allowing the resources tagged in the wn versions to be interlinked. this results in the creation of an enriched semantic graph that is highly suitable for nlp applications. fig shows how all the semantic resources (the taxonomies of sumo (in green), wnd (in purple) and wna (in red), the sc labels (in orange) and the swn descriptions (in grey) are linked to the wn synsets. particular aspects from each of the resources involved have therefore been taken into consideration. accepted manuscript a cc ep te d m a n u sc ri pt fig . architecture. with regard to integrating all the resources mentioned, mapping files are suitable for interlinking different wn versions, and are used by the semantic resources to map their labels with wn definitions. owing to the word sense granularity of the wn versions (word senses that exist in one version cannot exist in another) many word senses cannot be taken into account when a mapping occurs. a suitable solution to this problem is to navigate all the necessary mapping files in order to take into account the greatest quantity of possible interlinks. for instance, wna has been used in this proposal, and has been tagged with wn . , while swn has been tagged with . and . and sc with . . there is, therefore, no need to use mapping files to integrate them. it is worth noting that those semantic resources that are annotated with both wn . and . reduce the lost interlinks to %. fig shows a model that allows one of the two nuclei, wn . or . , to be chosen depending on the user‟s needs. this decision is not limited to the inclusion of other wn versions for further isr-wn approaches. as can be seen when comparing the second version to the first of wn, the former has new semantic labels and the option of selecting the wn nucleus in order to build the semantic graph on demand. this integration model therefore reuses all the semantic relations included in the princeton wn. furthermore, in isr-wn the sumo categories also keep their relations with their respective wn mappings; moreover, the semantic relations of wnd and wna are now hypernym (owing to the is_a relation) and hyponym (owing to the is_child relation) in order for their hierarchy taxonomy to make sense. pertainym is, correspondingly, the semantic relation used to link the wn synset to each label of both taxonomies (e.g. a synset can pertain to one or many wnd and wna labels). it is important to stress that the wna for version . has been populated with new affective relations, which do not exist in wna . (e.g. entailment and cause). new relations in wna . can therefore link verbs, adjectives and adverbs with those nouns from which these are derived. all these considerations have been taken into account when developing isr-wn resource. the resulting semantic interconnections of this knowledge base permit navigation among all the resources involved. for instance, if the english word atrocious is explored using wn . as a nucleus, a list of wn synsets can be obtained as regards this word. once a synset (e.g. offset ) has been selected from its representative list, the following information is retrieved.  atrocious [adjective] o similar-to ( alarming [adjective]) o pertainym (psychological_feature [domain]) o hypernym (subjetive_assessment_attribute [sumo]) accepted manuscript a cc ep te d m a n u sc ri pt o pertainym (emotion [affect] ) o pertainym (horror [affect] o sentiwn-description ( atrocious# pos: |neg: . |obj: . [sentiwordnet]) o cause ( horror [noun]) as will be noted, the offset, part-of-speech, synonym list and gloss are retrieved for each wn synset. the description of each semantic label from each particular resource linked to the synset being explored is also retrieved. as a search tool, isr-wn also allows labels to be discovered in all the resources included, thus providing theirs associated information and establishing semantic navigation in order to find interesting semantic paths. an interesting case may be discussed as regards the new features introduced in this integration resource for wna. for instance, in the case of exploring the label horror from wna, we find that this label is not linked directly with the third synset of atrocious (e.g atrocious# with offset ). moreover, it has been assumed that if a synset (e.g. atrocious# ) is linked with a noun (e.g. horror# ) by an affective relation (e.g. entailment, cause) which has been obtained from wna . , the synset will be linked to all the wna labels linked to its noun. upon applying this procedure the synset horror# , which is linked with the wna label horror provokes atrocious# to be additionally linked to horror. fig shows the aforementioned example after applying the previously mentioned procedure, focusing on the synset atrociuos# when linked to two wna labels (e.g. emotion and horror). fig . isr-wn including new affective relation from wna . . as part of the semantic improvement strategy for isr-wn, the semantic links suggested by xwn . and xwn . have also been considered. their semantic relations to the synsets for isr-wn have been tagged as follows:  xwn _relation_as_synset,  xwn _relation_as_gloss,  xwn _relation_as_synset,  xwn _relation_as_gloss. different features are involved in the resulting semantic graph depending on the nucleus selected (wn . or . ) in order to load the semantic knowledge base on demand. it is important to highlight that relation_as_synset and relation_as_gloss represent a bidirectional relation between synset and gloss. in the case of using wn . as a nucleus, , and , relations will be involved, and are from xwn . and xwn . , respectively. however, when wn . is used to load isr-wn , relations from xwn . will be involved, in addition to , relations from xwn . . table shows all the semantic relations included in this integration resource. note that xwn . is annotated with wn . , and it is therefore necessary to apply a conversion of wn versions in order to map their offset of version . onto the wn versions used as nucleus ( . or . ). the integration of xwn . , which is annotated with wn . , is performed in the same way. these mappings result in some relations being missed owing to the conversion issues mentioned above. an in-depth analysis of this can be found in the evaluation section. accepted manuscript a cc ep te d m a n u sc ri pt an exploration of isr-wn is presented as follows, in which wn . is used to load the knowledge base. this exploration reuses the example of the word love, in which an increase in the semantic information is evident as regards the target synset:  love [verb] o antonym ( hate, detest [verb]) o hyponym ( love [verb]) o hyponym ( care_for, cherish, hold_dear, teasure [verb]) o hyponym ( dote [verb]) o hyponym ( adore [verb]) o pertainym (factotum [domain]) o hypernym (intentional_psychological_process [sumo]) o pertainym (emotion [affect]) o pertainym (love [affect]) o stative ( love [noun]) o hypernym (love.v. [semanticclass]) o sentiwn-description ( love# pos: . |neg: |obj: . [sentiwordnet]) o xwn _relation_as_synset ( affection, affectionateness, fondness, tenderness, heart, warmheartedness [noun]) o xwn _relation_as_synset ( liking [noun]) o xwn _relation_as_synset ( have, have_got, hold [verb]) o xwn _relation_as_synset ( great [adjective]) o xwn _relation_as_gloss ( sweetheart, sweetie, steady, truelove [noun]) o xwn _relation_as_gloss ( patriot, nationalist [noun]) o xwn _relation_as_gloss ( bibliophile, booklover, book_lover [noun]) o xwn _relation_as_gloss ( amorist [noun] o xwn _relation_as_gloss ( lover [noun]) o xwn _relation_as_gloss ( brunnhilde, brynhild [noun]) o xwn _relation_as_gloss ( psyche [noun]) o xwn _relation_as_gloss ( adonis [noun]) o xwn _relation_as_gloss ( unloved ,not loved [adjective]) o xwn _relation_as_gloss ( hyperbaton [noun]) o xwn _relation_as_gloss ( lover [noun]) o xwn _relation_as_gloss ( bibliophile, booklover, book_lover [noun]) o xwn _relation_as_gloss ( strawflower, golden_everlasting, yellow_paper_daisy, helichrysum_bracteatum [noun]) o xwn _relation_as_synset ( affection, affectionateness, fondness, tenderness, heart, warmheartedness [noun]) o xwn _relation_as_synset ( liking [noun]) o xwn _relation_as_synset ( adore [verb]) o xwn _relation_as_synset ( idolize, worship, hero-worship, revere [verb]) as can be appreciated in this example, the exploration for the same word love (offset for the verb love) has been considerably enriched. the semantic information obtained combines many labels from different resources and establishes new relations among wn synsets. this integration of resources allows to discovery new semantic interconnections never before shown in a single tool. interesting features such as wnd or sumo concepts with positive or negative tendencies could also be mentioned. to return to the example of the word love, an exploration of one level in depth has been applied. however, if we were to navigate in greater depth, many more semantic elements would be discovered. table shows all the semantic relations. this table is used as a matrix to relate each resource by rows. note that in table each relation in isr-wn has one inverse relation. most of the relation pairs have been taken from wn and the other resources involved, while the rest (those marked with *) have been created in this research work. the aim of creating bilateral relations has been to allow forwards and backwards navigation through the semantic links. http://wordnet.princeton.edu/man/wninput. wn.html accepted manuscript a cc ep te d m a n u sc ri pt wn wnd wna sumo sc swn wn also see pertainym pertainym hypernym hypernym sentiwn_description* antonym hyponym hyponym attribute cause derivationally related form derived from adjective domain of synset - region domain of synset - topic domain of synset - usage entailment hypernym hyponym instance hypernym instance hyponym member holonym member meronym member of this domain - region member of this domain - topic member of this domain - usage part holonym part meronym participle of verb pertainym (pertains to noun) similar to substance holonym substance meronym verb group xwn _relation_as_gloss* xwn _relation_as_synset* xwn _relation_as_gloss* xwn _relation_as_synset* wnd pertainym hypernym hyponym wna pertainym hypernym hyponym sumo hypernym hypernym hyponym hyponym sc hypernym hypernym hyponym hyponym swn sentiwn_description* table . relations between semantic labels on isr-wn. bidirectional relations on wordnet pointer reflect antonym antonym hyponym hypernym hypernym hyponym instance hyponym instance hypernym instance hypernym instance hyponym holonym meronym meronym holonym similar to similar to attribute attribute verb group verb group derivationally related derivationally related domain of synset member of doman bidirectional relations added by isr-wn accepted manuscript a cc ep te d m a n u sc ri pt pointer reflect entailment cause participle participle pertainym pertainym xwn _relation_as_gloss* xwn _relation_as_synset* xwn _relation_as_gloss* xwn _relation_as_synset* sentiwn_description* sentiwn_description* stative cause synonymy synonymy table . bidirectional relations on isr-wn. . integration process evaluation in this section the results obtained are analysed by taking into consideration the semantic elements that are available for the integration and those eventually integrated. once developed isr-wn following the aforementioned process, an evaluation took place. this was carried out in order to obtain a statistic assessment of the quantity of wn synsets that should be linked as regards those established. two evaluations were applied based on the semantic knowledge interlinked depending on the nucleus (wn . or wn . ) selected for loading the semantic graph. . evaluation based on wn . as a nucleus table shows a distribution of the semantic labels involved in isr-wn with wn . as a nucleus. in this analysis the greatest possible quantity of labels has been extracted for inclusion, and the labels which have not been included are justified below. an important factor that supports the % of alignments of wnd, sumo and sc has been the reduction in the usage of mapping files. this was possible since these three resources were built based on wn . and the current evaluated nucleus is wn . . wnd . sumo wna . - . sc swn . xwn . xwn . mean # labels , , - - synsets to link n , , , , , - - a , , , - , - - v , , , , - - r , , - , - - total of synsets to link , , , , , , , linked synsets n , , , , , - - a , , , , - - v , , , , - - r , , , - - total of linked synsets , , , , , , , difference , , , linked % . . . . . . . . table . synsets linked to each resource by using wn . as a nucleus. in this integration resource we have considered both wna versions ( . and . ) to be a single resource and have mixed them. this mixture has been developed by involving the taxonomy of wna . and the wn mappings from both versions ( . and . ). however, a difference in labels (as regards all linked synsets with regard to those synsets that should be linked) between wna . and wna . has caused missing links (see table ). this special case is owing to the fact that most of the wna . labels are included in wna . , but several wna . labels are not included in wna . . some links have therefore been removed for the integration. all the affective labels not considered by isr-wn because they are not represented in the wna . taxonomy are shown as follows: attitude, emotional response, psy, man, sympathy, sta, softheartedness, joy-pride, identification, levity-gaiety, general-gaiety, empathy, positive-concern, compatibility, kindheartedness and buck-fever. note that the taxonomies used are the most recent for wnd and wna (i.e. wnd . and wna . ). accepted manuscript a cc ep te d m a n u sc ri pt wn mapping files have also been considered for the alignment of wna to wn. wna . has a peculiar feature consisting of the fact that only nouns are linked to affective labels. adjectives, verbs and adverbs are therefore linked to these nouns at sense level by means of derived relations (i.e. entailment and cause). the description of this issue can be found in the section “ . integration process”. furthermore, when evaluating swn integration based on wn . as a nucleus, it was found that the main difference between the mismatching of swn . alignments had occurred because this resource had been developed on the basis of a very different wn version, wn . . it is therefore possible to find that several swn features that are annotated with wn . do not exist for wn . . this justifies all the lost links shown in table . table also shows how the integration of xwn . reached . %. moreover, with xwn . it reached a lower integration of . %. a similar phenomenon occurred with swn . in which several mapping wn versions were involved. this is responsible for the low enrichment of isr-wn as regards xwn . when wn . is used as a nucleus. graph-based approaches, such as (gutiérrez, ) which uses isr-wn with xwn . , cannot therefore obtain the same relevant results as those obtained using xwn . . once isr-wn is loaded with wn . , it contains a total of , vertexes, of which , represent wn synsets, wnd domains, sumo categories, wna affective labels, , sc semantic classes and, finally, , swn descriptions. isr-wn therefore includes a total of , semantic relations with wn . as a nucleus. . evaluation based on wn . as a nucleus table shows a similar comparative study to table , but here wn . is considered as a nucleus. wnd . sumo wna . - . sc swn . xwn . xwn . mean # labels , , - - synsets to link n , , , , , - - a , , , - , - - v , , , , - - r , , - , - - total of synsets to link , , , , , , , linked synsets n , , , , , - - a , , , , - - v , , , , - - r , , , - - total of linked synsets , , , , , , , difference , , linked % . . . . . . . . table . synsets linked to each resource by using wn . as a nucleus. as can be seen in table , in isr-wn sc has some missing links, since it has been developed using wn . and has to connect with wn . through the use of mapping files. when this occurs, some synsets are not considered since some of them do not exist for both versions ( . and . ). it will, however, be noted that swn links increase in comparison to the previous distribution table because the mapping versions are reduced. wna . and wna . have been distributed together in both tables, since both have been seen as a single resource. all the wna . labels and the synset-label relations from wna . and . have therefore been considered. however, only all those wna . labels that do not exist in the wna . taxonomy have not been considered. both versions of wna are thus considered to be a single resource. as table shows, when using wn . as a nucleus the incorporation of xwn increases significantly, particularly in the case of xwn . . it is most appropriate to use isr-wn with wn . since the mean of the integration percentage was . %, while for wn . it was . %. it is therefore evident that the best integration of resources based on wn occurs when wn . is taken into account. in this case, , vertices have been computed, of which , are wn synsets, are wnd domains, are sumo categories, are wna affective domains, , are sc labels and , are swn descriptions. all these vertices are related to , , semantic relations. accepted manuscript a cc ep te d m a n u sc ri pt . nlp tasks description as we have mentioned before, our goal in this work is to provide a semantic framework able to enrich texts with semantic information and use this new information to classify, obtain opinions or recommend similar texts in open domains. to do this, we use isr-wn as knowledge source for different nlp tasks. basically, we have conducted four different tasks to add new information from different point of views: - word sense disambiguation (wsd). we need to obtain the correct sense of each word in order to obtain accurate results. due to the fact that one word has different meanings depending on the context, it is necessary to decide which word sense is the correct because it will determine a proper understanding. - semantic similarity. the goal of this task is to discover similar content. according to the words, senses, content structure, etc., we will be able to obtain similar opinions that could be expressed with different words but the meaning is still similar. - domain classification. this task will provide the most relevant domains of a text a set of texts. this information will be useful to annotate texts in order to recommend similar content. - sentiment analysis. the goal of this task is to decide whether an opinion is positive or negative automatically. moreover, many scientific researchers have used the semantic features that isr-wn provides to deal with different nlp tasks. some of these tasks and their contributions are shown below. notice that most of semantic nlp challenges are based on these wn versions, therefore this integration resource is very useful for dealing with these tasks (see http://senseval.org/). word sense disambiguation approaches (wsd): in order to create effective nlp systems it is necessary to transform the information extracted from the words in plain text into a conceptual level to detect meaningful word senses. in wsd the goal is to determine the words‟ senses in a text in which a word may have different senses depending on the context that it appears in. so that, the main purpose of the above listed approaches is to automatically choose the intended sense (meaning) of a word in a particular context. to be able to do that, it can be used two types of approaches: knowledge based approaches (zhibiao and martha, ), (leacock and chodorow, ) and corpus-based approaches (peter, ). the former requires lexical resources such as wordnet, roget's thesaurus , babelnet , dbpedia , etc. to obtain semantic similarities, while the latter uses co-occurrences to measure the similarity between words. in our case, by conducting knowledge based approaches through isr-wn, we find all the possible synsets for the word in wordnet and also their links to different concepts of the aforementioned resources (i.e. wnd, wna, sumo, sc) being able to group word synsets under semantic concepts facilitating their distinctions depending on the context. for example for the concept “sport” of wnd we can find three word senses with respect to the word “exercise”, two for “school”, one for “sociology” and one for “pedagogy”. so that, if we determine the context of the sentence where “exercise” is being used, we can reduce the number of possible meanings to assign. the advantage to use isr-wn is we can do this procedure not just using wnd, we can also utilise the rest of the conceptual resources at the same platform. wsd can be also considered a task for doing text‟s semantic enrichment, since by applying this technique and depending on the knowledge base used; the processed text can be tagged or linked with different sources. this phenomenon can be associated to the task entity linking in which the matching of a textual entity mention to a knowledge base is performed. this entity mention can be a wikipedia page, dbpedia entry, or a specific uri that identify an entity, that is a canonical entry for that entity (rao et al., ). the reason whereby we are introducing the entity linking task is because we also can integrate babelnet with isr-wn for dealing with it, even for different languages. such is the case of (gutiérrez et al., ) in semeval- for task “multilingual word sense disambiguation” . a list of wsd approaches that have been supported the conceptual integration of isr-wn are described below. among them, the second approach not just describes a word sense disambiguation approach, it also describes how by using isr-wn relevant conceptual trees of every conceptual resource http://www.thesaurus.com/roget http://babelnet.org/ http://dbpedia.org http://www.cs.york.ac.uk/semeval- /task / accepted manuscript a cc ep te d m a n u sc ri pt integrated in isr-wn, can be obtained. these trees can be interpreted as a manner of obtain those concepts that classify a text. for example, in the sentence taken from an imdb‟s user review “grey's anatomy stars ellen pompeo (who has starred in a few movies but was never really noticeable) as the narrator - meredith grey”, using this approach based on wnd this text can be classified with the following concepts [“cinema”, “art”, “humanities”]. the following researches, which are based on isr- wn, support the tasks mentioned:  participation of the umcc-dlsi research team in semeval- for task : all-words word sense disambiguation on specific domain (agirre et al., ). a description of this approach can be found in (gutiérrez et al., b).  research paper (gutiérrez et al., b) which was discussed in the ranlp conference and which combined isr-wn (considering , wn, wnd, wna and sumo) and word sense frequencies to solve word sense disambiguation. this proposal attained higher results in comparison to top proposals at a world level.  research paper (gutiérrez et al., d) which was discussed in the nldb conference and which used isr-wn (considering wn, wnd, wna and sumo) as a graph-based approach to deal with word sense disambiguation. this proposal attained important results in comparison with the reported submissions of senseval- (cotton et al., ). participation of the umcc-dlsi research team in semeval- for task “multilingual word sense disambiguation” . this approach attained the first position of the rank. a description of this approach can be found in (gutiérrez et al., ) in order to show the results obtained by wsd approaches supported by isr-wn we describe the two types used in this work: graph-based (gutiérrez et al., ) and tree-based (gutiérrez et al., b). the graph-based approach takes into account the whole isr-wn graph as configuration, thus it includes all isr-wn resources and their links for applying wsd tasks. as part of the evaluation of this approach we participated in the semeval- for task “multilingual word sense disambiguation” campaign. for example, for evaluating wsd in english language, , instances of babelnet, , of wikipedia and , of wordnet were considered. note that in this case isr-wn and babelnet, were aligned through wordnet, since wordnet is common for both. based on the corpus provided by task “multilingual word sense disambiguation”, our wsd approach was able to be placed at the top of the campaign ranking, reaching a f around . %, . % and . % for babelnet, wikipedia and wn respectively. more details about the approach and its experiments on semeval- campaign see (gutiérrez et al., ). resources files experiments wn wna sumo wnd mfs precision recall f f difference with the best system corpus d .txt [ instances] exp (mfs) x . . . exp x . . . exp x . . . exp x . . . exp x . . . exp (voting) all isr-wn as configuration x x x x x . . . whole corpus (d .txt[ instances], d .txt[ instances], d .txt[ instances]) exp (mfs) x . . . . exp (voting) all isr-wn as configuration x x x x x . . . . best system of senseval- . . . average of the senseval- results . . . worst system of senseval- . . . table . evaluation results of the tree based wsd approach with senseval- corpora http://www.imdb.com/ http://lml.bas.bg/ranlp / http://gplsi.dlsi.ua.es/congresos/nldb / http://www.cs.york.ac.uk/semeval- / https://www.cs.york.ac.uk/semeval- /task /index.php% fid=data.html accepted manuscript a cc ep te d m a n u sc ri pt in order to show how to use each resource individually (i.e. wn, wna, sumo and wnd) we describe another wsd approach: tree-based (gutiérrez et al., b). in next table we show the evaluation, using the corpora provided by the public rada mihalcea‟s repository. this repository includes , instances from the senseval- competition “the english all-words task” (cotton et al., ). table shows the results by using different configurations: each resource individually or all together. note that the most frequent word sense (mfs) is also considered as a dimension for the voting. more details in (gutiérrez et al., b). opinion mining approaches: nowadays, opinion mining (om), also known as sentiment analysis (sa), as part of an nlp task has become a popular discipline due to its wide-relatedness to social media behaviour studies. om is commonly used to analyse the comments that people post on social networks. also, it allows identifying the preferences and criteria of users about situations, events, products, brands, etc. in order to exploit the potential that provides isr-wn with the integration of semantic, conceptual, affective and sentiment scoring resources some approaches have been produced for the detection of opinions, relevance, and polarity classification and also measuring the impact of the sentiment polarity detection on textual entailment tasks. those approaches are the following:  the multidimensional point of view of isr-wn has been useful in the opinion mining area. in this case it was presented in the wassa‟ workshop , dealing with three opinion mining tasks in the moat (ntcir multilingual opinion analysis task ) competition system. these tasks consist of identifying whether or not a sentence represents an opinion. another task involves identifying the polarity of the opinion sentence, while the last one consists of aligning questions with topics. this proposal used isr-wn to extract relevant concepts which represent the sentences analysed. the opinion mining tasks were carried out by taking these concepts and linking them with swn (in enriched isr-wn) sentiment polarities. this research attained relevant results which could be comparable with the first places attained in the moat campaign (gutiérrez et al., c)  paper research “approaching textual entailment with sentiment polarity” (fernández et al., b), discussed in the icai‟ conference: the international conference on artificial intelligence. this takes isr-wn as a knowledge base and uses the (gutiérrez et al., c) approach to determine textual entailment with opinion mining techniques. semantic textual similarity (sts) approaches: sts is related to textual entailment (te) and paraphrase tasks. the main difference is that sts assumes bidirectional graded equivalence between the pair of textual snippets. in case of te, the equivalence is directional (e.g. a student is a person, but a person is not necessarily a student). in addition, sts differs from te and paraphrase in that, rather than being a binary yes/no decision, sts is a similarity-graded notion (e.g. a student is more similar to a person than a dog to a person). this graded bidirectional is useful for nlp tasks such as machine translation (mt), information extraction (ie), question answering (qa), and summarization. several semantic tasks could be added as modules in the sts framework, such as wsd and induction, lexical substitution, semantic role labelling, multiword expression detection and handling, anaphora and co- reference resolution, time and date resolution and named entity, among others”. below can be found different participations on international competitions where sts systems made use of isr-wn a semantic nucleus allowing judging textual similarities from multidimensional perspectives such as from domain and affective points of view and others. the following researches support sts tasks based on isr-wn:  participation of the umcc-dlsi research team in semeval- in task “semantic textual similarity”(agirre et al., ) . a description of this approach can be found in (fernández et al., a).  participation of the umcc-dlsi research team in semeval- task . “umcc_dlsi- (eps): paraphrases detection based on semantic distance” (dávila et al., ) discussed in second joint conference on lexical and computational semantics (*sem). it took isr-wn as a knowledge base to determine the semantic similarity of different short texts. http://gplsi.dlsi.ua.es/congresos/wassa / http://research.nii.ac.jp/ntcir/ntcir-ws /ws-en.html http://aclweb.org/aclwiki/index.php?title=recognizing_textual_entailment the adaptation or alteration of a text or quotation to serve a different purpose from that of the original. http://www.cs.york.ac.uk/semeval- /task / accepted manuscript a cc ep te d m a n u sc ri pt  participation of the umcc-dlsi research team in semeval- in task “semantic textual similarity” . a description of this approach can be found in (chávez et al., ). ongoing approach: in order to capture the semantics of a textual question, we pretend to use the aforementioned approaches (wsd and sts) supported by isr-wn, to make interpretations of both questions written in natural language and ontology node‟s descriptions (and also considering the name of the adjacent nodes) for finding further alignments with existing ontologies. reusing the work below, which has been the first approach by using isr-wn, we would be able to recover those ontology nodes that involve textual descriptions relevant to textual question. the following research is part of this ongoing approach.  paper research “semantic information extraction method on ontologies”(dávila et al., ) discussed in sepln‟ : xxviii congreso de la sociedad española para el procesamiento del lenguaje natural, in which semantic ontology searches are applied on the basis of isr-wn and natural language adaptation. next, we describe the annotation process through a case of study taken from imdb. . case study in order to enrich texts using isr-wn, we have developed different approaches to automatically process textual information and obtain the semantic knowledge associated to each word or to the general context. as a result, we provide a semantic framework suitable of being used as support of content-based recommender systems that is able to find contents that are similar to those already evaluated by the users by considering semantic information. in this section we illustrate the annotation process with an example obtained from imdb (internet movie database) and explain the different approaches used to conduct the annotation process. our experiments are carried out using textual information from imdb. imdb is the world‟s most popular and authoritative source for movie, tv and celebrity content. this website has more than million unique monthly visitors and offers a searchable database of more than million data items including more than million movies, tv and entertainment programs and more than million cast and crew members. as we have mentioned above, imdb provides a lot of information related to movies, tv shows or tv series. however, due to the fact that we are interested in people‟s interest and opinions we only focus our attention over the reviews & ratings section. in this case, for each movie or tv series the reviews & ratings section has a list of reviews from different users that give their opinions about general concepts, feelings, likes and dislikes, etc. our purpose is to collect textual information provided by the reviews of different users about tv series or movies in order to enrich it with semantic information and therefore being able to rate or recommend similar content according to people feelings, interests or likes. with this case study, we aim at demonstrating the practical use of our proposed resource. . semantic enrichment approaches in this section, we briefly describe how to enrich textual information by using isr-wn. as we have mentioned in section , we have integrated a set of different resources in order to obtain new relations and therefore being able to build new connections. with isr-wn we can enrich texts in different dimensions using: affection labels, domain labels, semantic classes, etc. to conduct our experiments and annotate texts we have developed a framework that uses different techniques in order to take into account each dimension of isr-wn. . wsd first of all, it is important to extract the correct senses of words. to do this, our framework uses an unsupervised multilingual approach of wsd. it works using a words window and applies a modification of the personalizing page rank (ppr) algorithm (agirre and soroa, ) considering as knowledge base the isr-wn semantic network. more in detail, it uses senses frequencies to rank the synsets of each http://ixa .si.ehu.es/sts/ http://www.imdb.com accepted manuscript a cc ep te d m a n u sc ri pt lemma (in a descending order) according to a calculated factor of relevance (ppr+freq algorithm, more details see in section or (gutiérrez et al., )). this approach was evaluated in semeval- obtaining the best results of this campaign. an example of the results after processing a text fragment, taken form imdb‟s user reviews, is shown in next table: original fragment format: [word][(word position in the sentence] wsd results format: [position of the word in the sentence][word sense key suggested as appropriated in this context][gloss of the word sense suggested] her mother( ) is( ) the famous( ) surgeon( ) and she is trying( ) to follow( ) in her mother's( ) footsteps( ) ( ) mother% : : :: "a woman who has given birth to a child (also used as a term of address to your mother); \"the mother of three children\" ( ) be% : : :: "have the quality of being; (copula, used with an adjective or a predicate noun); \"john is rich\"; \"this is not a good answer\" ( ) famous% : : :known: " widely known and esteemed; \"a famous actor\"; \"a celebrated musician\"; \"a famed scientist\"; \"an illustrious judge\"; \"a notable historian\"; \"a renowned painter\" surgeon% : : :: "a physician who specializes in surgery " ( ) be% : : :: " have the quality of being; (copula, used with an adjective or a predicate noun); \"john is rich\"; \"this is not a good answer\" ( ) try% : : :: " make an effort or attempt; \"he tried to shake off his fears\"; \"the infant had essayed a few wobbly steps\"; \"the police attempted to stop the thief\"; \"he sought to improve himself\"; \"she always seeks to do good in the world\" " ( ) follow% : : ::" to travel behind, go after, come after; \"the ducklings followed their mother around the pond\"; \"please follow the guide through the museum\" ( ) mother% : : :: " a woman who has given birth to a child (also used as a term of address to your mother); \"the mother of three children\" ( ) footstep% : : :: " the distance covered by a step; \"he stepped off ten paces from the old tree and began to dig\" " table : disambiguation example as table shows, each word (nouns, verbs, adjectives and adverbs) is disambiguated. the annotation process after disambiguating each word obtains: lemma and word sense (mother% : : ::) and a gloss or definition with examples ("a woman who has given birth to a child (also used as a term of address to your mother); \"the mother of three children\"). with this information we are able to discriminate among a set of senses and understand the meaning of words if they appear in different contexts. . concepts and polarities in order to obtain relevant concepts and extract sentiment information we propose an unsupervised knowledge-based approach, described in section before, that uses the rst (relevant semantic trees) technique (gutiérrez et al., c). the result is a set of relevant semantic trees associated to each sentence that provides sentiment polarity values. an example of the relevant semantic tree technique based on wordnet domains resource is shown in fig , where the number between parentheses of each concept indicates the order of relevance from to according to the next fragment. accepted manuscript a cc ep te d m a n u sc ri pt fig . relevant semantic tree (rst) based on wordnet domains (wnd) this tree represents the most relevant domains related to the next text fragment extracted from a review of grey‟s anatomy from imdb: “…the story revolves around her time as an intern and the people she meets and sort of is portrayed as "survival camp for medical students." the minute she arrives at work, she meets christina yang (sandra oh - flawless in her bitchy supporting role), george o'malley (t.r. knight - one word: breakthrough performer), izzie stevens (katherine heighl - very, very believable as a model who is more like the girl-next-door), and alex karev (justin chambers - plays sort of a not-so-likable person). most of all, there is dr. derek shepherd (patrick dempsey - very attractive), the man that meredith had a one- night stand with - he just happens to be her boss. this is a show that wants to be liked….”. as we can see in fig , there are seven domains that are closely related to the meaning of the context, starting from person to telecommunications. moreover, wnd has a hierarchy where all the domain labels ( labels) are connected and fig also shows how the relevant domains are arranged into the hierarchy. this information is useful in order to classify texts into different categories and to obtain related information through a previous categorization. . semantic textual similarity last step to enrich textual information is to extract semantic textual similarities. our approach is a machine learning system (mls) that uses several algorithms to extract all features: similarity measures, lexical-semantic alignment, semantic alignment (fernández et al., a). in order to extract the semantic features it uses the multidimensional resource isr-wn. thus our sts approach provides a value scale to decide whether a pair of contexts can be considered semantically similar or not. the scale has values that go from to . where indicates that there is no semantic relation between two pair of contexts and indicates that a pair of contexts is semantically equivalent. see the evaluation described in next section. . experiment setup and results due to the fact that we need real texts to evaluate our framework, our experiments have been conducted by using texts from imdb user reviews. these texts contain information related to user feelings, expectations, complains, acceptance, etc. moreover, each review has associated an overall rating. in order to achieve comprehensive results we have extracted information related to medical tv series: house m.d., grey anatomy, etc. accordingly to our data set, we randomly have selected reviews to show how texts are annotated. table shows a brief part of the original reviews that have been annotated. tv series/movies original texts grey‟s anatomy if you think abc can't get any better - you're wrong. with the great success over smash-hits, "desperate housewives" and "lost" they also picked up a few shows over mid-season hoping for more success. they got "jake in progress," "eyes" and "grey's anatomy" - but "grey's top level ( ) person doctrines history ( ) heraldry ( ) art ( ) theatre social science ( ) telecommunications ( ) cinema applied science ( ) medicine accepted manuscript a cc ep te d m a n u sc ri pt anatomy" is definitely considered the best out of those three. grey's anatomy stars ellen pompeo (who has starred in a few movies but was never really noticeable) as the narrator - meredith grey. her mother is the famous surgeon and she is trying to follow in her mother's footsteps.… house m.d. let me put it simply. i am a physician, and as an inviolable rule, i hate medical shows. granted, tv series tend to be one dimensional, due to inherent difficulties in the genre, but "doctor shows" are something i avoid like the proverbial plague. and then one evening i caught "house, md" and was completely drawn into the show. in house i find the anti-hero that i've been waiting for in a medical show. the guy who knows everything, but is wrong often enough to keep us all guessing. i enjoy the contrast of house and his cadre of young fresh faced colleagues… tomorrowland when the director of "the incredibles" signed for this film, i was looking forward to the same amount of humour and exhilaration present in that animated masterpiece, something similar to what the little boy expresses when he realizes he can run on water. nothing remotely close occurs in "tomorrowland", a film that suffers from having too big a budget and hardly any original or exciting thoughts. it is also hindered by the fact that almost all of the actors appear clueless and not quite matching their characters. there's something about george clooney… table : original reviews obtained from imdb imdb provides the overall rating for each tv show and movie in their database. for grey‟s anatomy the overall rating is . / from , users, for house m.d. is . / from , users and for tomorrowland . / from , users. after processing each text we obtain a set of different features that we use to enrich the original information. table shows the results of annotation for affect labels and domain labels including a rating of positiveness, negativeness and objectiveness, respectively: title affects p n o domains p n o grey‟s anatomy love annoyance liking anger anxiety . . . person theatre cinema art heraldry medicine telecommunication . . . house m.d. sensation love wonder . . . psychology telecommunication humanities literature theatre medicine person . . . tomorrowland closeness belonging joy behaviour exhilaration sensation . . . racing telecommunication sociology theatre cinema sport radio+tv . . . table : results of annotation of three imdb reviews once got over the annotation process we are able to infer whether the review is positive or negative and also which are the general domains of each context. this information is useful if we want to recommend similar shows to other people because we can analyse the general content and extract common domains. moreover, we are able to rate each show according to people‟s reviews gathering if they are positive, negative or objective. fig shows the annotation of sentiment positiveness of each imdb review in comparison with the real imdb rating: accepted manuscript a cc ep te d m a n u sc ri pt fig . annotation of sentiment positiveness of three imdb reviews as we can observe, our predictions for grey‟s anatomy and tomorrowland are very similar to real imdb rating. in these cases the textual enrichment process obtains very good results and the semantic classes and sumo dimensions are nearby to real ranking. in case of house m.d. there are two points of difference between our predictions and real rating. this is due to the review selected mixture a set of negative and positive reactions to the show. for example, it begins saying “i hate medical shows” but next says “i enjoy the contrast of house and his cadre of young fresh faced colleagues, complete with starched white lab coats, who struggle as much with their professionally imposed constraints, and sense of decorum, as they do with his personality”. so, in this case it is important to remark that despite the fact that there are some negative aspects our framework is able to establish that there is some positive things that must be taken into account. another dimension to enrich texts is provided by the sts (semantic textual similarity) approach. in sts, three features are used: syntactic (similar words, syntactic distances, etc), sentiment and semantics. as a result, we are able to identify similar texts by measuring them in a scale from to (where indicates that there is no semantic relation between a pair of contexts and indicates that both are semantically equivalents). in table we can observe the results obtained after applying sts. the results indicate that there is a slight relation between grey‟s anatomy and house m.d. ( . ) and there is no relation between grey‟s anatomy and tomorrowland ( ). semantic textual similarity grey's anatomy house m.d. tomorrowland grey's anatomy - . . machine learning technique support vector machine approach sts approach described in section table : sts results to measure contexts similarities in order to demonstrate how our framework would help to predict the rating with a set of multiple user reviews, we present an exhaustive evaluation over the first imdb reviews for each mentioned tv shows. table lists the title of each tv show considered on these evaluations and their ratings, which were manually set by the users. title no. review title imdb rating ( - ) grey's anatomy so far, so good addictive show . excellent series http://www.support-vector-machines.org/ . . . . . . . . . . . . . . . . . grey's anatomy house tomorrowland p affect p domain p sumo p semantic class p synset p mean real imdb rating accepted manuscript a cc ep te d m a n u sc ri pt waste of time . boring! . this show is absolutely wonderful!!! wonderful! can someone explain me why this is a top show? . started well, now its just more of the same... . i'm done--and shame on me for taking so long . house abrasive medical doctor saves lives - no matter what the cost. extremely formulaic fantasy medical detective show . ingenious and compelling!!! the greatest medical show ever best show ever! watch it for the characters, not the medicine! . one of the best things on the box . very good and getting better . everybody lies... pllllllease watch a full episode. tomorrowland do not listen to the critics on this one! a vastly under-appreciated tale of promise and hope. a fantastic sci-fi experience a spectacular disney sci-fi thrill ride . misunderstood . why do people hate this movie? maybe a bit childish but enjoyable sci-fi tale none the less. . upbeat positive story for the whole family. . a great movie! . fun kid safe sci fi movie that looks great . offensively shallow . table . list of reviews used for automatically rating these three movies once processed the reviews, positive, negative and objective scores were obtained by a tree based om approach (gutiérrez et al., c) using as configuration: wna (affect), wnd (domain), sumo, sc (semantic class) and wn (synsets). moreover, an additional experiment was conducted to take into account the harmonic mean among all positive outputs per review. as table shows different types of evaluations were performed. one evaluation measure is the pearson correlation that compares the real rating with the positiveness automatically obtained by our framework. another one is the combination of all the outputs with the harmonic mean where we reach a % of correlation. related to each resource, we can notice that wna (affect) obtains the best results of correlation. it means that this resource provides better knowledge to deal with om challenges due to the fact that wna conceptualize and link only affective words. on the other hand, if we pay attention to the evaluation shown in table , it shows the results of the mean difference between the real rating and the positiveness obtained by using each resource. in table , wna reaches the smallest mean difference with . perceptual points of margin. it is very difficult to compare this kind of reviews with om system outputs because the users usually set dramatic scores (e.g. or ) however; their reviews are not as cruel as their scores are. for example, we invite to take a look at the following imdb review: number of table titled “extremely formulaic fantasy medical detective show” with a score of . . in table it has been highlighted those cases in which the ratings are negatives. in those cases our om functionality is able to identify that the scoring should be lower than the others but it is not able to set how much. thus, as future work we propose to carry out new evaluations for introducing om outputs in a machine learning system. at this way the framework could be able to learn from the different patterns provided by theses outputs and suggest more balanced results. it is one of several kinds of average, and in particular one of the pythagorean means. typically, it is appropriate for situations when the average of rates is desired. accepted manuscript a cc ep te d m a n u sc ri pt tit. no. rating ( - ) affect domain sumo semantic class synset harmonic mean p n o p n o p n o p n o p n o p n o g r e y 's a n a to m y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . h o u se . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . t o m o r r o w la n d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . pearson correlation . . . . . . difference (rating-p) mean . . . . . . table . opinion mining evaluation over reviews another evaluation was carried out in order to compare the rating obtained by each movie (by considering reviews per movie) with the rating suggested by our framework. fig shows how our suggestions were very consistent regarding to grey‟s anatomy. however, there are some differences with the results for tomorrowland and house m.d, but not so distant. this is due to our results provide positive sentiment polarities and not specifically user ratings. even though there are some differences, we can still obtain accurate results according to wna or domains dimensions. fig . review rating (rating mean) and om rating (positiveness mean) comparison grey's anatomy house tomorrowland imdb rating (mean) . . . affect positiveness (mean) . . . domain positiveness (mean) . . . sumo positiveness (mean) . . . semantic class positiveness (mean) . . . synset positiveness (mean) . . . mean positiveness (mean) . . . . . . . . . . . . . . accepted manuscript a cc ep te d m a n u sc ri pt we want to remark that our framework is able to provide the om outputs as inputs to a recommender system by different ways. one way is to provide the individual analytics considering each semantic resource (in this case sumo, wnd, wna, sc, and the taxonomy of wn) separately. another way could be considering as input all resources to get a single sentiment analysis result. and finally, we can provide all these outputs with a mean calculation. . discussion according to the experiments conducted and the results obtained, we can conclude that our resource isr- wn and the different approaches that take advantage of it are suitable of being used to acquire enough knowledge to make recommendations for tv-shows or films. moreover, we are also able to classify texts into different labels with the purpose of ranking a set of tv-shows according to their similarities. moreover, due to the fact that user reviews have associated a set of feelings to indicate likes or dislikes we are also able to detect emotions and classify texts into three different sections: positive, negative or objective. in some cases, there are contexts that mixed a set of emotions and can confuse the framework. this is the case of the review for “house m.d.”. however, we can establish the appropriate feelings by taking into account all the context information. we would like to mention another similar work such as (briguez et al., ) that presents a novel framework for the specific domain of movie recommendation. this work proposes a complete set of postulates accounting for both quantitative and qualitative aspects of the movie domain implemented by means of delp (defeasible logic programming). related to our work, there are some differences: first of all, we want to remark that we present a semantic framework as a tool to help recommender systems so, taking into account our proposal, we can mention that the way we provide information to recommend similar movies is established by measuring different dimensions (i.e. feelings, sentiment polarities, word senses, textual similarities), on the contrary in (briguez et al., ) their system uses specific rules to determine which characteristics have to be shared. despite the fact that the process of building specific rules helps to obtain better user recommendations, it is designed to one specific domain. in case of applying this system to another domain it would be necessary to create additional rules to obtain accurate results. on the other hand, our approach can be easily adapted to other domains or languages (since wn is a nucleus for other wn languages ) with minimal changes. another difference with our work is that in (briguez et al., ) the system provides a combination of quantitative and qualitative aspects however, we only have focused on quantitative aspects. the results obtained after mixing qualitative and quantitative aspects demonstrate that the incorporation of qualitative aspects introduce significant improvements. in fact, it would be a reasonable feature to take into consideration for future works. related to evaluation results, we cannot provide an in depth comparison because our datasets are different. moreover, as far as we are concerned we have not found any example in the literature using exactly the same dataset employed in this work. even though our experiments have been carried out with user reviews for tv-shows we can deal with other domains. moreover, we can deal with different languages by using a specific version of wn for each language. this fact has been demonstrated with our participation in different competitions where our framework was evaluated over different languages. one interesting usage of our proposal is shown in fig . through our semantic labels and features provided by our framework, one user could navigate among different tv-shows or movies that have semantic similarities and thus, provide a great experience of knowing automatically which are the movies or shows more appropriated to each one. http://globalwordnet.org/wordnets-in-the-world/ accepted manuscript a cc ep te d m a n u sc ri pt fig . conceptual similarities among different imdb reviews . conclusions and future works in this work our goal has been to enrich texts by measuring similarities among different comments/reviews, grading the positiveness or negativeness by using textual information and using also this information to create a recommendation rate according to people interests. to do this, we have developed a framework that allows the integration of several semantic resources in order to provide to recommender systems enough knowledge for being able to: extract similarities among texts, measure sentiment polarities and obtain a semantic analysis to understand contexts meanings. in order to build the knowledge database of the framework we have selected a set of resources by conducting an in-depth analysis of the different semantic resources used in nlp. just like we have discussed in previous sections, several authors have created integrated resources but most of them have been based on linguistic features rather than on semantic or conceptual features. in this research work we have therefore considered previous studies in order to develop a multidimensional knowledge network that integrates different semantic resources. therefore, an integrated resource (isr-wn) has been proposed. this resource integrates semantic resources (wn, wnd, wna, sumo, sc and new semantic relations provided by xwn and a sentiment analysis resource (swn), using wn, from versions . and . ). the knowledge database obtained in the integration process was evaluated in order to detect any problems in the alignment by using different intermediate versions of wn. so, depending on the nucleus of wn, wnd and sumo used, an accuracy of % was obtained, while accuracies of . % and % were obtained for sc related to wn . and . , respectively. in addition to this research we have included a set of research works that contains interesting knowledge bases used to deal with nlp tasks. one of the most interesting resources is swn. it supplies the positivity of a sense, domain, category, emotion or semantic class. it is important to stress that this resource has been used in several research works to take advantage of its semantic multidimensionality. based on the semantic multidimensionality that isr-wn provides we are able to extract new information to classify, obtain opinions or recommend similar texts in open domains. basically, we have applied four different tasks to add new information from different point of views: extract similarities among texts, measure sentiment polarities and obtain a semantic analysis for understanding contexts meanings. moreover, we have illustrated a case of study with information related to movies and tv series reviews to demonstrate that our framework works properly. as future work we plan to enrich the isr- wn resource with collocation sense relations . this type of information provides better results in tasks such as wsd (gutiérrez, ). in (gutiérrez et al., c) isr-wn was used to combine conceptualizations with polarities. we propose adding resources such as micro-wnop (a corpus labelled with sentiments with around , wn synsets). moreover, we are working to align isr-wn with wns in other languages in order to create a multilingual resource. synset pairs that commonly appear together in corpus. http://www.unipv.it/micrownop love love sensation n sensation medicine medicine accepted manuscript a cc ep te d m a n u sc ri pt in order to add more knowledge to isr-wn, we propose the integration of different ontologies by using the rdf model to align wn synsets with ontological concepts. this kind of information will help us to apply different techniques (i.e. domain classification) with specific domains such as: medicine, technology, tourism, pharmaceutical, etc. moreover, we want to integrate babelnet (navigli and ponzetto, ) (a resource that connects the multilingual web encyclopaedia wikipedia with wn) with isr-wn. babelnet was considered as a sense inventory in semeval in task : multilingual word sense disambiguation. finally, we plan to use the isr-wn functionalities in order to help summarization systems to customize text summaries using different domains, categories, emotions, sentiment polarities, etc. acknowledgments this research work has been partially funded by the university of alicante, generalitat valenciana, spanish government and the european commission through the projects, tin - -c - -r, tin - -r, sam (fp - ), and prometeoii/ / . references agirre, e., cer, d., diab, m. & gonzalez-agirre, a. ( ) semeval- task : a pilot on semantic textual similarity. {*sem }: the first joint conference on lexical and computational semantics -- volume : proceedings of the main conference and the shared task, and volume : proceedings of the sixth international workshop on semantic evaluation {(semeval )}. montreal, canada, association for computational linguistics. agirre, e., lacalle, o. l. d., fellbaum, c., hsieh, s.-k., tesconi, m., monachini, m., vossen, p. & segers, r. ( ) semeval- task : all-words word sense disambiguation on a specific domain. proceedings of the th international workshop on semantic evaluation. los angeles, california, association for computational linguistics. agirre, e. & soroa, a. ( ) personalizing pagerank for word sense disambiguation. proceedings of the th conference of the european chapter of the association for computational linguistics (eacl- ). athens, greece. atserias, j., villarejo, l., rigau, g., agirre, e., carroll, j., magnini, b. & vossen, p. ( ) the meaning multilingual central repository. proceedings of the second international global wordnet conference (gwc’ ). . brno, czech republic. baccianella, s., esuli, a. & sebastiani, f. ( ) sentiwordnet . : an enhanced lexical resource for sentiment analysis and opinion mining. in , l. (ed.) th language resources and evaluation conference. valletta, malta. bentivogli, l., forner, p., magnini, b. & pianta, e. ( ) revising the wordnet domains hierarchy: semantics, coverage and balancing. proceedings of coling workshop on "multilingual linguistic resources". geneva, switzerland. . briguez, c. e., budan, m. c. d., deagustini, c. a. d., maguitman, a. g., capobianco, m. & simari, g. r. ( ) argument-based mixed recommenders and their application to movie suggestion. expert systems with applications, , - . brill, e. ( ) transformation-based error-driven learning and natural language processing: a case study in part-of- speech tagging. mit press. cotton, s., edmonds, p., kilgarriff, a. & palmer, m. ( ) english all word. in linguistics, a. f. c. (ed.) senseval- : second international workshop on evaluating word sense disambiguation systems. toulouse, france, association for computational linguistics. chávez, a., dávila, h., gutiérrez, y., collazo, a., abreu, j. i., fernández orquín, a., montoyo, a. & muñoz, r. ( ) umcc_dlsi: textual similarity based on lexical-semantic features. second joint conference on lexical and computational semantics (*sem), volume : proceedings of the main conference and the shared task: semantic textual similarity. atlanta, georgia, usa, association for computational linguistics. dávila, h., fernández, a., gutiérrez, y., muñoz, r., montoyo, a. & vázquez, s. ( ) semantic information extraction method on ontologies. sepln : xxviii congreso de la sociedad espaÑola para el procesamiento del lenguaje natural. castellón, spain. dávila, h., fernández orquín, a., chávez, a., gutiérrez, y., collazo, a., abreu, j. i., montoyo, a. & muñoz, r. ( ) umcc_dlsi-(eps): paraphrases detection based on semantic distance. second joint conference on lexical and computational semantics (*sem), volume : proceedings of the seventh international http://www.cs.york.ac.uk/semeval- / accepted manuscript a cc ep te d m a n u sc ri pt workshop on semantic evaluation (semeval ). atlanta, georgia, usa, association for computational linguistics. dorr, b. j. & castellón, m. a. m. a. i. ( ) spanish eurowordnet and lcs-based interlingual mt. amta/sig-il first workshop on interlinguas. san diego, ca. esuli, a. & sebastiani, f. ( ) sentiwordnet: a publicly available lexical resource for opinion mining. in , l. (ed.) fifth international conference on languaje resources and evaluation genoa - italy. fellbaum, c. ( ) wordnet. an electronic lexical database, university of cambridge. fernández, a., gutiérrez, y., dávila, h., chávez, a., gonzález, a., estrada, r., castañeda, y., vázquez, s., montoyo, a. & muñoz, r. ( a) umcc_dlsi: multidimensional lexical-semantic textual similarity. {*sem }: the first joint conference on lexical and computational semantics -- volume : proceedings of the main conference and the shared task, and volume : proceedings of the sixth international workshop on semantic evaluation {(semeval )}. montreal, canada, association for computational linguistics. fernández, a., gutiérrez, y., muñoz, r. & montoyo, a. ( b) approaching textual entailment with sentiment polarity. icai' - the international conference on artificial intelligence. las vegas, nevada, usa. forner, p. ( ) wordnet domains . . itc-irst, povo-trento, italy. gediminas, a. & alexander, t. ( ) toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions. ieee educational activities department. genesereth, m. r. & fikes, r. e. ( ) knowledge interchange format. in university, c. s. d. s. (ed.) version . reference manual ed. stanford, computer science department. gliozzo, a., strapparava, c. & dagan, i. ( ) unsupervised and supervised exploitation of semantic domains in lexical disambiguation. computer speech and language. gutiérrez, y. ( ) análisis semántico multidimensional aplicado a la desambiguación del lenguaje natural. departamento de lenguajes y sistemas informáticos. alicante, alicante. gutiérrez, y., castañeda, y., gonzález, a., estrada, r., piug, d. d., abreu, j. i., pérez, r., fernández orquín, a., montoyo, a., muñoz, r. & camara, f. ( ) umcc_dlsi: reinforcing a ranking algorithm with sense frequencies and multidimensional semantic resources to solve multilingual word sense disambiguation. second joint conference on lexical and computational semantics (*sem), volume : proceedings of the seventh international workshop on semantic evaluation (semeval ). atlanta, georgia, usa, association for computational linguistics. gutiérrez, y., fernández, a., montoyo, a. & vázquez, s. ( a) integration of semantic resources based on wordnet. in , s. (ed.) xxvi congreso de la sociedad española para el procesamiento del lenguaje natural. universidad politécnica de valencia, valencia, sepln . gutiérrez, y., fernández, a., montoyo, a. & vázquez, s. ( b) umcc-dlsi: integrative resource for disambiguation task. proceedings of the th international workshop on semantic evaluation. uppsala, sweden, association for computational linguistics. gutiérrez, y., fernández, a., montoyo, a. & vázquez, s. ( a) enriching the integration of semantic resources based on wordnet. procesamiento del lenguaje natural, , - . gutiérrez, y., vázquez, s. & montoyo, a. ( b) improving wsd using isr-wn with relevant semantic trees and semcor senses frequency. proceedings of the international conference recent advances in natural language processing . hissar, bulgaria, ranlp organising committee. gutiérrez, y., vázquez, s. & montoyo, a. ( c) sentiment classification using semantic features extracted from wordnet-based resources. proceedings of the nd workshop on computational approaches to subjectivity and sentiment analysis (wassa . ). portland, oregon., association for computational linguistics. gutiérrez, y., vázquez, s. & montoyo, a. ( d) word sense disambiguation: a graph-based approach using n- cliques partitioning technique. in muÑoz, r., montoyo, a. & mÉtais, e. (eds.) natural language processing and information systems. springer berlin / heidelberg. izquierdo, r. ( ) una aproximación a la desambiguación del sentido de las palabras basada en clases semánticas y aprendizaje automático. departamento de lenguajes y sistemas informáticos. alicante, universidad de alicante. izquierdo, r., suárez, a. & rigau, g. ( ) a proposal of automatic selection of coarse-grained semantic classes for wsd. procesamiento del lenguaje natural, , - . izquierdo, r., suárez, a. & rigau, g. ( ) gplsi-ixa: using semantic classes to acquire monosemous training examples from domain texts proceedings of the th international workshop on semantic evaluation. uppsala, sweden, association for computational linguistics. leacock, c. & chodorow, m. ( ) using corpus statistics and wordnet relations for sense identification. computational linguistics. luisa bentivogli, p. f., bernardo magnini, emanuele pianta ( ) revising the wordnet domains hierarchy: semantics, coverage and balancing. itc-irst – istituto per la ricerca scientifica e tecnologica via sommarive , povo – trento, italy, . magnini, b. & cavaglia, g. ( ) integrating subject field codes into wordnet. proceedings of third international conference on language resources and evaluation (lrec- ). magnini, b., satrapparava, c., pezzulo, g. & gliozzo, a. (july ) the role of domains informations in word sense disambiguatios. treto, cambridge university press. magnini, b., strapparava, c., pezzulo, g. & gliozzo, a. ( ) comparing ontology-based and corpus-based domain annotations in wordnet. proceedings of the first international wordnet conference. mysore, india. accepted manuscript a cc ep te d m a n u sc ri pt marco de, g., pasquale, l., giovanni, s. & pierpaolo, b. ( ) integrating tags in a semantic content-based recommender. proceedings of the acm conference on recommender systems. lausanne, switzerland, acm. miller, g. a., beckwith, r., fellbaum, c., gross, d. & miller, k. ( ) five papers on wordnet. princenton university, cognositive science laboratory. navigli, r. ( ) word sense disambiguation: a survey. acm comput. surv., , : -- : . navigli, r. & ponzetto, s. p. ( ) babelnet: building a very large multilingual semantic network. proceedings of the th annual meeting of the association for computational linguistics. uppsala, sweden, association for computational linguistics. niles, i. ( ) mapping wordnet to the sumo ontology. teknowledge corporation. niles, i. & pease, a. ( ) origins of the ieee standard upper ontology. working notes of the ijcai- workshop on the ieee standard upper ontology. seattle, washington, usa. niles, i. & pease, a. ( ) linking lexicons and ontologies: mapping wordnet to the suggested upper merged ontology. pease, a. ( ) standard upper ontology knowledge interchange format. perner, p., candillier, l., meyer, f. & boulle, m. ( ) comparing state-of-the-art collaborative filtering systems. machine learning and data mining in pattern recognition. springer berlin heidelberg. peter, d. t. ( ) mining the web for synonyms: pmi-ir versus lsa on toefl. proceedings of the th european conference on machine learning. springer-verlag. pianta, e., bentivogli, l. & girardi, c. ( ) multiwordnet. developing an aligned multilingual database. proceedings of the st international wordnet conference. mysore, india. rao, d., mcnamee, p. & dredze, m. ( ) entity linking: finding extracted entities in a knowledge base. multi- source, multilingual information extraction and summarization. springer. russell, s. & norvig, p. ( ) a modern, agent-oriented approach to introductory artificial intelligence. sanda m. harabagiu, g. a. m., dan i. moldovan ( ) wordnet -a morphologically and semantically enhanced resource. siglex : standardizing lexical resources. sara, t. & daniele, p. ( ) new features for framenet: wordnet mapping. proceedings of the thirteenth conference on computational natural language learning. boulder, colorado, association for computational linguistics. Ševčenko, m. ( ) online presentation of an upper ontology. ctu prague, dept of computer science. shinde, s. k. & kulkarni, u. ( ) hybrid personalized recommender system using centering-bunching based clustering algorithm. expert systems with applications, , - . sowa, j. f. ( ) knowledge representation: logical, philosophical, and computational foundations. course technology. strapparava, c. & valitutti, a. ( ) wordnet-affect: an affective extension of wordnet. proceedings of the th international conference on language resources and evaluation (lrec ). lisbon. turney, p. d. & littman, m. l. ( ) measuring praise and criticism: inference of semantic orientation from association. acm transactions on information systems (tois), , – . udi, m., ash, p. & john, r. ( ) experience with personalization of yahoo! , acm. valitutti, a., strapparava, c. & stock, o. (eds.) ( ) developing affective lexical resources, itc-irst, trento, italy, psychnology journal. vázquez, s. ( ) resolución de la ambigüedad semántica mediante métodos basados en conocimiento y su aportación a tareas de pln. depto. de lenguajes y sistemas informáticos. alicante, spain., universidad de alicante. vázquez, s., montoyo, a. & rigau, g. ( ) using relevant domains resource for word sense disambiguation. ic-ai’ . proceedings of the international conference on artificial intelligence. ed: csrea press. las vegas, e.e.u.u. vossen, p. ( ) eurowordnet: a multilingual database with lexical semantic networks, dordrecht, kluwer academic publishers. vossen, p., peters, w. & gonzalo, j. ( ) towards a universal index of meaning. proceedings of the acl- siglex workshop. university of maryland. walter, c.-n., maria luisa, h.-a., rafael, v.-g. & francisco, g.-s. ( ) social knowledge-based recommender system. application to the movies domain. pergamon press, inc. zhibiao, w. & martha, p. ( ) verbs semantics and lexical selection. proceedings of the nd annual meeting on association for computational linguistics. las cruces, new mexico, association for computational linguistics. zouaq, a., gagnon, m. & ozell, b. ( ) a sumo-based semantic analysis for knowledge extraction. proceedings of the th language & technology conference. poznań, poland. pii: - ( ) -x international journal of production economics, - ( ) - elsevier an expert system for a local planning environment g.j. meester laboratory of production management and logistics, department of mechanical engineering, university of twente, p.o. box , ae enschede. the netherlands abstract in this paper, we discuss the design of an expert system (es) that supports decision making in a local plannmg system (lps) environment. the lps provides the hnk between a high level factory planning system (rough cut capacity planning and material coordination) and the actual execution of jobs on the shopfloor. by specifymg a detailed workplan. it is divided m two hterarchtcal layers: planmng and scheduling. at each level, a set of different algorithms and heuristics is available to anticipate different situations. the expert system (which is a part of the lps) supports decision making at each of the two lps layers by evaluatmg the planning and scheduling conditions and, based on this evaluation, advising the use of a specific algorithm, and evaluatmg the results of using the proposed algorithm. the expert system is rule-based while knowledge (structure) and data are separated (which makes the es more flexible m terms of fine-tuning and adding new knowledge). knowledge is furthermore separated in algorithmic knowledge and company specific knowledge. in this paper we discuss backgrounds of the expert system in more detail. an evaluation of the expert system is also presented . introduction production control has become increasingly complex during the last few decades. on the one hand market requirements tend towards more diversity, a higher quality and reliable due dates. on the other hand the rapid ad- vancements in information technology have increased the possibilities on the shopfloor, thereby leading to more complex planning and control problems. in this paper we concentrate on discrete parts manufacturing environments. it is com- monly accepted that the complexity of the overall planning problems should lead to a hierarchical planning and control framework, where demand management and long term capacity planning are high level decisions, correspondence to: g.j. meester, laboratory of produc- tion management and logistics, department of mech- anical engineering, university of twente, p.o. box , ae enschede, the netherlands. materials coordination is at the medium level, while detailed shopfloor planning and control can be found at the lower levels. for the high and medium level decisions we assume that a global planning system (gps) has been implemented. in practice we often may find an mrp ii system (manufacturing resource planning system) at this place. however, it is well known that global planning systems such as e.g. an mrp ii sys- tem are not adequate in translating a planned set of orders into detailed work instructions at the shopfloor level. to this end, a local plann- ing system (lps) has recently been developed at the university of twente (cf. refs [l- ]). hence, this lps acts as shopfloor planning and control system and as such bridges the gap between the global planning system and the actual execution of jobs on the shopfloor. to highlight the functioning of the lps in some more detail, we assume (reflecting common industrial practice) that the factory shopfloor can be decomposed into relatively - / /$ . elsevier science publishers b.v. all rights reserved. core metadata, citation and similar papers at core.ac.uk provided by universiteit twente repository https://core.ac.uk/display/ ?utm_source=pdf&utm_medium=banner&utm_campaign=pdf-decoration-v g. j. meesterilocal planning tmmnn~e~~t autonomous groups of machines (this may be a functional arrangement but also a cell in a cellular manufacturing environment). now, the global planning system organizes the materials flow between these machine groups (to be called factory systems (fs) in the se- quence). subsequently, the local planning sys- tem attempts to match this flow with the critical resources within each fs, such as tools, fixtures, machine capacities and operators. hence, the fs is the local planning environment for the lps. in performing these tasks, a large number of decisions have to be made while facing several, sometimes conflicting, objectives. to support this decision making an expert system may be helpful, in particular to evaluate specific work- ing conditions, to suggest planning and sched- uling algorithms and to judge the final results. the objective of this paper is to describe what such an expert system should look like, and to demonstrate that it can support an operator to substantially improve the performance of the lps. in other words: it is demonstrated how we can formalize knowledge about planning and scheduling and how we can use this knowl- edge operationally. other research on expert systems in the type of production situations discussed here can be found in refs. [ - . for general discussions on expert systems the reader is referred to refs. [s- ]. in the remainder of this article we will focus on the expert system. first the expert system concept in general will be outlined and some properties of the concept of knowledge are discussed. next the architecture, its tasks and functions are described for the particular lps situation. the lps-expert system combi- nation has been evaluated in four different situ- ations. the evaluation of the expert system is presented in section . conclusions and sugges- tions for further research are given in section . . the expert system: principles, tasks and knowledge the basic idea behind an expert system for an industrial company is that the company should save its investments in knowledge, even if skilled employees should leave the company. expert systems can be used when we have to deal with problems in such fields as [ ] : ( ) interpretation and understanding of a complex data structure, ( ) ordering/scheduling, ( ) the evaluation of situations, ( ) diagnostics, ( ) the tracing of failures. there are different types of expert systems which can be used to solve these problems [ ] such as: ( ) rule based expert systems, ( ) expert systems based on first order predi- cate logic, ( ) expert systems based on structured objects (e.g. frames and trees). looking at the characteristics of the tasks that an expert system should perform in the particular environment, studied in this paper, we have selected a rule based system. rules can represent a unit piece of knowledge. advan- tages of a rule based es are: ~ it is easy to work with in test situations, _ it is easy to implement in a control concept (as will be explained in the next sections). a more elaborate discussion of arguments leading to the choice of a rule based es is given in ref. [ ]. the structure of such an expert system is described in detail in ref. [ ]. knowledge is represented in what we call “knowledge rules” (if-then-else). these rules are mutually independent. one of the properties of an expert system in general is that the knowledge (structure) and the data are strictly separated. the advantage of an expert system in contrast with a tradi- tional computer program is that if something changes in the problem environment, it will not cause much problems in the es. one should simply add a new rule or new data. in a traditional program, the entire source code may have to be changed. the structure of an expert system will be briefly explained now. an expert system is con- structed from three basic elements: ) the database, ) the knowledge base, ) the infer- ence engine. in our research framework we g.j. meesterjlocal planning encironment have added some slightly more advanced con- cepts like a statistics module (and in the future a simulation module). see fig. . the inference engine matches patterns from the “rule base” or knowledge base, with recom- mended data from the database. the combina- tion represents the knowledge. in fig. we have also included the module “statistics”. by changing data in the database, it implicitly changes the knowledge. to clarify what is meant by knowledge, how it can be used opti- mally and how it should be represented in the rule based expert system, some techniques are illustrated by using an example. example: suppose the “expert” knows that computer time will increase exponentially fast when the algorithm “complete enumeration” (this algorithm calculates every possible sched- ule and gives always the optimal solution) will be used by a set of jobs larger than . we can implement this knowledge as follows: if ord# ~ (condition] then advise algorithm = complete enumeration (action}, certainty factor = {certainty}. i : simulation it__. statistics :______: i i_____________, fig. . expert system, structure. ii knowledgebase !+ if a > b then action. i one can distinguish three main parts in the knowledge rules, which are explained in short: ( ) ( ) ( ) the condition part: the condition part can be true or false. if it is true, action is taken, if it is false, no action will be taken. the condition part may find the values of the corresponding parameters (constants, variables, etc.) in the database or from other rules in the rule base (tree search). for instance in fig. action will be taken if a > b. the value of the parameter a is found in the database. however, b is not and therefore the inference engine now searches in the knowledge base and finds b in the second rule. since now d is needed, the inference engine will search for it. in this way a tree of rules can be checked. the action part: this part of the rule de- scribes the action that has to be taken. an action can be: _ changing data in the database (for con- trol applications), - placing at inferences disposal of data for tree search activities, - proposing an algorithm. the certainty part: this part specifies the value of the so-called certaintyfactor. basi- cally the value of the certainty factor indi- cates the probability that a recommended action will lead to satisfactory results. we will come back to the construction and role of the certainty factor in more detail below. under complex environmental conditions it is not always obvious what may be expected from the performance of a given rule or algo- rithm (e.g. priority dispatching rules). statis- tical knowledge about the performance of such a rule in the past, under similar conditions, may therefore be helpful to predict its behavior if d then b = , if c <> then d = fig. . example knowledge representation. g.j. meester/local planmng encironrnent in the future. in this way a sort of learning capability is included in the expert system. this idea has motivated the introduction of the certaintyfactor. basically its value specifies the probability that an algorithm performs sat- isfactorily. since any knowledge about recent performances should be included, a reasonable up-date procedure for the value of the cer- tainty factor is as follows: cf,,, = (cfold*n + x ) n+l . ( ) here cfold and cf,,, represent the initial and the updated values of the certainty factor, respectively. furthermore, we set = if the solution determined by the relevant algorithm in the last cycle is satisfactory, and = other- wise. n is the number of previously recorded tests upon which the certainty factor has been based so far. for practical reasons the value of n should neither be too low nor too high. when n is very low the certainty factor will be strongly influenced by each additional test. when n is high the influence becomes negli- gible. when the rule is tested often, n is high and the value will become very stable. as cir- cumstances may change, considering all for- mer tests has its disadvantages. therefore, both a lower bound and an upper bound are given to (in some cases even a constant value will do). having thus motivated the use of the cer- tainty factor (basically to incorporate learning effects) we next explain its use. given a particu- lar planning or scheduling problem faced by the lps, we first check the environmental con- ditions and objectives to find what algorithms or rules might be applicable. next, we simply choose the rule with the highest cf-value, where ties are broken arbitrarily. another way to incorporate knowledge about the performance of certain rules or algo- rithms is by adjusting the conditions under which it may be applied. these conditions, such as the condition in our example, may be changed as a result of recent performance evaluations. e.g. in our example let us replace the factor by a factor wf. the condition hence becomes ord d wf. in this case, as a re- sult of statistical analysis one may adapt the value of the factor wf and write this new value in the database. complex statistical pro- cedures may be required to find good values for such factors. an expert system will only perform well if the knowledge added to the data and knowl- edge base is well defined. this means: the pro- cess of knowledge acquisition must be carried out thoroughly and accurate. a formal de- scription of the knowledge acquisition process is given in ref. [ ]. for our particular situ- ation we list a few important steps. note that knowledge acquisition basically concerns the representation of knowledge rules (fig. ). ( ) first the relevant relations in combination with the relevant indicators have to be traced. to get these relations one may use a knowledge relation chart (see fig. ). in such a chart the relations between the lps algorithms and the potential environment indicators are given. after an interview with an expert or analyses done on the lps algorithms, the relations between the indicators and the lps-algorithms found can be added into the chart. ( ) after finding the important relations, the effective area of the indicators must be found (i.e. in our example the effective area was the interval [ , . values of or more are not relevant). possibilities here are to test the lps algorithms for different input values. the condition part can be defined in combination with the action part (lps-algorithm). fig. . knowledge acquisition chart. g.j. meesterllocal planning environment ( ) finally a certainty factor must be set. when there is no information about the performance of the lps-algorithm an ar- bitrary value must be added. in the long run the certainty factor converges to a stable value (in a stable environment). with this information the condition part, action part and certainty part can be projected on the knowledge representation structure (if-then-else structure) which yields the knowledge rules that can be implemented. . an expert system for the lps a local planning system contains several modules, structured in a hierarchical frame- work. the two most important ones are the “planning module” and the “scheduling module”. the planning module performs the follow- ing functions: _ it receives a set of orders (see fig. ) from the gps that have to be produced in a certain time period of one to three weeks, say. - it checks capacity and handles utilization planning on a high aggregation level. this fig. . lps. means in particular: _ moving orders in time, _ selecting alternative process plans, avail- able in “technical data” (cf. fig. ). _ it selects a subset of orders for the next day from the above described order set. this sub- set forms the input for the scheduling module. the scheduling module performs the follow- ing functions: _ it creates the schedule for the next day by assigning all operations from the selected subset of orders to machines and by sequencing all operations assigned to that specific machine. - it creates task lists for the resources (e.g. tools, machines, fixtures, operators, etc.). both levels contain a range of different algo- rithms and heuristics. for most particular situations there is an algorithm or heuristic available that performs best under these circumstances. this means that an lps oper- ator (planner) has to interpret the situation (given by the gps and the actual situation on the shopfloor) and has to select an algorithm that yields a satisfactory planning/scheduling solution. the algorithms and heuristics available on “planning” level are: ( ) for utilization planning: _ moving orders in time: _ sorting algorithms to select jobs on specific properties, _ clustering algorithms on technical constraints (for tools and fixtures), _ clustering algorithms with precedence relations, release and due dates. - process plan selection: - complete enumeration (workload bal- ancing), _ mixed integer programming (work- load balancing), - linear programming, continuous rounded (workload balancing), _ two-opt heuristic (pair-wise inter- change) (workload balancing), _ sorting on resource load (workload balancing), ( ) for workloading: - linear programming. the module scheduling contains algorithms and heuristics to generate schedules which focus on the utilization of resources, short term due-date performance, throughput times, and schedule robustness. ( ) before using actual dispatching and sched- uling algorithms, it is possible to use a preparation tool to let the dispatching and scheduling algorithms perform better. for example: _ add all operator relations, _ add relations for a specific fixture type, _ add all machine-tool relations for a spe- cific machine. ( ) the dispatching may be used to accom- modate e.g.: _ order due dates, _ operations due dates, - priorities. the actual scheduling procedure is the shifting bottleneck procedure developed by adams et. al. [ ]. ( ) after having generated the schedule, there is a possibility to improve the schedule somewhat further by one of the secondary scheduling procedures: _ make feasible, _ adapt for gaps. the heuristics and algorithms are described more extensively in ref. [ ]. the expert system should support the deci- sion making in selecting algorithms on differ- ent levels. it may either support the operator or operate as an autonomous program with- out intervention of the operator. let us return to the specific expert system structure de- veloped for the lps environment. the lps basically consists of two layers: a planning level and a scheduling level. an important rea- son to split the system in such a hierarchical structure is to be able to deal with the com- plexity of the problems encountered. also the es developed here is used in an experimental environment, with a data and knowledge base on each separate level it seems easier to test. having just one level where all decisions should be made it may become harder to evalu- ate the effects of any new rule or alteration. as mentioned earlier, the expert system has to perform the following tasks. on both, the planning and the scheduling level it has to: _ evaluate the situation in the planning and scheduling environment, _ support the operator by choosing the right algorithm or heuristic in a specific situation (generate), _ evaluate the result of one algorithm or heu- ristic (are we satisfied or not?) (evaluate). the generate-evaluation concept (depicted in fig. ) is implemented on both lps levels. on each level the expert system has to advise. all relevant data according to this advice and to the result of this advice is temporarily saved (as best). the evaluation part checks this data. if “evaluation” is satisfied, the expert system permits to move to a lower level of the lps. if “evaluation” is not satisfied, the expert system has to generate a new solution by sug- gesting a new algorithm. remark i. when “evaluation” is not satis- fied, it will explain to the advice-part what’s wrong. so the advice part is able to re-advise “intelligently”. we call this control loop the short control loop. remark . when “evaluation” continues to return “not satisfied” it may advise to go to ____________ evalua :ion part advise part j w o on with ‘best’ to next stag & fig. . control mechanism. g.j. meesterllocal planning environment a lower lps-level after a certain number of loops. this is to protect against infinite loops. (temporary solution “best” is given to the lower level). remark . the evaluation system on the lps-scheduling level has the possibility to send information to the lps-planning level. since we deal with a hierarchic system, it may possibly happen that the planning module presents an order subset to the scheduling module which does not allow to meet a specific criterion under any schedule. both advice and evaluation part have to be fitted into the expert system (see section ). . implementation the expert system is implemented on a hp workstation. it operates together with the lps interface - , i lps under an x-window representation man- ager on a unix operating system. both are written in standard pascal, see ref. [ ]. the choice for pascal is motivated by the fact that in this type of research one has to deal with many new concepts and changes. in pascal it is easy to develop new ideas. we also have investigated different software packages, like the ibm expert shell [ ] and the delfi ii expert environment. both of the packages are professional expert environments but the link between these expert systems and the lps is difficult to make. also extensions on the ex- pert system framework the hardly to realize. once having discussed the control mechan- ism and the knowledge structure we are now ready to present the complete expert system structure. the structure is depicted in fig. . observe the knowledge rules embedded in this concept. expert system i gps i i i if_tkn-els.. ifjknelse.. lzzza= tnference engine fig. . control chart. g.j. meester/lod plunnmg ewirorwnent not mentioned till now but important for the es-concept is the interface unit (see fig. ). the interfaces are mechanisms which are able to compress files of information in a special way. in particular, a file given by the gps can be transformed into a vector of indicators. in that way it is easier for expert system to interpret the gps situation. for a more elab- orate treatment of the interfaces we refer to ref. [ ]. to further highlight fig. one, step of the advisory and evaluation process of the expert system will be described. step i. the gps file is translated into an indicator vector. the mass of available data will be compressed to a set of indicators which represent the important characteristics of the file. the vector is stored in the database. step . after a request for advice is received, the inference engine matches the data and the rules and presents the advice. it also reports to the module statistics. ( ) ( ) ( ) ( ) step . the result of the advice is imple- mented. step . the planning file (containing the results of following the advice) is transformed in a way similar to the one described in step . the combination tested in an operation oriented fs situation. this case concerns an fs which contains rather similar milling/drilling machines (the only differ- ence between the machines is the number of driven rotation and translation axes). the second case concerns a part oriented fs consisting of machines of three differ- ent types. we may think about families of products visiting this machining centre (part flow). for details see below. the third case is taken from ref. [is]. actually this fms situation concerns a pilot plant in a dutch research institute (ipl-tno, apeldoorn). it consists of a turning centre and a fs. the machines and a tool preparation cell are integrated in a cim environment (the tn super- visory control system). the final case is a comparison made be- tween opis/isis, a scheduling system de- veloped by fox et al. [ ], and the lps-es. the test situation used in this case is de- scribed in ref. [ ]. special in this case is that the fs consists of a number of cells, where each cell contains a number of sim- ilar machines. step . there are two possibilities. the re- sult of the evaluation is: “ok”: -+ we continue with scheduling; go to step , “not ok”: --t a new advice is given. the differ- ence is that the evaluation changes the data in the database on the advice level (see fig. ). go to step . in our expert system we have implemented the knowledge evaluation by means of the cer- tainty factor. the module “statistics” takes care of the handling of the factor n. for every rule which generates an advice a value of n is maintained. how the lps-es combination works in a particular environment (recall that the es implementation is situation-specific) is ex- plained in more detail, although the lps-es combination is extensively tested for each of the four cases described above, at this point, for the sake of brevity only the knowledge acquisition and the evaluation of the second case are described. . . outline of the second case . evaluation of the expert system . . . the factory system the fs consists of machines: two of type a, two of type b and one of type c. each machine has a tool magazine with a capacity of forty tools. to evaluate the power and usefulness of the each job follows a predetermined route lps-es (local planning system-expert sys- along the machine types: a + b + c. the tem) combination, it is implemented in four maximum number of different operations different environments. within one job is three on machine a and b g.j. meesterllocal planning environment and two on machine c. the total number of operations within one job is at most five. two shifts ( hours) per day are assumed. . . . jobs jobs have to be produced in one week. every job has to be produced in to days, with a mean of . jobs also can have a priority index ranging from to where means a rush job. every job consists of one or more products of one available type. for every product type there are two process plans avail- able which can be selected for production. other job characteristics are: the batch size: to with a mean of , the number of operations: to with a mean of , the process time: to minutes with a mean of , special tools per operation: to with a mean of . a file generator has been developed to gen- erate a gps file with the preferred factory and job characteristics. . . knowledge acquisition for the second case rithm, based on the situation given by the gps and depending on the status of the fs. for this information is required (cf. kl . . . kn in fig. ). this information must be compressed into a limited set of indicators (see also fig. ). based on the results of step and given the information stored in the gps data file, we have chosen the following set of indicators for the selection of planning algorithms: ( ) the total workload per machine, ( ) total workload released/total workload, ( ) the number of different parts, ( ) the average number of alternative process plans, at this point the knowledge acquisition pro- ( ) the total number of operations, cess (in analogy of section ) for the expert ( ) the average order priority, system for the second case is described. we ( ) the average release date, confine with the description of the acquisition ( ) the average due date, process for the lps-“planning level”. ( ) the number of orders. . . . step i: dejinition of useful algorithms, heuristics and solution rules first step in the knowledge acquisition pro- cess is to select a subset of useful algorithms (cf. pl . . . pn in fig. ) from the nine algorithms available on the planning level. in this case the objective on the planning level is to balance the workload among the machines. according to this objective we have, based on the knowledge and experience of the lps-builder, selected the following algo- rithms: for the process plan selection: - full enumeration, _ a heuristic on workload balancing, _ a mixed integer programming formulation, _ a heuristic for sorting on the workload of one specific machine. for work loading: - linear programming with a utilization of %, - a heuristic with a tool capacity bound and a utilization of %. . . . step : dejinition of indicators for the advice part the expert system has to advise an algo- . . . step : definition of indicators for the evaluation part the evaluation is an important part of the expert system framework. when an algo- rithm is selected and executed, the results may be judged to be poor or even not accept- able by the evaluation part of the expert sys- tem, according to some specified criteria. these criteria may be simple operational cri- teria such as the average or maximum lateness but may also concern more “fuzzy” conditions such as the capabilities of operators to deal with the resulting solution. to this end some control indicators are needed. we have g.j. merster/local platmng twwontntwt selected: ( ) ( ) ( ) c a parameter indicator whether or not the result is almost satisfactory, what exactly is meant by “satisfactory” is discussed below. the number of attempts so far. the algorithms evaluated earlier. these parameters are written into the database. they provide sufficient information to yield a good advice. generally, the pro- cess to find the proper parameters for practical situations will be iterative. _ j. . step : dejinition sf the knowledge rules now that we have defined a set of important indicators in combination with a set of useful algorithms in the previous steps, we may con- struct the knowledge rules by using a chart as depicted in fig. . after the determination of the indicator intervals and the certainty factors our knowledge acquisition process is com- pleted. we implemented two types of knowledge. we can distinguish: ( ) ( ) theoretical- knowledge: this is knowledge about the actual lps-algorithms. this means that the relation between the input data characteristics and the output perfor- mance of each algorithm and its properties in different situations are stored in knowl- edge rules. this knowledge is generally usable, situation specific knowledge which may concern specific company knowledge or fs-configuration constraints (e.g.: tool x is not available for job of type y). this knowl- edge is specific for each of the four cases and therefore different in each of the cases. . . evaluation of the espert sqxtern in the second case . . . hon. to evaluate an expert system the expert system is implemented to sup- port the operator in decision making among the different algorithms, rules and heuristics available on the different levels in the lps. so, what we are interested in, is what the quality is of the advice to select a specific algorithm, as given by the expert system. did the algorithm, suggested by the expert system, perform well? before we can give judgement about a satisfac- tory performance of an advised algorithm, we first must know what exactly is meant by satis- factory. in other words: how can one measure the quality of an advice. . . . the qualit), of an advice to establish the quality of an advice one has to evaluate the actual output of the lps, which is actually a complete production sched- ule. a schedule can be evaluated by perfor- mance indicators like maximum lateness, tardiness, utilization of machines or through- put times. these performance indicators can be compared with lower bounds. lower bounds can be found by using e.g the shifting bottleneck procedure (cf. ref. [ ] ) for the lateness criterium and a mathematical programming formulation for the utilization criterium. deviation of the performance indi- cators from the lower bound gives a measure for the quality of a production schedule. in this research, the quality of the produc- tion schedules is measured for every advice generated by the expert system. the reason for not presenting the data to match, is that the data gives more information about the quality of the used algorithms themselves instead of given straight the information about the qual- ity of the decision of choosing the algorithm. so, in our evaluation we have used the follow- ing performance measure: did the advised al- gorithm lead to a solution (i.e. a workload balancing or a production schedule) which could not be improved by applying any of the other available algorithms according to the performance measures described above (in particular the lateness criterium). . . . evaluation results of’ the second case the lps expert system is tested on dif- ferent gps order sets. g.j. meester,ilocul planthg etwironment in cases the expert system advised these algorithms which lead the lps to an “opti- mal” production schedule at once. in cases the expert system used its evalu- ation optimization mechanism (cf. figs. and ) on planning level only ( times) on schedul- ing level only ( times) and on both levels at the same time ( times) to find the optimal combination of algorithms leading to the “op- timal” production schedule. in cases the expert system advised to re- turn to the lps-planning level for a complete new lps session, after recognizing infeasibility at the lps-scheduling level. these cases show that the control mechanism as depicted in figs. and , works well. in one case the expert system did not advise the algorithm that finally led to the “optimal” production schedule. . . . remarks ( ) ( ) ( ) the results of these tests reflects the quality and the completeness of the know- ledge gathered during the acquisition phase. although only the second case is outlined here, equivalent results have been found in the other three test environments. the fact that most of the times an optimal production schedule has been found indi- cates that the situation presented by the gps is interpreted well by the expert system. . . . conclusions extracted from the test cases include ( ) ( ) in all four test cases, the lps performs well on the advices given by the es. we only once could find a better production sched- ule by using a different algorithm then the one advised by the expert system in the second case. in the various, often complex, situations and circumstances the expert system appears to be a useful tool in supporting decision making to control the lps. ( ) the theoretical knowledge (not presented here but described in ref. [ ]) added till now is functioning well. ( ) it appears to be very easy to add knowl- edge, and make the universal lps situ- ation specific (e.g. by adding practical constraints). . conclusions and suggestions for further research in this paper we have outlined the structure and functions of a rule-based expert system (es), which has been developed to assist deci- sion making by a local planning and schedul- ing system (lps), operating at a production cell level. in particular, we have underlined the need to recognize and interpret complex situ- ations in the local planning environment, as well as the fact that any advice from the lps should be based on both theoretical knowl- edge (algorithms and heuristics) and factory specific practical knowledge. a prototype ver- sion of the expert system has been pro- grammed. finally, we have evaluated the performance to the es by considering some typical cases; this evaluation in particular ad- dresses the quality of the advisory function, not primarily the quality of the underlying algorithms (which is quite a different topic). further research might be concentrated on number of aspects, including: the influence of alterations in the condition part of the underlying rules. the knowledge evaluation has been based primarily on the values of the certainty factors. changing the range of values for which a condition of a specific rule should hold may severely in- fluence the ultimate performance of that rule, the extension of the es with a simulation part (cf. fig. ). such an extension yields the possibility to generate additional knowledge by simulating several alternatives. see ref. [ ] for some more elaborate ideas, knowledge acquisition in a real manufactur- ing environment. this will be the next topic in forthcoming research. g.j. meesterllocal planning environment acknowledgement isbn - - - . the author is grateful to professor w.h.m. zijm for detailed comments on earlier versions of this paper which have substantially im- proved the presentation of the material. c cl references iill jackson, p., . expert systems. addison- wesley, reading. horvitz, e.j., heckerman, d.e. and langlotz, c.p., . a framework for comparing alternative formalisms for plausible reasoning. in: proceed- ings fifth national conference on artificial intel- ligence. vol. : science. morgan kaufmann publishers, los altos, pp. , . turksen, i.b., . an approximate reasoning framework for aggregate production planning. nato as series, vol. , c.i.m, pp. , . levesque, h.j. and brachman, r.j., . a funda- mental tradeoff in knowledge representation and reasoning, pp. , . buchanan, b.c. and duna, r., . principles of rulebased expert systems. stanford university memo, hpp- - . meester, g.j., . een expertsysteem voor planning- en schedulingsalgoritmen, master thesis university of twente. saywer, b. and foster, d., . an expert system in pascal. vi c c c c c c pi bossink, g.j., . a framework for planning and scheduling manufacturing systems. in: proceed- ings workshop on decisional architectures in automated manufacturing, genova, italy. bossink, g.j., . planning and scheduling (flex- ible) manufacturing systems. in: iscie proceed- ings, japan-u.s.a. symposium on flexible automation, vol. iii, pp. . . bossink, g.j., . planning and scheduling for flexible discrete parts manufacturing. phd. thesis. isbn - - . university of twente. enschede, the netherlands. fox, b.p.. allen, m.s., smith, s.f. and strohm. g.a., . isis: a constraint directed reasoning approach to job shop scheduling. carnegie mel- lon univerity, cmu-ri-tr- - , pitsburg, usa. bensana, e., bel, g. and dubois, d., . opal. a knowledge based system for industrial jobshop scheduling. nato as series, vol. c.i.m, pp. , . wu, sd., . an expert system approach for the control and scheduling of flexible manufacturing cells. worden, b. and weber, e., . erfurungen emes expertsystems in konstructions berreich. in: v.d.i. berichte : expertsysteme in entwicklung und konstruktion, pp. , . bonnet, a., . l’in~elligence artificielle; promesses et r&alit&s. intereditions, paris, france, wi cl cl cl cl cl cl cl c buchanan, b.c., bartlow, d. and benet. j., . constructing an expert system. (eds. hayes-roth, waterman and lenat), chapter . adams, j., balas, e. and zwawack, d., . the shifting bottleneck procedure for job shop sched- uling. manage. sci., ( ). aanen, e., . planning and scheduling in flex- ible manufacturing systems. phd thesis, isbn - - , university of twente, enschede, the netherlands. fox, m.s., chang, w.y. and peng. s.o.. . fac- tory model and test data descriptions: opis experiments. internal report. carnegie mellon university. cmu-ri-tr- - , pitsburg, usa. ibm, knowledge processing series, expert system development environment, expert system con- sultation environment, and expert system refer- ence manual. naar de inhoud springen terms of use login zoeken naar: zoeken je moet eerst inloggen, daarna kun je een nieuwe site maken. technische universiteit delft auteursrecht © alle rechten voorbehouden.thema: flash door themegrill. aangedreven door wordpress symbiotic filtering for spam email detection clotilde lopes a paulo cortez a,∗ pedro sousa b miguel rocha b miguel rio c adep. of information systems/algoritmi, university of minho, - guimarães, portugal bdep. of informatics/cctc, university of minho, - braga, portugal cdep. of electrical engineering, university college london, wc e je uk abstract this paper presents a novel spam filtering technique called symbiotic filtering (sf) that aggregates distinct local filters from several users to improve the overall perfor- mance of spam detection. sf is an hybrid approach combining some features from both collaborative (cf) and content-based filtering (cbf). it allows for the use of social networks to personalize and tailor the set of filters that serve as input to the filtering. a comparison is performed against the commonly used naive bayes cbf algorithm. several experiments were held with the well-known enron data, under both fixed and incremental symbiotic groups. we show that our system is competitive in performance and is robust against both dictionary and focused con- tamination attacks. moreover, it can be implemented and deployed with few effort and low communication costs, while assuring privacy. key words: anti-spam filtering, naive bayes, collaborative filtering, content-based filtering, word attacks introduction unsolicited bulk email, widely known as spam, has become a serious prob- lem for network administrators and for internet users in general. according to maawg ( ), it accounts for % to % of all email messages sent and it consumes resources (i.e. time spent reading messages, bandwidth, cpu and disk) that are far away from negligible. spam is also an intrusion of privacy ∗ corresponding author. e-mail pcortez@dsi.uminho.pt; tel.: + ; fax: + . preprint submitted to elsevier november and used to spread malicious content (e.g. online fraud or viruses). the cost of sending these emails is, however, very close to zero. internet connections are cheap and with the advent of botnets (e.g. bobax worms), criminal orga- nizations have access to potentially millions of infected computers and thus send emails from what has to be regarded as legitimate users (ramachandran and feamster, ; kanich et al., ). in the last years, the most popular anti-spam solutions have been based in content-based filtering (cbf) (in particular, bayesian filtering) (garriss et al., ; guzella and caminhas, ). these class of algorithms use mes- sage features (e.g. word frequencies) for statistically discriminating email into legitimate (ham) and illegitimate (spam) messages. however, cbf presents some drawbacks. often, there is a large gap between the high-level concept (e.g. spam image) and the low-level message features (e.g. bit colors). also, cbf tends to give weak performances for new users, as it requires a large number of representative examples. moreover, spammers can mix spam with normal words (often not visible to the final user), in what is known as dictio- nary or focused attacks (nelson et al., ). when users flag these messages as spam, their training set is contaminated and the cbf performance is heavily reduced. as an alternative, collaborative filtering (cf) is a distinct anti- spam strategy, where information (e.g. ip or message fingerprints) is shared about spam messages (zhong et al., ). yet, pure cf often suffers from several issues (e.g. first-rater, sparsity of data and privacy, see section ). to solve the cbf drawbacks, we propose a novel distributed approach, termed symbiotic filtering (sf), that combines features from cbf and cf. symbiosis is a close interaction among different entities and this phenomenon is present not only in biological species but also in business enterprises (alocilja, ). we prefer the term symbiotic rather than collaborative or cooperative, since in this case each individual may have a distinct goal, as spam by definition is a personal concept. the idea is to promote a cooperation among distinct entities (e.g. email users) interested on personalized filtering (e.g. spam detection). rather than exchanging messages, these entities will share information about what each local filter has learned (e.g. bayesian model). the aim of sf is to foster mutual relationships, where all or most members benefit. under sf, a given user is interested in improving filtering at a personal level. the internet is used to gather collaborators among these (high number of) users. high group dynamics are expected, as members may join or leave the collaboration, and also there are privacy issues regarding what can be shared. sf is different from the centralized cbf-cf works (e.g. (yu et al., )) since sf data and models are distributed through different entities. hence, there are issues of user management (e.g. adding or removing a user), privacy, security and motivation (e.g. each user should benefit from the collaboration). the main contributions of this paper are: i) we propose the new sf concept that combines filters from distinct entities in order to improve local filtering while assuring privacy; ii) we apply sf to spam detection and compare it with a local cbf filter (i.e. naive bayes); iii) the spam detection performance is measured under distinct scenarios, to test the effect of using fixed and incremental symbiotic groups and also to access the robustness to dictionary and focused attacks. this paper is structured as follows. section presents the related work. next, we introduce the individual and symbiotic filtering methods (section ). the results are presented in section . finally, closing conclusions are drawn (section ). related work several solutions have been proposed to fight spam, which can fall into three main categories (garriss et al., ; méndez et al., ): designing new mail protocol systems (nmps), collaborative filtering (cf) and content-based filtering (cbf). examples of the first approach are: digitally signing mail, where recipients authenticate the sender’s address and mail content (wong, ); requiring the sender to “pay” (e.g. give a small fee or solve a com- putational puzzle) for each message sent (loder et al., ) or limiting the amount of email any sender may send (walfish et al., ). yet, none of these solutions is currently adopted in a massive fashion and we are still far from a worldwide acceptance of a nmps. also, the proposal of payment systems did not take into account the effect spamming botnets, where a large number ma- chines are controlled for malicious messaging (ramachandran and feamster, ; kanich et al., ). cf is based on sharing information about spam messages and it can be based on lists (e.g. blacklists with ip addresses of known spammers), or di- gest/fingerprints extracted from spam messages. often, dns-based blackhole lists (dnsbls) work in a centralized fashion, being vulnerable to denial-of- service (dos) attacks. also, they may blacklist legitimate users, with a false negative rate of about %, and spammers that use bgp spectrum agility techniques are rarely listed in dnsbls (ramachandran and feamster, ). several cf systems based on social networks have also been proposed. for example, kong et al. ( ) suggest a system where users manually identify spam and then publish a digest through her/his social network. garriss et al. ( ) propose the propagation of whitelists among socially connected users. zhong et al. ( ) introduce a large-scale privacy cf based on digests. cbf filters use a text classifier, such as the popular naive bayes algorithm (used by the thunderbird client), that learns to discriminate spam from mes- sage features (e.g. common spam words). at the present time, cbf is the most used anti-spam solution (garriss et al., ). current research relies mainly on improving individual classifier performance, by a better preprocess- ing (méndez et al., ) or enhancement of the learning algorithm (chang et al., ; guzella and caminhas, ). ensembles that combine distinct spam classifiers have also been proposed (hershkop and stolfo, ). both cf and cbf have drawbacks. cf often suffers from first-rater, sparsity of data and privacy problems. the first issue is due to the difficulty of classify- ing emails that have not been rated before, the second problem is present when users rate few messages and the last problem depends on what is shared. for example, while presenting a better privacy digest protocol (when compared with previous cf solutions), the approach of zhong et al. ( ) is still vul- nerable to privacy breaches. in addition, people have personal views of what is spam and cf often discards this issue (gray and haahr, ). on the other hand, cbf frequently suffers from lack of sufficient training messages and is highly vulnerable to contamination attacks (as explained in the previous section). by fusing the cf and cbf views there is a potential for a better personalized filtering. however, the number of studies that unify cbf and cf is scarce and mainly focused towards recommendation systems that run at centralized systems (yu et al., ). garg et al. ( ) describe a spam strategy based on sharing of email filters among collaborating users. the authors point out that exchanging filters requires less communication than cf systems, as the need to share filters is relatively rare when compared with exchanging digests each time an email is received. yet, they fail to acknowledge motivational (i.e. each user should benefit from the collaboration), temporal (i.e. how to syn- chronize several distinct filters), privacy and security issues regarding filter sharing. more recently, lai et al. ( ) presented a collaborative approach to exchange spam rules. however, this approach was designed for rule sharing at the server level, demanding secure channels between all these servers. also, the authors only explored simple attributes (e.g. message length) and not word content. moreover, explicit (and simple) human understandable rules are re- quired, thus approaches such as neural networks or support vector machines are not suitable. in contrast, our sf approach is more flexible, since it can address any type of cbf filter. also, it is suited for the web . paradigm, where users can exchange filters from their social networks. the sf approach should also be more robust to contamination when compared to pure cbf, since it aggregates responses from several (possible unknown) users and tar- geting a specific victim may be easy but contaminating the whole symbiotic group is not. filtering methods . naive bayes filtering we will address only textual content (i.e. word frequencies) of email mes- sages. this popular approach (e.g. thunderbird filter) has the advantage of being generalizable to wider contexts, such as spam instant messaging (spim) detection. while different data mining algorithms can be adopted for spam filtering, such as support vector machines (cheng and li, ), we will use the simpler naive bayes (nb), which is widely adopted by anti-spam filtering tools (metsis et al., ; guzella and caminhas, ). as both individual and symbiotic strategies will be compared using the same learning algorithm, we believe that most of the results presented in this paper can be extended to other text classifiers. we will also adopt the preprocessing proposed in (kosmopoulos et al., ): ( ) the word frequencies are extracted from the subject and body message (with the html tags previously removed). each message j is encoded into a vector xj = (x j, . . . ,xmj), where xij is the number of occurrences of token xi in the text. ( ) the feature selection is applied, which consists in ignoring any words when xij < in the training set and then selecting up to the most relevant features according to the mutual information (mi(xi)) crite- rion: mi(xi) = ∑ c∈{s,¬s} p(xi|c) log( p(xi|c) p(xi)p(c) ) ( ) where c is the message class (s - spam or ¬s - ham), p(xi|c) is the probability of finding token xi in emails from class c, p(xi) and p(c) are the proportions of xi terms and c class examples present in the data. ( ) each xij value is transformed into: x ′ ij = log(xij + ) (tf transform), x′′ij = x ′ ij ·log(k/ ∑ k δik) (idf transform) and x ′′′ ij = x ′′ ij/ √∑ l(xlj) (length normalization), where δik is if the token i exists in the message k and otherwise. the nb computes the probability that a document j is spam (s) for a filter trained over du email data from user u, according to: p(s|xj,du) = α ·p(s|du) m∏ i p(xi|s,du) ( ) where α is normalization constant that ensures that p(s|x,du)+p(¬s|x,du) = , p(s|du) is the p(s) of dataset du. the p(xi|s,du) estimation depends on the nb version. in this work, we will use the multi-variate gauss nb (as implemented in the r tool, see section ) (metsis et al., ): p(xi|c,du) = σi,c √ π exp(− (x′′′ij −µi,c) σ i,c ) ( ) where µi,s and σi,s are the mean and standard deviation estimated from the c = s or c = ¬s messages of du. in (nelson et al., ), it has been shown that local spam filters are vulnerable to dictionary and focused contamination attacks. the former attack is used to reduce the cbf efficiency, leading the victim to read spam, while the latter can be used to prevent the victim from reading an important email. both attacks can be achieved by sending spam messages mixed with normal words. once the victim labels these messages as spam, the training set is contaminated and the filter will be affected the next time it is retrained. a dictionary aggression consists in sending a large amount of normal words, while the focused assault assumes that the attacker has some knowledge of a specific message that the victim will receive in the future (e.g. a competing offer for a given contract). . symbiotic filtering in our proposed sf, the individual predictions can be combined by using a collaborative ensemble of the local filters. to tackle the concept drift (i.e. the learning tasks changes through time) nature of spam (fdez-riverola et al., ), the ideal symbiotic combination function should be dynamic. to achieve this we propose a hierarchical learning, where the outputs of the distinct fil- ters are used as the inputs of another (meta-level) learner. hence, each user has a local meta-learner that is responsible for aggregating the distinct filter responses. this meta-learner is dynamically trained to get a high accuracy on the user past data, thus it assigns different weights to the cbf filters through time (see figure ). while several algorithms could be used for this hierarchi- cal learning (e.g. svm), we will adopt the same nb described in the previous section. the rationale is that nb is commonly adopted by anti-spam solutions, thus incorporating sf into these tools would be simpler by reuse of code. we assume that each user u trains a local filter θu,t over her/his du training data. filters can be trained asynchronously and l filters will be available for each user at time t: {θ ,t, . . . ,θl,t}. the symbiotic nb (snb) meta-model spam probability is given by: p(s|xj,d′u) = α ·p(s|d′u) ∏l i= p(θi,t|s,d′u) p(θi,t|c,d′u) = σi,c √ π exp(−(p(s|xj,θi,t,d ′ u)−µi,c) σ i,c ) ( ) where d′u is the snb training set and p(s|xj,θi,t,d′u) is the probability given by the filter θi,t, as computed in equation . to reduce memory and compu- tational requirements, we allow that d′u ⊆ du, where m = |d′u| denotes the most recent messages from u mailbox. it should be noted that any token from xj that is not considered by θi,t will simply be discarded by the filter from user i. similarly, any input attribute from θi,t that is not included in xj will be set to . while sharing models is less sensitive than exchanging email messages, there are still privacy issues to be considered. for instance, if user a has access to the filter of user b, then a may feed a given token (or set of tokens) into the model and thus know some probability that such token was classified by b as spam or ham. our privacy solution resides in an anonymous exchange of the filters, which can occur under a centralized server or a peer-to-peer (p p)-like application. under the first option, all users register into a centralized and secure service. this service could be implemented by large companies or email providers (e.g. gmail or hotmail), when all emails are stored at a given server. for scalabil- ity, user profiles could be defined (e.g. country or profession) and clustering algorithms could be used to group users with similar interests. another vari- ant would be the definition of social networks, where users could choose their “friends”. these systems could, for example, be implemented by social net- working websites (e.g. facebook or myspace). alternatively, when the mes- sages are stored locally at the client side, the server would be responsible for a blind exchange of the filters, using secure transfers (left of figure ). to exchange the filters, a standard format should be adopted, such as the pre- dictive model markup language (pmml) (grossman et al., ), which is compatible with a large number of data mining tools. in should be noted that exchanging filters requires less communication costs. for example, a filter built from a millions of emails can be described by a few hundreds or thousands of bytes (depending on the filter algorithm used) (garg et al., ). when a new email is received, the user can easily compute the sf, as a copy of all filters is available locally. the above systems have the disadvantage of having to depend and trust in a centralized service. as an alternative, the use of a p p-like distribution scheme is also possible to be adopted (right of figure ). under this solution all peers may donate, store and fetch filters among each other. this approach anonymousanonymous filter filter user a user c user b anonymous filter filter anonymous filter a filter b user c user buser a server secure fig. . anonymous exchange of filters by using a secure server (left) or a p p-like application (right) could be implemented as a trusted and secure plug-in of an email client (e.g. thunderbird). the filter sharing process among the peers could work similarly to the explained in the secure server scheme (i.e. using pmml). further pri- vacy increase could be achieved if each individual does not know the symbiotic group composition. however, for some scenarios it may also be attractive that the symbiotic group composition could be assessed by all the participants in a given group. yet, even in such scenarios it might be very difficult to “guess” who created each particular model, as in sf there will be typically a large number of users that dynamically may join or leave the collaboration. experiments and results . spam data to evaluate sf, ideally there should be real mailboxes collected from distinct users (possibly from a social network) during a given time period. yet, due to logistic and privacy issues, it is quite difficult to obtain such data (in par- ticular personal messages) and make it public. hence, we will use a synthetic mixture of real spam and ham messages, in a strategy similar to what has been proposed in (metsis et al., ; zhong et al., ). the ham messages will come from the enron email collection, which was originally used for a global evaluation of filters by merging all enron user messages into a single corpus. in particular, we will use the cleaned-up form provided by (becker- mann et al., ) and we will select the five enron employees with the largest mailboxes collected during the same time period: kaminski-v (kam), farmer-d (far), beck-s (bec), lokay-m (lok) and kitchen-l (kit). since these employees worked at the same organization, it is reasonable to assume that they would know each other, i.e. belong to a social network. we will also use the spam collection of bruce guenter (http://untroubled.org/spam/), which is based in spam traps (i.e. fake emails published in the web), during the years of and (our dataset was built in ). only messages with latin character sets were selected, because the ham messages use this type of character coding and non-latin mails would be easy to detect. also, since this collection con- tains several copies of the same messages (due to the use of multiple traps), we removed duplicates by comparing md signatures of the body messages. we propose a mixture algorithm that is based on the time that each message was received (date field, using the gmt time zone). by preserving the tem- poral order of the emails, we believe that a more realistic mixture is achieved than the sampling procedures adopted in (metsis et al., ) or (zhong et al., ). since the enron data is from a previous period (see table ), we first added years to the date field of all ham messages. let st denote a spam mes- sage received at time t, si,f = (sti,sti+ , . . . ,stf ) the time ordered sequence of the bruce guenter spam, hu,i,f and su,i,f the sequences of ham and spam messages for user u from time ti to tf . for a given time period t ∈ (ti, . . . , tf ), the algorithm randomly selects |s′i,j| spam messages from si,j. then, su,i,j is set by sampling messages from s′i,j with a probability of p for each message selection. the size of s′i,j (cardinality) is given by: |s′i,j| = r· ∑l i= |hu,i,j| p ·l ( ) where l denotes the total number of users available at the time period and r is the overall (i.e. including all user and time data) spam/ham ratio. since the time periods are different for each user (table ), four time sequences (i.e. ti and tj values) were used by the algorithm (figure ). table the s-enron corpus main characteristics user ham spam time spam size size period /ham kam [ / , / ] . far [ / , / ] . bec [ / , / ] . lok [ / , / ] . kit [ / , / ] . the mixture is affected by the r and p parameters. since a high number of ex- periments is addressed in this work, we will fix these parameters to reasonable values. while the global spam/ham ratio is r = , the individual ratios range from . to . . also, the spam/ham ratios fluctuate through time (as shown in figure ). on the other hand, the probability of spam selection affects the percentage of common spam between users. if two users have similar profiles (e.g. email exposure), then they should receive similar spam. we assume that this scenario is expected for the enron employees and thus set p = . . under this setup and for a given time period, any users will receive around % of date (year) kam far bec lok kit l= l= l= l= a b c fig. . time view of the s-enron mailboxes similar spam, users will share around % of spam and so on. the result- ing corpus is named s-enron and it is publicly available in its raw form at: http://www .dsi.uminho.pt/pcortez/s-enron. . . . . batch (x emails) ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● fig. . evolution of the spam/ham ratio for the user far and scenario c . evaluation as explained in section . , spam detection suffers from concept drift (fdez- riverola et al., ). features such as the amount of spam received, the ham/spam ratio and even the content itself evolve through time. hence, to evaluate spam filters, we will adopt the more realistic incremental retraining evaluation procedure, which periodically trains and tests filters. under this procedure, a mailbox is split into batches b , . . . ,bn of k adjacent messages (|bn| may be less than k) (metsis et al., ). for i ∈ { , . . . ,n − }, the filter is trained with du = b ∪ . . .∪bi and tested with the messages from bi+ (figure ). this procedure is more realistic than the simple % train/test split adopted in (zhong et al., ). the predicted class for a probabilistic filter is given by s if p(s|xj,du) > d, where d ∈ [ . , . ] is a decision threshold. for a given d and test set, it is k |b | test set test set test set training set tr. set training set n− training set ts. set ...... runs b b b batch b n ... mailbox kkk n fig. . the incremental retraining procedure possible to compute the true (tpr) and false (fpr) positive rates: tpr = tp/(tp + fn) fpr = fp/(tn + fp) ( ) where tp , fp , tn and fn denote the number of true positives, false posi- tives, true negatives and false negatives. the receiver operating characteristic (roc) curve shows the performance of a two class classifier across the range of possible threshold (d) values, plotting fpr (x-axis) versus tpr (y-axis) (fawcett, ). the global accuracy is given by the area under the curve (auc= ∫ rocdd). a random classifier will have an auc of . , while the ideal value would be . . since the cost of losing normal email (fp) is much higher than receiving spam (fn), d is usually set to favor points in the low false-positive region of the roc. thus, we will also adopt the normalized auc (nauc), which is the auc area in the section fpr ≤ r, divided by r (chang et al., ). typically, the target fpr rate (r) is close to . . we will also compute the relative gain of the symbiotic performance over the local filter: gain= ξsnb/ξnb − , where ξ is the evaluation metric (i.e. auc or nauc) and snb and nb are the symbiotic and individual filters. with the incremental retraining procedure, one roc will be computed for each bi+ batch and the overall results will be presented by adopting the vertical aver- aging roc (i.e. according to the fpr axis) algorithm presented in (fawcett, ). statistical confidence will be given by a paired t-student test, at the % confidence level (flexer, ). . experimental setup all experiments were conducted in the r environment, an open source and high-level programming language for data analysis (r development core team, ). in particular, the nb algorithm described in section . is implemented by the naivebayes function of the e r package, while the text prepro- cessing uses several functions from the tm package (feinerer et al., ). during the all experiments, we set k = (a reasonable value also adopted in (metsis et al., )). for the snb, we used a similar number for the hierarchical training set size, i.e. m = . this small value has the advantage of reducing memory requirements (the user only needs messages in his mailbox) and some initial experiments with larger values of m revealed no gain in performance. the auc, nauc and gain values will be shown in percentage. all nauc values will be computed with r = . ( %). . fixed symbiotic group two distinct scenarios will be tested, according to the time periods a and b of figure . given the s-enron corpus characteristics, in this work we will explore a small number of fixed symbiotic users: l = for a and l = for b. the incremental retraining method (section . ) was applied to both sce- narios, by considering all messages within the corresponding time period. thus, the number of kam, far and bec batches (n) will be different for a and b. the obtained results are summarized as the mean of all test sets (bi+ , i ∈ { , . . . ,n − }) and shown in table and figure . the best values are in bold, while underline denotes a statistical significance (i.e. p-value< . ). in figure , bars denote % t-student confidence intervals and only the most interesting region of fpr is shown for the roc curves. for the first scenario (a), the symbiotic strategy outperforms the local filter for all users and metrics, except for lok and nauc. a similar behavior occurs for the b setting, where snb is better than nb except for kam and auc. as false positives have higher costs in spam detection, the nauc results are particularly important. thus, it is interesting to notice that there is a high auc improvement (i.e. gain) given by the symbiotic method in several cases (kam, far and kit for a and bec for b). to demonstrate the snb dynamics, figure shows the first two consecutive graphs of the snb input importances under scenario ii. each edge represents the influence (in %) of the nb filter (the origin) in the symbiotic model (the destination), as measured by applying a sensitivity analysis procedure (kewley et al., ). the text in bold (e.g. b ) denotes the last batch used to train the nb classifier. for example, the first snb model of user far (left graph) uses a nb filter from kam that was trained using messages (dkam = b ∪ b ). table the results for scenarios a and b scenario a auc nauc user n nb snb gain nb snb gain kam . . . . far . . . . bec . . . . lok . . . . - kit . . . . scenario b auc nauc user n nb snb gain nb snb gain kam . . - . . . far . . . . . bec . . . . . . incremental symbiotic group a more realistic scheme is adopted for the time period c, where users join the symbiotic group in an incremental fashion, at different time stages according to figure . thus, l will grow from to . the results are presented in table . as expected, the symbiotic strategy (snb) clearly favors newcomers, which have small mailboxes and thus benefit from the collaboration. in effect, the nauc differences are quite large, such as in bec and lok for l = and l = ; and kit for l = . for demonstration purposes, the roc curves are plotted for kam, far and bec, when l = and l = (figure ). however, the results show that even “veteran” users benefit from the symbiotic relation when the number of users grow. for instance, the kam and far nauc results for l = improve, with a gains of % and %, respectively. . contamination attacks we will repeat the experiments of section . , by considering only scenario a and user bec to test the effects of mailbox contamination. the dictionary assault is simulated by replacing the first spam emails at batch from bec by the gnu aspell (http://aspell.net/) english dictionary (version . , with tokens). in figure , gray lines denote the behavior of nb and snb without the attack (i.e. results of section . ), black lines show the per- formance under the attack and the dot-dashed vertical line shows when the . . . . . . . . . . . . kam (a) ● ● ● ● ● ● ● ● ● ● ● ● snb nb . . . . . . . . . . . . far (a) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . bec (a) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . lok (a) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . kit (a) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . kam (b) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . far (b) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . bec (b) ● ● ● ● ● ● ● ● ● ● ● fig. . roc curves for scenarios a and b bec kam far bec kam far b b b b b b b b b b % % % % % % % % % % % % % % % b % b b b b b fig. . examples of snb input importances for b attack starts. the filters of bec are not trained with contaminated messages at batch , yet these messages appear in the test set and thus the nb and snb performances suffer a moderate decay. the true effect of the attack is only visible at batch , where local cbf is highly affected. only messages were replaced and yet the filter detection capability is reduced to a random classi- fier (since auc= . ) through all remaining batches. in contrast, the symbiotic . . . . . . . . . . . . kam (l= ) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . kam (l= ) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . far (l= ) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . far (l= ) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . bec (l= ) ● ● ● ● ● ● ● ● ● ● ● . . . . . . . . . . . . bec (l= ) ● ● ● ● ● ● ● ● ● ● ● fig. . roc curves for users kam, far and bec (l= at left and l= at right) method is only initially affected, since as time goes by the performance gets closer to the no attack scenario. also, the remaining symbiotic users maintain their spam detection capabilities, as shown by the far results, which is a rep- resentative example. this behavior is explained by the snb algorithm, which simply discards a given filter if it does not help to predict the recent past m messages of the user. hence, this experiment shows that sf is robust also to saboteurs, i.e. if a particular user intentionally feeds the group with a random or bad filter then this filter will be simply ignored. . . . . . . bec test set batch (x e−mails) a u c ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● snb nb snb (attack) nb (attack) . . . . . . far test set batch (x e−mails) a u c ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● snb snb (attack) fig. . the effect of the dictionary attack the dictionary attack can be solved by performing a rollback (i.e. returning table the results for scenario c l= auc nauc user n nb snb gain nb snb gain kam . . - . . far . . - . . - l= auc nauc user n nb snb gain nb snb gain kam . . - . . far . . . . . bec . . . . l= auc nauc user n nb snb gain nb snb gain kam . . . . . . far . . . . . bec . . . . lok . . . . l= auc nauc user n nb snb gain nb snb gain kam . . - . . . far . . . . . bec . . . . lok . . . . kit . . . . to the previous filter) or using the roni defense (nelson et al., ), which rejects training examples that have a large negative impact in spam detection. yet, focused assaults are more difficult to prevent and finding a defense is still an open problem (nelson et al., ). we believe sf is an interesting solution due to the same rationale presented for the dictionary aggression, i.e. the combination of multiple filters should overcome the limitations of a single model contamination. a new set of experiments was devised, using again scenario a and bec mailbox. during a given run, a legitimate message was randomly selected, from batches to , as the target text. we assume that the attacker is confident about the target content and thus can guess % of the target words. at batch , spam emails were replaced by the contaminated messages. we repeated this procedure during runs. the effect of this attack on spam is minimal and thus we will only show the effect on the target ham emails. figure plots the filter spam probability (y-axis) for each target message. the obtained probability for each run is plotted along the x-axis (total of runs). since all target messages are ham, a robust filter should present low spam probabilities, near the zero horizontal axis. the results show that local filter (nb) is much more vulnerable to focused attacks than the symbiotic strategy (snb). the spam probability mean values of nb and snb are . and . (the differences are statistically significant). for example, when using a decision threshold of d = . , (of ) messages are classified by nb as spam, while this number lowers to for snb. even if d is raised to . , nb predicts spam emails and snb only detects . . . . . . . runs sp a m p ro b a b ili ty f o r ta rg e t te xt ( h a m ) ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● snb nb fig. . the effect of the focused attack conclusions email has long been one of the most important and widely used internet ap- plications. however, spam emerged quickly after email itself and nowadays it accounts for the majority of the email traffic. thus, several anti-spam tech- niques were developed. this paper proposes a novel symbiotic filtering (sf) approach, which combines several features from collaborative filtering (cf) and content-based filtering (cbf). it takes the advantage of use of social networks, where users with the same or related interests have the opportunity to form mutually beneficial alliances with the aim of enhancing spam detec- tion techniques. instead of sharing messages or digests, the idea is to share filters. to combine the individual probabilities, the proposed solution uses the concept of hierarchical learning, where a meta-learner is dynamically trained to improve its accuracy. after describing the proposed model, we compared the effectiveness of the sf versus cbf, using for that purpose a realistic mixture of real spam and ham messages. promising results were obtained by the sf, which outperformed the local filtering for a small number of users (from to ). moreover, we have shown that sf is more robust to word attacks (e.g. dictionary or focused as- saults). furthermore, we proposed several deployment scenarios for sf, under a secure server or p p settings. there is a continuous race between spammers and anti-spammers and a local classifier that is currently perfect will be eventually defeated. we believe that a stronger protection is achieved by adopting a dynamic cooperation of filters from distinct users. in future work, we intent to apply sf to other personalized filtering scenarios, such as web page blocking (e.g. sensitive content). also, we will explore scalability issues. under a large group, this could be achieved by adopting user selection algorithms (e.g. clustering user profiles). acknowledgments this work is supported by fct grant ptdc/eia/ / . references alocilja, e. ( ). principles of biosystems engineering. erudition books, massachusetts, usa. beckermann, r., mccallum, a., and huang, g. ( ). automatic catego- rization of email into folders: benchmark experiments on enron and sri corpora. ir- , university of massachusetts amherst. chang, m., yih, w., and meek, c. ( ). partitioned logistic regression for spam filtering. in th acm sigkdd int. conference on knowledge discovery and data mining, pages – . cheng, v. and li, c. ( ). personalized spam filtering with semi- supervised classifier ensemble. in ieee/wic/acm international con- ference on web intelligence. fawcett, t. ( ). an introduction to roc analysis. pattern recognition letters, ( ): – . fdez-riverola, f., iglesias, e. l., dı́az, f., méndez, j. r., and corchado, j. m. ( ). applying lazy learning algorithms to tackle concept drift in spam filtering. expert systems with applications, ( ): – . feinerer, i., hornik, k., and meyer, d. ( ). text mining infrastructure in r. journal of statistical software, ( - ). flexer, a. ( ). statistical evaluation of neural networks experiments: minimum requirements and current practice. in proceedings of the th european meeting on cybernetics and systems research, volume , pages – , vienna, austria. garg, a., battiti, r., and cascella, r. ( ). may i borrow your filter? exchanging filters to combat spam in a community. in advanced infor- mation networking and applications, . aina . th international conference on, volume . garriss, s., kaminsky, m., freedman, m., karp, b., mazières, d., and yu, h. ( ). re: reliable email. in proceedings of the rd conference on networked systems design and implementation (nsdi), pages – , san jose, ca. usenix association berkeley, ca, usa. gray, a. and haahr, m. ( ). personalised, collaborative spam filtering. in st conference on e-mail and anti-spam ceas. grossman, r., hornick, m., and meyer, g. ( ). data mining standards initiatives. communications of acm, ( ): – . guzella, t. and caminhas, w. ( ). a review of machine learning ap- proaches to spam filtering. expert systems with applications, : – . hershkop, s. and stolfo, s. ( ). combining email models for false positive reduction. in th acm sigkdd int. conference on knowledge discovery and data mining, pages – . kanich, c., kreibich, c., levchenko, k., enright, b., voelker, g., paxson, v., and savage, s. ( ). spamalytics: an empirical analysis of spam mar- keting conversion. in computer and communications security conference (ccs’ ), pages – . acm. kewley, r., embrechts, m., and breneman, c. ( ). data strip mining for the virtual design of pharmaceuticals with neural networks. ieee trans neural networks, ( ): – . kong, j., boykin, p., rezaei, b., sarshar, n., roychowdhury, v., rothenstein, b., damian, i., paramonov, p., lyuksyutov, s., barrat, a., et al. ( ). let your cyberalter ego share information and manage spam. eprint arxiv: physics/ . kosmopoulos, a., paliouras, g., and androutsopoulos, i. ( ). adaptive spam filtering using only naive bayes text classifiers. in ceas - fifth conference on email and anti-spam. lai, g., chen, c., laih, c., and chen, t. ( ). a collaborative anti-spam system. expert systems with applications, : – . loder, t., van alstyne, m., and wash, r. ( ). an economic answer to unsolicited communication. in proceedings of the th acm conference on electronic commerce, pages – . acm new york, ny, usa. maawg ( ). email metrics program: the network operators’ perspec- tive. report # – third and fourth quarter , messaging anti-abuse working group, s. francisco ca, usa. méndez, j., cid, i., glez-peña, d., rocha, m., and fdez-riverola, f. ( ). a comparative impact study of attribute selection techniques on näıve bayes spam filters. in springer, editor, th industrial conference on data mining, volume lnai , pages – . metsis, v., androutsopoulos, i., and paliouras, g. ( ). spam filtering with naive bayes – which naive bayes? in third conference on email and anti-spam (ceas), pages – . nelson, b., barreno, m., chi, f., joseph, a., rubinstein, b., saini, u., sutton, c., tygar, j., and xia, k. ( ). exploiting machine learning to subvert your spam filter. in st usenix workshop on large-scale exploits and emergent threats, pages – . acm press. r development core team ( ). r: a language and environment for statis- tical computing. r foundation for statistical computing, vienna, austria, isbn - - - , http://www.r-project.org. ramachandran, a. and feamster, n. ( ). understanding the network- level behavior of spammers. in acm, editor, sigcomm’ , pages – . walfish, m., zamfirescu, j., balakrishnan, h., karger, d., and shenker, s. ( ). distributed quota enforcement for spam control. in proceedings of the rd conf. on networked systems design and implementation (nsdi), pages – , san jose, ca. usenix association berkeley. wong, m. ( ). sender authentication: what to do. white paper, july. yu, k., schwaighofer, a., tresp, v., ma, w., and zhang, h. ( ). collab- orative ensemble learning: combining collaborative and content-based information filtering via hierarchical bayes. in th int. conf. on uncer- tainty in artificial intelligence (uai), pages – . acm. zhong, z., ramaswamy, l., and li, k. ( ). alpacas: a large-scale privacy-aware collaborative anti-spam system. in ieee infocom, pages – . author information pack apr www.elsevier.com/locate/eswa expert systems with applications an international journal author information pack table of contents . xxx . • description • audience • impact factor • abstracting and indexing • editorial board • guide for authors p. p. p. p. p. p. issn: - description . expert systems with applications is a refereed international journal whose focus is on exchanging information relating to expert and intelligent systems applied in industry, government, and universities worldwide. the thrust of the journal is to publish papers dealing with the design, development, testing, implementation, and/or management of expert and intelligent systems, and also to provide practical guidelines in the development and management of these systems. the journal will publish papers in expert and intelligent systems technology and application in the areas of, but not limited to: finance, accounting, engineering, marketing, auditing, law, procurement and contracting, project management, risk assessment, information management, information retrieval, crisis management, stock trading, strategic management, network management, telecommunications, space education, intelligent front ends, intelligent database management systems, medicine, chemistry, human resources management, human capital, business, production management, archaeology, economics, energy, and defense. papers in multi- agent systems, knowledge management, neural networks, knowledge discovery, data and text mining, multimedia mining, and genetic algorithms will also be published in the journal. reproducibility badge initiative and software publication reproducibility badge initiative (rbi) is a collaboration with code ocean (co), a cloud based computational reproducibility platform that helps the community by enabling sharing of code and data as a resource for non-commercial use. co verifies the submitted code (and data) and certifies its reproducibility. code submission will be verified by the code ocean team for computational reproducibility by making sure it runs, delivers results and it is self-contained. for more information please visit this help article. note that an accepted paper will be published independently of the co application outcome. however, if the paper receives the reproducibility badge, it will be given additional exposure by having an attached r badge, and by being citable at the co website with a doi. we invite you to convert your open source software into an additional journal publication in software impacts, a multi-disciplinary open access journal. software impacts provides a scholarly reference to software that has been used to address a research challenge. the journal disseminates impactful and re-usable scientific software through original software publications which describe the application of the software to research and the published outputs. for more information contact us at: software.impacts@elsevier.com benefits to authors https://help.codeocean.com/en/articles/ -code-ocean-s-verification-process-for-computational-reproducibility https://www.journals.elsevier.com/software-impacts https://www.journals.elsevier.com/software-impacts author information pack apr www.elsevier.com/locate/eswa we also provide many author benefits, such as free pdfs, a liberal copyright policy, special discounts on elsevier publications and much more. please click here for more information on our author services. please see our guide for authors for information on article submission. if you require any further information or help, please visit our support center audience . knowledge engineers, managers, systems engineers, automation and control engineers, electronic and electrical engineers. impact factor . : . © clarivate analytics journal citation reports abstracting and indexing . cambridge/computer and information abstracts web of science research alert current contents - engineering, computing & technology science citation index expanded scopus inspec editorial board . editor-in-chief binshan lin, louisiana state university in shreveport, shreveport, louisiana, united states of america founding editor jay liebowitz, university of maryland global campus, adelphi, maryland, united states of america associate editors rodrigo agerri, university of the basque country upv/ehu, hitz center - ixa, donostia-san sebastian, spain information extraction, sequence labelling, natural language processing stuart barnes, king’s college london, king’s business school, london, united kingdom sharing economy, consumer behaviour, sustainability, digital innovation, machine learning vincent charles, pontifical catholic university of peru centrum graduate business school, lima, peru management science, artificial intelligence, data science, big data kwai-sang chin, city university of hong kong department of systems engineering and engineering management, hong kong, hong kong decision support systems, management intelligence, quality and engineering management sumali conlon, university of mississippi, university park, mississippi, united states of america business intelligence, expert and intelligent systems, machine learning, data science, natural language processing, intelligent systems and ai in business & erik cuevas, university of guadalajara university center of exact sciences and engineering, guadalajara, mexico research topics/areas, optimization, metaheuristic methods, evolutionary computation, swarm intelligence, image processing pierpaolo d'urso, university of rome la sapienza department of social and economic sciences, roma, italy fuzzy clustering, clustering of spatial data, clustering of time series, clustering of spatio-temporal data, clustering of complex structures of data, imprecise data science reggie davidrajuh, university of stavanger, stavanger, norway discrete-event systems, petri nets, modeling, simulation, performance evaluation, algorithms zhi-hong deng, peking university school of electronics engineering and computer science, beijing, china machine learning, natural language processing, data mining debora di caprio, university of trento department of economics and management, trento, italy knowledge-based systems, algorithm, statistic computing, fuzzy system, soft computing, intelligent systems and ai in business & social science https://www.elsevier.com/authors/author-services https://www.elsevier.com/journals/expert-systems-with-applications/ - /guide-for-authors https://service.elsevier.com/app/home/supporthub/publishing/ author information pack apr www.elsevier.com/locate/eswa giorgio maria di nunzio, university of padua department of information engineering, padova, italy interactive machine learning, visual data mining, information retrieval theory, big data, linked open data, digital geolinguistics sofia balula dias, ciper, faculdade de motricidade humana, universidade de lisboa, cruz quebrada, portugal intelligent learning management systems, artificial intelligence, serious games, blended-learning, fuzzy logic, deep learning, higher education, exergames, affective learning, collaborative learning, online learning elisabetta fersini, university of milan-bicocca department of informatics systems and communication, milano, italy machine learning, natural language processing, data science iztok jr fister, university of maribor, maribor, slovenia computational intelligence, data mining, sport, swarm intelligence c. bryan foltz, the university of tennessee at martin, college of business and global affairs, martin, tennessee, united states of america information systems security, privacy, web development francisco guijarro, polytechnic university of valencia, valencia, spain business intelligence, trading rules, portfolio management, valuation jin-kao hao, university of angers, angers, france combinatorial optimization, metaheuristics, large scale optimization, applications li-an ho, tamkang university, taipei, taiwan knowledge management, project management, information management, human resources management, human capital, instructional design alessio ishizaka, neoma business school, mont st aignan, france multi-criteria decision analysis, supply chain management, decision support systems, operations management, decision analysis dragi kocev, jozef stefan institute, ljubljana, slovenia ensemble learning, structured output prediction, multi label classification, hierarchical multi label classification, semi-supervised learning, feature selection salim lahmiri, concordia university, john molson school of business, department of supply chain and business technology management, montréal, quebec, canada data science, data analytics, signal processing, intelligent systems, pattern recognition voon-hsien lee, tunku abdul rahman university - perak campus, kampar, malaysia business intelligence, intelligent systems and ai in business & social science, quality management, supply chain management chunping li, tsinghua university, school of software, beijing, china artificial intelligence, data analysis and data mining, machine learning and automated reasoning chinho lin, national cheng kung university department of industrial and information management and institute of information management, tainan, taiwan knowledge management, strategic management, supply chain management, total quality management peter love, curtin university, perth, western australia, australia infrastructure, safety, resilience engineering, asset management, information systems evaluation, quality management, computer vision, engineering informatics, deep learning, machine learning yannis marinakis, school of production engineering & management, chania, greece swarm intelligence, vehicle routing problem, supply chain management, metaheuristic algorithms, optimization yalda mohsenzadeh, massachusetts institute of technology, cambridge, massachusetts, united states of america rui neves, university of lisbon higher technical institute, lisboa, portugal algorithms applied to the stock market luiz eduardo de oliveira, federal university of parana, curitiba, brazil pattern recognition, machine learning, image processing keng-boon ooi, ucsi university, kuala lumpur, malaysia artificial neural networks for machine learning, decision support systems, data mining joanna paliszkiewicz, warsaw university of life sciences, warszawa, poland trust management, knowledge management, social media, leadership, trust zbigniew pastuszak, maria curie-sklodowska university in lublin faculty of economics, lublin, poland business intelligence, knowledge-based systems, data science, decision support systems, prediction systems and early warning systems, data driven optimization, intelligent systems and ai in business & social science guojie song, peking university, beijing, china data mining, social network analysis, intelligent transportation system ying tan, peking university school of electronics engineering and computer science, beijing, china author information pack apr www.elsevier.com/locate/eswa fields of specialization- deep neural networks, deep learning, swarm intelligence, fireworks algorithms, data mining, financial predication joao manuel r.s. tavares, university of porto faculty of engineering, porto, portugal biomechanics, medical imaging, computational vision, gait &, posture, biodevices, image processing &, analysis, ndt manoj kumar tiwari, national institute of industrial engineering, mumbai, india decision support models, simulation model, computational intelligence, machine learning, supply chain and logistics, manufacturing systems, , conrad tucker, carnegie mellon university, pittsburgh, pennsylvania, united states of america machine learning , intelligent systems and ai in engineering, expert and intelligent systems julien velcin, lumière university lyon , lyon, france artificial intelligence, machine learning, data mining, natural language processing, social media analysis, digital humanities ling wang, tsinghua university department of automation, beijing, china computational intelligence, scheduling, meta-heuristics, soft computing, intelligent optimization ping wang, james madison university, department of computer information systems & business analytics, harrisonburg, virginia, united states of america supply chain management models and analysis, structural equation modeling, multivariate analysis, classification problems june wei, university of west florida college of business, pensacola, florida, united states of america business intelligence, knowledge-based systems petros xanthopoulos, stetson university school of business administration, deland, florida, united states of america business analytics, machine learning, operations research hua xu, tsinghua university department of computer science and technology, beijing, china natural language processing, intelligent optimization algorithms, intelligent information processing, multimodal information processing and mining steve yang, stevens institute of technology wesley j howe school of technology management, hoboken, new jersey, united states of america microstructure, behavioral finance, portfolio optimization, algorithmic trading, financial systemic risk editorial advisory board pornthep anussornnitisarn, kasetsart university, bangkok, thailand dianne berry, university of reading, reading, united kingdom bernt bremdal, university of oslo, oslo, norway francisco javier cantu, monterrey institute of technology and higher education, monterrey, mexico kimiz dalkir, mcgill university, montreal, quebec, canada randall davis, massachusetts institute of technology, cambridge, massachusetts, united states of america edward feigenbaum, stanford university, stanford, california, united states of america roar fjellheim, computas as, oslo, norway mark fox, university of toronto, toronto, ontario, canada olli-pekka hilmola, lut university - kouvola unit, kouvola, finland jae kyu lee, korea advanced institute of science and technology college of business, seoul, south korea dusan lesjak, international school for social and business studies, celje, slovenia ana gabriela maguitman, national university of the south, bahia blanca, argentina shawn mankad, weill cornell medicine, new york, new york, united states of america vladimir milacic, university of belgrade, belgrade, serbia arcot desai narasimhalu, singapore management university, singapore, singapore daniel o'leary, university of southern california, los angeles, california, united states of america zbigniew pastuszak, maria curie-sklodowska university in lublin faculty of economics, lublin, poland cezar scarlat, polytechnic university of bucharest, bucuresti, romania randall shumaker, university of central florida, orlando, florida, united states of america james slagle, university of minnesota, minneapolis, minnesota, united states of america ching y. suen, concordia university, montréal, quebec, canada jie tang, tsinghua university, beijing, china efraim turban, university of hawai'i at manoa, honolulu, hawaii, united states of america ismail burhan turksen, university of toronto, toronto, ontario, canada jan vanthienen, ku leuven, leuven, belgium shouhong wang, university of massachusetts dartmouth, dartmouth, massachusetts, united states of america mark zielinski, harvard medical school, boston, massachusetts, united states of america author information pack apr www.elsevier.com/locate/eswa guide for authors . introduction expert systems with applications is a refereed international journal whose focus is on exchanging information relating to expert and intelligent systems applied in industry, government, and universities worldwide. the thrust of the journal is to publish papers dealing with the design, development, testing, implementation, and/or management of expert and intelligent systems, and also to provide practical guidelines in the development and management of these systems. types of paper the journal will publish papers in expert and intelligent systems technology and application in the areas of, but not limited to: finance, accounting, engineering, marketing, auditing, law, procurement and contracting, project management, risk assessment, information management, information retrieval, crisis management, stock trading, strategic management, network management, telecommunications, space education, intelligent front ends, intelligent database management systems, medicine, chemistry, human resources management, human capital, business, production management, archaeology, economics, energy, and defense. papers in multi-agent systems, knowledge management, neural networks, knowledge discovery, data and text mining, multimedia mining, and genetic algorithms will also be published in the journal. contact details for submission authors are requested to submit their original papers using the submission website. to submit your paper online, please go to https://www.editorialmanager.com/eswa/default.aspx. authors interested in online submission are requested to go to the website and upload their manuscript and its associated artwork. an electronic (pdf) proof is generated and the reviewing process is carried out using that pdf. the pdf file may be edited after acceptance to follow journal standards. authors and editors send and receive all correspondence by email via the website and no paper correspondence is performed. for the name and address of the editor-in-chief, please refer to the list of editors in each issue of the journal or on the journal's homepage. submission checklist you can use this list to carry out a final check of your submission before you send it to the journal for review. please check the relevant section in this guide for authors for more details. ensure that the following items are present: one author has been designated as the corresponding author with contact details: • e-mail address • full postal address all necessary files have been uploaded: manuscript: • include keywords • all figures (include relevant captions) • all tables (including titles, description, footnotes) • ensure all figure and table citations in the text match the files provided • indicate clearly if color should be used for any figures in print graphical abstracts / highlights files (where applicable) supplemental files (where applicable) further considerations • manuscript has been 'spell checked' and 'grammar checked' • all references mentioned in the reference list are cited in the text, and vice versa • permission has been obtained for use of copyrighted material from other sources (including the internet) • a competing interests statement is provided, even if the authors have no competing interests to declare • journal policies detailed in this guide have been reviewed • referee suggestions and contact details provided, based on journal requirements author information pack apr www.elsevier.com/locate/eswa for further information, visit our support center. before you begin ethics in publishing please see our information pages on ethics in publishing and ethical guidelines for journal publication. declaration of competing interest all authors must disclose any financial and personal relationships with other people or organizations that could inappropriately influence (bias) their work. examples of potential conflicts of interest include employment, consultancies, stock ownership, honoraria, paid expert testimony, patent applications/ registrations, and grants or other funding. authors should complete the declaration of competing interest statement using this template and upload to the submission system at the attach/upload files step. note: please do not convert the .docx template to another file type. author signatures are not required. if there are no interests to declare, please choose the first option in the template. this statement will be published within the article if accepted. more information. submission declaration and verification submission of an article implies that the work described has not been published previously (except in the form of an abstract, a published lecture or academic thesis, see 'multiple, redundant or concurrent publication' for more information), that it is not under consideration for publication elsewhere, that its publication is approved by all authors and tacitly or explicitly by the responsible authorities where the work was carried out, and that, if accepted, it will not be published elsewhere in the same form, in english or in any other language, including electronically without the written consent of the copyright- holder. to verify originality, your article may be checked by the originality detection service crossref similarity check. preprints please note that preprints can be shared anywhere at any time, in line with elsevier's sharing policy. sharing your preprints e.g. on a preprint server will not count as prior publication (see 'multiple, redundant or concurrent publication' for more information). use of inclusive language inclusive language acknowledges diversity, conveys respect to all people, is sensitive to differences, and promotes equal opportunities. content should make no assumptions about the beliefs or commitments of any reader; contain nothing which might imply that one individual is superior to another on the grounds of age, gender, race, ethnicity, culture, sexual orientation, disability or health condition; and use inclusive language throughout. authors should ensure that writing is free from bias, stereotypes, slang, reference to dominant culture and/or cultural assumptions. we advise to seek gender neutrality by using plural nouns ("clinicians, patients/clients") as default/wherever possible to avoid using "he, she," or "he/she." we recommend avoiding the use of descriptors that refer to personal attributes such as age, gender, race, ethnicity, culture, sexual orientation, disability or health condition unless they are relevant and valid. these guidelines are meant as a point of reference to help identify appropriate language but are by no means exhaustive or definitive. please note that changes to the list of contributors are not permitted after the article has been accepted. authors' affiliations must be the institutions where the research presented in the article took place. please note that changes to the author affiliations are not permitted once the corrected proof is published online. author contributions for transparency, we encourage authors to submit an author statement file outlining their individual contributions to the paper using the relevant credit roles: conceptualization; data curation; formal analysis; funding acquisition; investigation; methodology; project administration; resources; software; supervision; validation; visualization; roles/writing - original draft; writing - review & editing. authorship statements should be formatted with the names of authors first and credit role(s) following. more details and an example changes to authorship authors are expected to consider carefully the list and order of authors before submitting their manuscript and provide the definitive list of authors at the time of the original submission. any addition, deletion or rearrangement of author names in the authorship list should be made only https://service.elsevier.com/app/home/supporthub/publishing/ https://www.elsevier.com/about/policies/publishing-ethics https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/declaration-of-competing-interests https://service.elsevier.com/app/answers/detail/a_id/ /supporthub/publishing/ https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/editors/perk/plagiarism-complaints/plagiarism-detection https://www.elsevier.com/editors/perk/plagiarism-complaints/plagiarism-detection https://www.elsevier.com/about/policies/sharing/preprint https://www.elsevier.com/about/policies/sharing https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/authors/journal-authors/policies-and-ethics/credit-author-statement author information pack apr www.elsevier.com/locate/eswa before the manuscript has been accepted and only if approved by the journal editor. to request such a change, the editor must receive the following from the corresponding author: (a) the reason for the change in author list and (b) written confirmation (e-mail, letter) from all authors that they agree with the addition, removal or rearrangement. in the case of addition or removal of authors, this includes confirmation from the author being added or removed. only in exceptional circumstances will the editor consider the addition, deletion or rearrangement of authors after the manuscript has been accepted. while the editor considers the request, publication of the manuscript will be suspended. if the manuscript has already been published in an online issue, any requests approved by the editor will result in a corrigendum. article transfer service this journal is part of our article transfer service. this means that if the editor feels your article is more suitable in one of our other participating journals, then you may be asked to consider transferring the article to one of those. if you agree, your article will be transferred automatically on your behalf with no need to reformat. please note that your article will be reviewed again by the new journal. more information. copyright upon acceptance of an article, authors will be asked to complete a 'journal publishing agreement' (see more information on this). an e-mail will be sent to the corresponding author confirming receipt of the manuscript together with a 'journal publishing agreement' form or a link to the online version of this agreement. subscribers may reproduce tables of contents or prepare lists of articles including abstracts for internal circulation within their institutions. permission of the publisher is required for resale or distribution outside the institution and for all other derivative works, including compilations and translations. if excerpts from other copyrighted works are included, the author(s) must obtain written permission from the copyright owners and credit the source(s) in the article. elsevier has preprinted forms for use by authors in these cases. elsevier supports responsible sharing find out how you can share your research published in elsevier journals. role of the funding source you are requested to identify who provided financial support for the conduct of the research and/or preparation of the article and to briefly describe the role of the sponsor(s), if any, in study design; in the collection, analysis and interpretation of data; in the writing of the report; and in the decision to submit the article for publication. if the funding source(s) had no such involvement then this should be stated. open access please visit our open access page for more information. elsevier researcher academy researcher academy is a free e-learning platform designed to support early and mid-career researchers throughout their research journey. the "learn" environment at researcher academy offers several interactive modules, webinars, downloadable guides and resources to guide you through the process of writing for research and going through peer review. feel free to use these free resources to improve your submission and navigate the publication process with ease. language (usage and editing services) please write your text in good english (american or british usage is accepted, but not a mixture of these). authors who feel their english language manuscript may require editing to eliminate possible grammatical or spelling errors and to conform to correct scientific english may wish to use the english language editing service available from elsevier's author services. submission our online submission system guides you stepwise through the process of entering your article details and uploading your files. the system converts your article files to a single pdf file used in the peer-review process. editable files (e.g., word, latex) are required to typeset your article for final publication. all correspondence, including notification of the editor's decision and requests for revision, is sent by e-mail. https://www.elsevier.com/authors/article-transfer-service https://www.elsevier.com/about/policies/copyright https://www.elsevier.com/about/policies/copyright/permissions https://www.elsevier.com/__data/assets/word_doc/ / /permission-request-form.docx https://www.elsevier.com/authors/journal-authors/submit-your-paper/sharing-and-promoting-your-article https://www.elsevier.com/journals/expert-systems-with-applications/ - /open-access-options https://researcheracademy.elsevier.com/ https://webshop.elsevier.com/language-editing-services/language-editing/ https://webshop.elsevier.com/language-editing-services/language-editing/ author information pack apr www.elsevier.com/locate/eswa additonal information papers submitted to expert systems with applications, may also be posted on the computer science preprint server http://www.compscipreprints.com. such posting on the computer science preprint server is in conformity with elsevier copyright policy and in no way conflicts with submission to expert systems with applications. preparation peer review this journal operates a single anonymized review process. all contributions will be initially assessed by the editor for suitability for the journal. papers deemed suitable are then typically sent to a minimum of two independent expert reviewers to assess the scientific quality of the paper. the editor is responsible for the final decision regarding acceptance or rejection of articles. the editor's decision is final. editors are not involved in decisions about papers which they have written themselves or have been written by family members or colleagues or which relate to products or services in which the editor has an interest. any such submission is subject to all of the journal's usual procedures, with peer review handled independently of the relevant editor and their research groups. more information on types of peer review. use of word processing software it is important that the file be saved in the native format of the word processor used. the text should be in single-column format. keep the layout of the text as simple as possible. most formatting codes will be removed and replaced on processing the article. in particular, do not use the word processor's options to justify text or to hyphenate words. however, do use bold face, italics, subscripts, superscripts etc. when preparing tables, if you are using a table grid, use only one grid for each individual table and not a grid for each row. if no grid is used, use tabs, not spaces, to align columns. the electronic text should be prepared in a way very similar to that of conventional manuscripts (see also the guide to publishing with elsevier). note that source files of figures, tables and text graphics will be required whether or not you embed your figures in the text. see also the section on electronic artwork. to avoid unnecessary errors you are strongly advised to use the 'spell-check' and 'grammar-check' functions of your word processor. article structure subdivision - numbered sections divide your article into clearly defined and numbered sections. subsections should be numbered . (then . . , . . , ...), . , etc. (the abstract is not included in section numbering). use this numbering also for internal cross-referencing: do not just refer to 'the text'. any subsection may be given a brief heading. each heading should appear on its own separate line. introduction state the objectives of the work and provide an adequate background, avoiding a detailed literature survey or a summary of the results. material and methods provide sufficient details to allow the work to be reproduced by an independent researcher. methods that are already published should be summarized, and indicated by a reference. if quoting directly from a previously published method, use quotation marks and also cite the source. any modifications to existing methods should also be described. theory/calculation a theory section should extend, not repeat, the background to the article already dealt with in the introduction and lay the foundation for further work. in contrast, a calculation section represents a practical development from a theoretical basis. results results should be clear and concise. discussion this should explore the significance of the results of the work, not repeat them. a combined results and discussion section is often appropriate. avoid extensive citations and discussion of published literature. http://www.compscipreprints.com https://www.elsevier.com/reviewers/what-is-peer-review https://www.elsevier.com/reviewers/what-is-peer-review https://www.elsevier.com/authors/journal-authors/submit-your-paper author information pack apr www.elsevier.com/locate/eswa conclusions the main conclusions of the study may be presented in a short conclusions section, which may stand alone or form a subsection of a discussion or results and discussion section. essential title page information • title. concise and informative. titles are often used in information-retrieval systems. avoid abbreviations and formulae where possible. • author names and affiliations. where the family name may be ambiguous (e.g., a double name), please indicate this clearly. present the authors' affiliation addresses (where the actual work was done) below the names. indicate all affiliations with a lower-case superscript letter immediately after the author's name and in front of the appropriate address. provide the full postal address of each affiliation, including the country name and, the e-mail address of each author. • corresponding author. clearly indicate who will handle correspondence at all stages of refereeing and publication, also post-publication. ensure that phone numbers (with country and area code) are provided in addition to the e-mail address and the complete postal address. contact details must be kept up to date by the corresponding author. • present/permanent address. if an author has moved since the work described in the article was done, or was visiting at the time, a 'present address' (or 'permanent address') may be indicated as a footnote to that author's name. the address at which the author actually did the work must be retained as the main, affiliation address. superscript arabic numerals are used for such footnotes. highlights highlights are optional yet highly encouraged for this journal, as they increase the discoverability of your article via search engines. they consist of a short collection of bullet points that capture the novel results of your research as well as new methods that were used during the study (if any). please have a look at the examples here: example highlights. highlights should be submitted in a separate editable file in the online submission system. please use 'highlights' in the file name and include to bullet points (maximum characters, including spaces, per bullet point). abstract a concise and factual abstract is required. the abstract should state briefly the purpose of the research, the principal results and major conclusions. an abstract is often presented separately from the article, so it must be able to stand alone. for this reason, references should be avoided, but if essential, then cite the author(s) and year(s). also, non-standard or uncommon abbreviations should be avoided, but if essential they must be defined at their first mention in the abstract itself. keywords immediately after the abstract, provide a maximum of keywords, using american spelling and avoiding general and plural terms and multiple concepts (avoid, for example, 'and', 'of'). be sparing with abbreviations: only abbreviations firmly established in the field may be eligible. these keywords will be used for indexing purposes. abbreviations define abbreviations that are not standard in this field in a footnote to be placed on the first page of the article. such abbreviations that are unavoidable in the abstract must be defined at their first mention there, as well as in the footnote. ensure consistency of abbreviations throughout the article. acknowledgements collate acknowledgements in a separate section at the end of the article before the references and do not, therefore, include them on the title page, as a footnote to the title or otherwise. list here those individuals who provided help during the research (e.g., providing language help, writing assistance or proof reading the article, etc.). formatting of funding sources list funding sources in this standard way to facilitate compliance to funder's requirements: funding: this work was supported by the national institutes of health [grant numbers xxxx, yyyy]; the bill & melinda gates foundation, seattle, wa [grant number zzzz]; and the united states institutes of peace [grant number aaaa]. https://www.elsevier.com/authors/journal-authors/highlights author information pack apr www.elsevier.com/locate/eswa it is not necessary to include detailed descriptions on the program or type of grants and awards. when funding is from a block grant or other resources available to a university, college, or other research institution, submit the name of the institute or organization that provided the funding. if no funding has been provided for the research, please include the following sentence: this research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. math formulae please submit math equations as editable text and not as images. present simple formulae in line with normal text where possible and use the solidus (/) instead of a horizontal line for small fractional terms, e.g., x/y. in principle, variables are to be presented in italics. powers of e are often more conveniently denoted by exp. number consecutively any equations that have to be displayed separately from the text (if referred to explicitly in the text). footnotes footnotes should be used sparingly. number them consecutively throughout the article. many word processors can build footnotes into the text, and this feature may be used. otherwise, please indicate the position of footnotes in the text and list the footnotes themselves separately at the end of the article. do not include footnotes in the reference list. artwork electronic artwork general points • make sure you use uniform lettering and sizing of your original artwork. • embed the used fonts if the application provides that option. • aim to use the following fonts in your illustrations: arial, courier, times new roman, symbol, or use fonts that look similar. • number the illustrations according to their sequence in the text. • use a logical naming convention for your artwork files. • provide captions to illustrations separately. • size the illustrations close to the desired dimensions of the published version. • submit each illustration as a separate file. • ensure that color images are accessible to all, including those with impaired color vision. a detailed guide on electronic artwork is available. you are urged to visit this site; some excerpts from the detailed information are given here. formats if your electronic artwork is created in a microsoft office application (word, powerpoint, excel) then please supply 'as is' in the native document format. regardless of the application used other than microsoft office, when your electronic artwork is finalized, please 'save as' or convert the images to one of the following formats (note the resolution requirements for line drawings, halftones, and line/halftone combinations given below): eps (or pdf): vector drawings, embed all used fonts. tiff (or jpeg): color or grayscale photographs (halftones), keep to a minimum of dpi. tiff (or jpeg): bitmapped (pure black & white pixels) line drawings, keep to a minimum of dpi. tiff (or jpeg): combinations bitmapped line/half-tone (color or grayscale), keep to a minimum of dpi. please do not: • supply files that are optimized for screen use (e.g., gif, bmp, pict, wpg); these typically have a low number of pixels and limited set of colors; • supply files that are too low in resolution; • submit graphics that are disproportionately large for the content. color artwork please make sure that artwork files are in an acceptable format (tiff (or jpeg), eps (or pdf) or ms office files) and with the correct resolution. if, together with your accepted article, you submit usable color figures then elsevier will ensure, at no additional charge, that these figures will appear in color online (e.g., sciencedirect and other sites) in addition to color reproduction in print. further information on the preparation of electronic artwork. https://www.elsevier.com/authors/author-schemas/artwork-and-media-instructions https://www.elsevier.com/authors/author-schemas/artwork-and-media-instructions https://www.elsevier.com/authors/author-schemas/artwork-and-media-instructions author information pack apr www.elsevier.com/locate/eswa figure captions ensure that each illustration has a caption. supply captions separately, not attached to the figure. a caption should comprise a brief title (not on the figure itself) and a description of the illustration. keep text in the illustrations themselves to a minimum but explain all symbols and abbreviations used. text graphics text graphics may be embedded in the text at the appropriate position. if you are working with latex and have such features embedded in the text, these can be left. see further under electronic artwork. tables please submit tables as editable text and not as images. tables can be placed either next to the relevant text in the article, or on separate page(s) at the end. number tables consecutively in accordance with their appearance in the text and place any table notes below the table body. be sparing in the use of tables and ensure that the data presented in them do not duplicate results described elsewhere in the article. please avoid using vertical rules and shading in table cells. references citation in text please ensure that every reference cited in the text is also present in the reference list (and vice versa). any references cited in the abstract must be given in full. unpublished results and personal communications are not recommended in the reference list, but may be mentioned in the text. if these references are included in the reference list they should follow the standard reference style of the journal and should include a substitution of the publication date with either 'unpublished results' or 'personal communication'. citation of a reference as 'in press' implies that the item has been accepted for publication. web references as a minimum, the full url should be given and the date when the reference was last accessed. any further information, if known (doi, author names, dates, reference to a source publication, etc.), should also be given. web references can be listed separately (e.g., after the reference list) under a different heading if desired, or can be included in the reference list. data references this journal encourages you to cite underlying or relevant datasets in your manuscript by citing them in your text and including a data reference in your reference list. data references should include the following elements: author name(s), dataset title, data repository, version (where available), year, and global persistent identifier. add [dataset] immediately before the reference so we can properly identify it as a data reference. the [dataset] identifier will not appear in your published article. references in a special issue please ensure that the words 'this issue' are added to any references in the list (and any citations in the text) to other articles in the same special issue. reference management software most elsevier journals have their reference template available in many of the most popular reference management software products. these include all products that support citation style language styles, such as mendeley. using citation plug-ins from these products, authors only need to select the appropriate journal template when preparing their article, after which citations and bibliographies will be automatically formatted in the journal's style. if no template is yet available for this journal, please follow the format of the sample references and citations as shown in this guide. if you use reference management software, please ensure that you remove all field codes before submitting the electronic manuscript. more information on how to remove field codes from different reference management software. users of mendeley desktop can easily install the reference style for this journal by clicking the following link: http://open.mendeley.com/use-citation-style/expert-systems-with-applications when preparing your manuscript, you will then be able to select this style using the mendeley plug- ins for microsoft word or libreoffice. reference style text: citations in the text should follow the referencing style used by the american psychological association. you are referred to the publication manual of the american psychological association, seventh edition, isbn - - - - , copies of which may be ordered online. https://citationstyles.org https://citationstyles.org https://www.mendeley.com/reference-management/reference-manager/ https://service.elsevier.com/app/answers/detail/a_id/ / https://service.elsevier.com/app/answers/detail/a_id/ / https://apastyle.apa.org/products/publication-manual- th-edition author information pack apr www.elsevier.com/locate/eswa list: references should be arranged first alphabetically and then further sorted chronologically if necessary. more than one reference from the same author(s) in the same year must be identified by the letters 'a', 'b', 'c', etc., placed after the year of publication. examples: reference to a journal publication: van der geer, j., hanraads, j. a. j., & lupton, r. a. ( ). the art of writing a scientific article. journal of scientific communications, , – . https://doi.org/ . /j.sc. . . reference to a journal publication with an article number: van der geer, j., hanraads, j. a. j., & lupton, r. a. ( ). the art of writing a scientific article. heliyon, , article e . https://doi.org/ . /j.heliyon. .e . reference to a book: strunk, w., jr., & white, e. b. ( ). the elements of style ( th ed.). longman (chapter ). reference to a chapter in an edited book: mettam, g. r., & adams, l. b. ( ). how to prepare an electronic version of your article. in b. s. jones, & r. z. smith (eds.), introduction to the electronic age (pp. – ). e-publishing inc. reference to a website: powertech systems. ( ). lithium-ion vs lead-acid cost analysis. retrieved from http://www.powertechsystems.eu/home/tech-corner/lithium-ion-vs-lead-acid-cost-analysis/. accessed january , reference to a dataset: [dataset] oguro, m., imahiro, s., saito, s., & nakashizuka, t. ( ). mortality data for japanese oak wilt disease and surrounding forest compositions. mendeley data, v . https://doi.org/ . / xwj nb r. . reference to a conference paper or poster presentation: engle, e.k., cash, t.f., & jarry, j.l. ( , november). the body image behaviours inventory- : development and validation of the body image compulsive actions and body image avoidance scales. poster session presentation at the meeting of the association for behavioural and cognitive therapies, new york, ny. journal abbreviations source journal names should be abbreviated according to the list of title word abbreviations. video elsevier accepts video material and animation sequences to support and enhance your scientific research. authors who have video or animation files that they wish to submit with their article are strongly encouraged to include links to these within the body of the article. this can be done in the same way as a figure or table by referring to the video or animation content and noting in the body text where it should be placed. all submitted files should be properly labeled so that they directly relate to the video file's content. in order to ensure that your video or animation material is directly usable, please provide the file in one of our recommended file formats with a preferred maximum size of mb per file, gb in total. video and animation files supplied will be published online in the electronic version of your article in elsevier web products, including sciencedirect. please supply 'stills' with your files: you can choose any frame from the video or animation or make a separate image. these will be used instead of standard icons and will personalize the link to your video data. for more detailed instructions please visit our video instruction pages. note: since video and animation cannot be embedded in the print version of the journal, please provide text for both the electronic and the print version for the portions of the article that refer to this content. data visualization include interactive data visualizations in your publication and let your readers interact and engage more closely with your research. follow the instructions here to find out about available data visualization options and how to include them with your article. supplementary material supplementary material such as applications, images and sound clips, can be published with your article to enhance it. submitted supplementary items are published exactly as they are received (excel or powerpoint files will appear as such online). please submit your material together with the article and supply a concise, descriptive caption for each supplementary file. if you wish to make changes to supplementary material during any stage of the process, please make sure to provide an updated file. do not annotate any corrections on a previous version. please switch off the 'track changes' option in microsoft office files as these will appear in the published version. https://www.issn.org/services/online-services/access-to-the-ltwa/ https://www.sciencedirect.com https://www.elsevier.com/authors/author-schemas/artwork-and-media-instructions https://www.elsevier.com/authors/tools-and-resources/data-visualization author information pack apr www.elsevier.com/locate/eswa research data this journal encourages and enables you to share data that supports your research publication where appropriate, and enables you to interlink the data with your published articles. research data refers to the results of observations or experimentation that validate research findings. to facilitate reproducibility and data reuse, this journal also encourages you to share your software, code, models, algorithms, protocols, methods and other useful materials related to the project. below are a number of ways in which you can associate data with your article or make a statement about the availability of your data when submitting your manuscript. if you are sharing data in one of these ways, you are encouraged to cite the data in your manuscript and reference list. please refer to the "references" section for more information about data citation. for more information on depositing, sharing and using research data and other relevant research materials, visit the research data page. data linking if you have made your research data available in a data repository, you can link your article directly to the dataset. elsevier collaborates with a number of repositories to link articles on sciencedirect with relevant repositories, giving readers access to underlying data that gives them a better understanding of the research described. there are different ways to link your datasets to your article. when available, you can directly link your dataset to your article by providing the relevant information in the submission system. for more information, visit the database linking page. for supported data repositories a repository banner will automatically appear next to your published article on sciencedirect. in addition, you can link to relevant data or entities through identifiers within the text of your manuscript, using the following format: database: xxxx (e.g., tair: at g ; ccdc: ; pdb: xfn). mendeley data this journal supports mendeley data, enabling you to deposit any research data (including raw and processed data, video, code, software, algorithms, protocols, and methods) associated with your manuscript in a free-to-use, open access repository. during the submission process, after uploading your manuscript, you will have the opportunity to upload your relevant datasets directly to mendeley data. the datasets will be listed and directly accessible to readers next to your published article online. for more information, visit the mendeley data for journals page. data in brief you have the option of converting any or all parts of your supplementary or additional raw data into a data article published in data in brief. a data article is a new kind of article that ensures that your data are actively reviewed, curated, formatted, indexed, given a doi and made publicly available to all upon publication (watch this video describing the benefits of publishing your data in data in brief). you are encouraged to submit your data article for data in brief as an additional item directly alongside the revised version of your manuscript. if your research article is accepted, your data article will automatically be transferred over to data in brief where it will be editorially reviewed, published open access and linked to your research article on sciencedirect. please note an open access fee is payable for publication in data in brief. full details can be found on the data in brief website. please use this template to write your data in brief data article. methodsx you have the option of converting relevant protocols and methods into one or multiple methodsx articles, a new kind of article that describes the details of customized research methods. many researchers spend a significant amount of time on developing methods to fit their specific needs or setting, but often without getting credit for this part of their work. methodsx, an open access journal, now publishes this information in order to make it searchable, peer reviewed, citable and reproducible. authors are encouraged to submit their methodsx article as an additional item directly alongside the revised version of their manuscript. if your research article is accepted, your methods article will automatically be transferred over to methodsx where it will be editorially reviewed. please note an open access fee is payable for publication in methodsx. full details can be found on the methodsx website. please use this template to prepare your methodsx article. https://www.elsevier.com/authors/tools-and-resources/research-data https://www.elsevier.com/authors/tools-and-resources/research-data/data-base-linking https://www.elsevier.com/authors/tools-and-resources/research-data/data-base-linking#repositories https://www.elsevier.com/authors/tools-and-resources/research-data/mendeley-data-for-journals https://www.journals.elsevier.com/data-in-brief/about-data-in-brief/video-discover-the-benefits-of-publishing-your-research-data https://www.elsevier.com/journals/data-in-brief/ - /open-access-journal https://www.journals.elsevier.com/data-in-brief https://www.elsevier.com/__data/assets/word_doc/ / /datainbrief_template.docx https://www.journals.elsevier.com/methodsx https://www.journals.elsevier.com/methodsx https://www.elsevier.com/__data/assets/word_doc/ / /methodsx-article-template.docx author information pack apr www.elsevier.com/locate/eswa data statement to foster transparency, we encourage you to state the availability of your data in your submission. this may be a requirement of your funding body or institution. if your data is unavailable to access or unsuitable to post, you will have the opportunity to indicate why during the submission process, for example by stating that the research data is confidential. the statement will appear with your published article on sciencedirect. for more information, visit the data statement page. after acceptance online proof correction to ensure a fast publication process of the article, we kindly ask authors to provide us with their proof corrections within two days. corresponding authors will receive an e-mail with a link to our online proofing system, allowing annotation and correction of proofs online. the environment is similar to ms word: in addition to editing text, you can also comment on figures/tables and answer questions from the copy editor. web-based proofing provides a faster and less error-prone process by allowing you to directly type your corrections, eliminating the potential introduction of errors. if preferred, you can still choose to annotate and upload your edits on the pdf version. all instructions for proofing will be given in the e-mail we send to authors, including alternative methods to the online version and pdf. we will do everything possible to get your article published quickly and accurately. please use this proof only for checking the typesetting, editing, completeness and correctness of the text, tables and figures. significant changes to the article as accepted for publication will only be considered at this stage with permission from the editor. it is important to ensure that all corrections are sent back to us in one communication. please check carefully before replying, as inclusion of any subsequent corrections cannot be guaranteed. proofreading is solely your responsibility. offprints the corresponding author will, at no cost, receive a customized share link providing days free access to the final published version of the article on sciencedirect. the share link can be used for sharing the article via any communication channel, including email and social media. for an extra charge, paper offprints can be ordered via the offprint order form which is sent once the article is accepted for publication. both corresponding and co-authors may order offprints at any time via elsevier's author services. corresponding authors who have published their article gold open access do not receive a share link as their final published version of the article is available open access on sciencedirect and can be shared through the article doi link. author inquiries visit the elsevier support center to find the answers you need. here you will find everything from frequently asked questions to ways to get in touch. you can also check the status of your submitted article or find out when your accepted article will be published. © copyright elsevier | https://www.elsevier.com https://www.elsevier.com/authors/tools-and-resources/research-data/data-statement https://www.elsevier.com/authors/journal-authors/submit-your-paper/sharing-and-promoting-your-article/share-link https://www.sciencedirect.com/ https://webshop.elsevier.com/article-services/article-offprints/ https://service.elsevier.com/app/home/supporthub/publishing https://service.elsevier.com/app/answers/detail/a_id/ /supporthub/publishing/kw/status+submitted+article/ https://service.elsevier.com/app/answers/detail/a_id/ /kw/ /p/ /supporthub/publishing https://service.elsevier.com/app/answers/detail/a_id/ /kw/ /p/ /supporthub/publishing document resume ed cs author willson, victor l. title methodological limitations of the application of expert systems methodology in reading. pub date dec note p.; paper presented at the annual meeting of the national reading conference ( th, austin, tx, december - , ). pub type information analyses ( ) -- speeches/conference papers ( ) edrs price mf /pc plus postage. descriptors evaluation problems; *reading research; *research methodology; *research problems; *research utilization; *theory practice relationship abstract methodological deficiencies inherent in expert-novice reading research make it impossible to draw inferences about curriculum change. first, comparisons of intact groups are often used as a basis for making causal inferences about how observed characteristics affect behaviors. while comparing different groups is not by itself a useless activity, progressing directly to training is premature at best. second, the think-aloud protocol technique is often used for inferring a subject's cognitive structure of subject matter. this method is inappropriate because it assumes that the organization of this structure resides consciously in a person's mind and can be verbally reproduced. third, retrospective methods have been employed to infer causality by selecting groups currently differing and discovering differences in their past on putative causal variables, which are then inferred to have caused the present differences. while this technique must be used in historical analyses, it becomes suspect when the inferences are used to speculate on implications for current practice. finally, techniques employed in naturalistic inquiry often confuse a change in methodology with a change in the discipline being studied, and rely heavily on impressionistic, one-shot observation for many facts. (jd) **********************************************************************x reproductions supplied by edrs are the best that can be made from the original document. ********************************************************************** u.s. department of education office of educational research and improvement educational resources information center (eric) $ this document has been reproduced as received from the person or organization originating it. c minor changes have been made to improve reproduction quality. points of view or opinions stated in this dap- ment do not necessarily represent official oeri position or policy. methodological limitations of the application of expert systems methodology in reading victor l. willson texas a & m university "permission to reprodijce this material has been granted by victor l. willson to the educational resources informal ion center (eric)." paper presented at the annual meeting of the national reading conference, austin tx, december methodological limitations of the application of expert systems methodology in reading the comparison between the expert and inexpert, whether in organizations or in humans, has become a major research technique in education in the last decade. in research on problem solving comparisons have been made between physicists and physics stu- dents. in reading research the basic comparison is between good and poor readers for much of the information-processing theory being advanced. in educational administration successful schools are compared with unsuccessful schools. creative students are compared with normal students. in all of these areas a similar paradigm is being used: the expert system is compared on a number of attributes with a less or inexpert system. some attributes are assumed to be causal (such as program differences or strategies employed) and others are assumed to be outcomes, such as achieve- ment or time to solution of a problem. differences between the expert and inexpert on the causal variables are then assumed to be evidence for causation of the variables, and these salient variables are prcmoted as efficacious for remedying the deficien- cies of the inexpert. this paradigm is an important departure from the dominant experimental paradigm used in education, and this paper presents a critical examination of its methodological limitations. it is an intaresting situation that the expertise methodology has sprung from two quite different areas of research, cognitive psychology and curriculum behaviorism. cogni- tive psychology has drawn from and has itself influenced artifi- cial intelligence (ai) research in machine computing. as chi, feltovitch & glaser ( ) has noted a shift took place in ai research from power strategies to knowledge strategies. the best knowledge strategies available for stuay were found in human beings, so that new computer programs were developed to imitate the way human experts organized and processed information, as could be determined in comparisons with inexpert humans. the effective schools movement, not directly influenced by artificial intelligence in any obvious way, sought to study the school environment. one outcome of such study, conducted by anthropolo- gists, psychologists, and educational researchers, was that some schools (classrooms, teachers, administrators, etc.) were super- ior in performance to others. comparisons between variables de- fined as input, or causal,' and output or dependent, led to pre- scriptions for change in inexpert schools to make them more like the best schools being observed. for example clark & mccarthy ( ) reported on a cohort sequential type design in which volunteer new york city schools implemented a new program based on effective schools literature. since most of the expert-novice comparisons were presented in the methodological trappings of experimental research (anova statistical analysis and interpretation) there have been only one or two serious evaluations of the causal logic underlying the method and its basis for making causal inferences (rowan, bossert dwyer, ). the main thesis presented here is that all re- search applying this method suffers from internal validity flaws sufficiently serious to render it uninterpretable. furthermore, expertise method is incapable of supporting causal inference regarding change in the inexpert without true experimental re- search. techniques employed the techniques being drawn upon in research on expertise include, but are not limited to, the following: comparisons between inact groups; think-aloud protocol; naturalistic inquiry, including ethnographic field method; and retrospective research. there are researchers who are employing experimental research as part of their research strategies, and their applications, specifically exempted from the criticism levelled here, will be mentioned as exemplars of appropriate or adequate research. comparison between intact groups. this technique is widely used in expertise research. in problem-solving research chase & simon ( ) compared the ability of chess masters and novices to chunk board groupings. they found more elements in the chunks of mas- ters than in the novices. simon & simon ( ) found differences in the problem-solving behaviors of physicists and physics stu- dents using verbal protocols of the tasks each performed when solving novel problems. the effective schools movement has used comparisons between schools defined as outstanding or excellent and those defined as inferior or deficient to make programmatic decisions about how schools ought to be run. a well-constructed criticism of the effective schools reseat-ch methodology was made by rowan, bossert, & dwyer ( ). their specific points will be incorporated into this review; these points include difficulty with causal ordering, instrumentation, limitations of generaliza- tion, and nonequivalent control group comparisons. the effective schools research of the u's was employed in examining reading at both the school and classroom level. teacher effectiveness has been particularly emphasized (rupley, wise, & logan, ). brophy's ( ) work on process-product research with primary grade teachers is a widely cited example; later important studies include medley ( ) and rosenshine ( ); the latter study raised the problem of little experimental veri- fication for effectiveness research. the stanford program on teaching effectiveness (crawford, gage, corno, staybrook, mitman, schunk, stallings, baskin, hanvey, austin & newman, ), and the first grade reading group study (anderson, evertson & brophy, ) are experimental or quasi-experimental studies based on initial observations contrasting good and poor teachers (rupley et al, ). it is important to note that in these studies curriculum recommendations were made after comparative interven- tion was made, not directly on the basis of the original compari- sons. much of the recent research on reading from a cognitive perspective is based on comparisons between good and poor read- ers. for example, in a recent article by underwood & zola ( ) good and poor readers were compared on letter recognition span. in this study no differences were found and no particular instructional inferences were made. in other studies this has not been the case: mcgee ( ) compared good and poor fifth grade readers and poor third grade readers, finding differences in recall of text structure ordered from good to poor fifth graders to third graders. mcgee concluded that young readers "benefit from following the top-level structure of text to guide reading and remembering passage information..." even though a disclaimer below this quote suggests the need for more research on efective- ness of instruction, there is a clear message that the observed difference.a are caused by what good readers do, and that poor readers will be helped by some strategy based on the good read- ers' processes. while the study is itself limited because of the text reading level (third grade), it is part of a chain of re- search related to automaticity (laberge & samuels, ) which is itself based in part on these same good-poor reader differences. there is simply no basis for assuming that the poor readers can be made to perform like the good readers or that their processing will oecome automatic, or if automatic in the same way that the good readers' process is automatic. another such study is due to sannomiya ( ) in which poor third grade comprehenders were compared with good sixth grade comprehenders on text comprehen- sion under auditory or visual conditions. in this study both age and ability are confounded. again, there is no evidence that the poor comprehenders can be made to look like good sixth graders, or that different modes of presentation will change reading performance in this direction. intact groups are also used as the basis for inferring developmental change. for example, baldwin & coady ( ) com- pared fifth graders and college students on their use of punctua- tion as clues to meaning in isolated sentences. they found dif- ferences between the groups and inferred developmental differ- ences in use of punctuation as clues to meaning. it is common to see studies that mix age and reading ability. juel ( ) com- pared grade two, grade five, and upper division undergraduates; good and poor readers were identified for the elementary groups. in this study the word adult ' eis used interchangeably with the college sample, the implication being that these readers are a norm for adult performance. this assumption is most definitely wrong, and an assumption that the elementary students are likely to or can become like these adults is unwarranted. juel never- theless suggests that presenting children with practice words with similar letter combinations would help to develop versatil- ity in decoding. that may be true, but the comparisons made in her study do not support such conclusions. there are many other child-adult comparisons in recent literature in which the adults are high ability college students (mcgee, ; schwartz, ; taylor, ), or secondary students (fletcher, satz, & scholes, ). the use of intact groups has been repeatedly criticized in the educational research methodology literature from campbell & stanley ( ) onward with respect to the inference of causality for observed characteristics affecting behaviors. in the case of good versus poor readers, the inference is that what good readers do, poor readers can do, and that instruction directly oriented toward the discrepancy will remediate deficiencies in the poor readers. the good readers are the experts, and the poor readers the novices.' the critical assumption is that the good readers were themselves in the poor readers' state at some point. often, since the two groups are age matched, this is not true. the good readers were never like the poor readers. consequently, the inference that the observed differences in condition can lead to training is by itself without basis. similarly, developmental studies are susceptible to the same difficulty, especially when they involve elementary, secondary, and college populations. in the baldwin & coady ( ) study a comparison between fifth grade and college students is meaningless, because differences may be due to selection: if one were able to select the fifth graders who will eventually qo to college, would we still see the dif- ferences in use of punctuation clues? even if we did find the differences, how cnmfortable would we be in ignoring any other differences that remain between the prospective college-bound fifth graders and the college students. any variables upon which tne two groups differ become possible alternative causal vari- ables, and training in the absense of experimental demonstration is merely yuesswork. a similar problem exists for comparisons with secondary students when dropout rate becomes appreciable (after grade ), or when students begin self-selecting into courses (grade ). differential maturation and history are other threats from the campbell & stanley list which are relevant. finally, regression threats due to selection of extremes are not only omnipresent in good-poor comparisons, their effects should always be estimated statistically just to provide a comparison with the observed differences. comparing different groups is not by itself a useless activity, but progressing directly to training is premature at best. differences between good and poor readers, or between developmentally different groups of readers, is useful for supplying clues or hints for more careful investigation. the tendency to assume a causal shortcut, permitting the ignoring of experimentation, is unfortunate; while the technique may prove correct in a few instances, our experience in educational research with intact groups is lengthy enough to predict many erroneous conclusions and wasted resourc,_ s if the method is allowed to predominate. similarly, research on developmental differences has largely opted for cross-sectional designs, not wishing to do the hard research implicit in true longitudinal study of development. in the good reader-poor reader research this is particularly telling, for we have little data on long term development of either group from a cognitive, information processing theoretic perspective. 'his is the causal ordering problem that rowan et al ( ) pointed out; cross-sectional designs that substitute for longitudinal designs almost always have this difficulty. think-aloud protocol. this technique has tieen used in the study of expert and novice organization of knowledge and was eloquently and favorably defended by ericsson & simon ( ) as a valid means to record information that humans are attendlng to in short-term memory. it was attacked by phillips ( ) as an inappropriate technique to infer human's cognitive structure of subject matter. the core of phillips' argument is that the external organization imposed in the learning required for a task may require a person to reproduce it verbally, but there is no evidence that that organization resides internally in the per- son's mind. similarly, the content and organization of a ques- tion, with perhaps the exception of free response, imposes an organization on the su'jject's response that does not necessarily mirror the internal representation of the response. the use of think aloud method, while it occurs in a variety of research contexts, is a major technique in naturalistic or ethnographic research. a recent study by nicholson ( ) in which minutes of inerviewing with junior high students was conducted is an example in point. this study will be examined in more detail below, but interview techniques in the comparison of experts and novices are likely to suffer from many difficulties. in reading it is particularly problematical because the researchers usually share the same culture (reading, education, etc.) as both the experts and the novices. this is usually a drawback for ethnogra- phers, who are attempting to view the culture with fresh eyes. in nicholson's work the experts were teachers, and the similarity between researcher and expert was far greater than between re- searcher and novices (teenagers). the commonness of a shared language of educationese is quite troublesome for a researher in such conditions and the trustworthiness of such interviewing must be questioned; it is not that interviewing cannot be done well, it is that areat care must be taken to support the evidence presented in such a context. retrospective studies. in research on creativity the compari- son between creative and noncreative individuals has led to the formulation of programs to teach creativity (van tassel-baska, ). also, researchers on creativity have employed retrospec- tive methods to examine prior differences between more and less creative individuals and then to propose ch-nges in education which are expected to engender the same effects in young students as were observed in the creative adults. segal, busse & mansfield ( ) compared retrospectively two groups of biologists, highly cited and nonhighly cited, using self-report survey technique. they found post-doctoral productivity to be related to pre doctoral productivity and high school science interest. as noted by example aove this research technique is used to infer causality by selecting groups currently differing and dis- covering differences in their past en putative causal variables, which are then inferred to have caused the present differences. this technique is apparently not used search, for a search over the last study, by castagna ( ) in which a influential persons in western history and autobiographies. the implication changed these people and that some was very much in reading re- ten years found only one historical examination of was made using biographies is that decisive reading purposeful, some was not. of course, historical analyses must use such methods; it is only if an implication for current practice is made that the analysis becomes suspect. naturalistic inquiry. this body of techniques, attempting to become a method in educational research ,in kaplan's ( ) sense, draws upon ethnographic research from cultural anthropology, but then leaves it in a philosophical sense. recent apologies by harste (undated) and weaver ( b) liken the use of naturalisatic inquiry to a paradigm shift, citing kuhn's now dated and largely refuted work ( ). while this debate more properly belongs in a different critical paper, the use of the techniques in the expert-novice studies requires a small aside. the appeal to a paradigm shift has been misunderstood and mislaid to boot. the shift occurred in psychology in the late s and is often tied to neisser's ( ) resurrection of internal mental representational constructs, the shift being away from behaviorism. this paradigm shift has flowed into educational research rapidly and convincingly, predating the widespread interest in ethnographic techniques by a decade. the latter interest, it is presumed, was an outgrowth of the real paradigm shift. paradigm shifts occur in disciplines w.en the prevailing theories are overturned by new, revolutionary ones, that nevertheless account for the facts and relationships previously learned. in paradigm shifts the old is not discarded, it is reinterpreted. there is no such change occurring in reading, notwithstanding the wishful thinking of harste (undated). the mistake is in confusing a change in methodology with a change in the discipline. methodologies cannot and never will drive disciplines to the extent that the naturalistic inquirers maintain that they do; recent arguments by kuhn ( ) himself have backpedalled on the theory-ladenness argument of data. cooke & campbell ( ) attack the emphasis by philosophers of science on the preeminence of theory, relegating,facts to an unwarranted secondary status. that is, facts ar'e observed by resrearchers working from different methodological perspectives. they must reconcile them; their methodologies become more suspect than the facts, which are interobserver confirmable. if the facts are not confirmable, then they cannot be admitted. this latter issue becomes the main problem for the naturalistic researchers, for they rely heavily on impressionistic, one-shot observation for many facts. many researchers using this method deny intersubjective confirmability, but they abandon science for art. they are not wrong, they merely inquire in another domain. a number of naturalistic studies in reading have been published in the last few years. the study by nicholson ( ) is the primary study i have encountered which purported to compare experts and novices. the study actually examined the structures of teenagers' understanding of' classroom material; teachers were apparently ignored, although there is an appeal to teachers as experts at athe end of the study. a small section on low achievers was also tacked on. the catchy title was misleading or there was a serious editing problem because there was no comparaison between experts and novices in this study. if there had been it would have told us nothing about how to change stu- dents' conceptions. this is a common problem in naturalistic studies. one gets whatever one happens to find in the setting. if there is nothing very interesting going on little of use will be brought out. also, naturalistic studies are limited by what passes for actual practice, not by what is possible. it is quite possible that most of what will occur in education in the next century is being tested in laboratory schools, industrial set- tings, and nontraditional educational locations. the public schools are likely to be the last places to find out about these changes, whether through experimental or noturalistic means. naturalistic research on expert-novice differences in read- ing is limited by selection, ie. the choice of locations; hj history, the context of the location; by instrumentation, espec- ially changes in the observer/interviewer; and by te.voral limi- tations in when the study is conducted and for how long. it is not argued here that naturalistic inquiry is less appropriate than the quasi-experimental research described earlier. neither is likely to be able to draw valid conclusions regarding curriculum change in the absense of careful experimental manipulation of variables. summary this paper has sought to draw attention to methodological deficiencies inherent in expert-novice research with respect to drawing inferences about curriculum change. much credit must be given to the reading research community for generally not leaping to conclusions from such literature, in comparison with some fields of psychology, engineering, and.science education. while some reading studies seem to overreach their conclusions, far more have used the observed differences to probe experimentally hypotheses generated by the observations. this approach cannot be faulted, even if one cannot resist challenging the original premise: that good readers can tell us anything about how poor readers ought to proceed. the methodological threats to internal validity of such research ventures should make us pause to consider if good-poor or expert-novice comparisons are really of value: history, selection, instrumentation, maturation, and regression. while no study necessarily is damned due to possible internal invalidity thrats, the weight of methodological argument certainly should make us pause. ex post facto methods, such as meta analysis, can never rectify the poor initial choice of field of explorat ion. if we want to see how poor readers can be made into good readers we ought to find examples, or better yet, create examples, and then work to find out what is replicable. that is good science and good research. references anderson, l.m., evertson, c.m. & brophy, j.e. ( ). the first grade reading group study: technical report of experimen- tal effects and process-outcome relationships, report no. . austin:research and development center for teacher education, university of texas. baldwin,r.s. & coady, j.m. ( ). psycholingusitic approaches to a theory of punctuation. journal of reading behavior, , - . brophy, j.e. ( ). stability of teacher effectiveness. american educational research journa , , - . campbell,'d.t. & stanley, j.c. ( ).experimental and guasi- experimental designs for research on teaching. chicago: rand-mcnally. castagna, e. ( ).caught in the act: the decisive readinl of some notable men and women and its influence on their actions and attitudes.metuchen,nj: the scarecrow press. chase w.g. & simon, h. a. ( ). perception in chess. cognitive psychology, , - . chi, m.t.h., feltovitch, p.j. & glaser, r. ( ). categorization and representation of physics problems by experts and nov- ices.cognitive science, , - . clark, t.a. & mccarthy, d.p. ( ). school improvement in new york city: the evolution of a project. educational re- searcher , - . crawford, j., gage, n., corno, l., staybrook, n., mitman, a., schunk, d., stallings, j., baskin, e., hanvey, p., austin, d., & newman,. r. an experiment on teacher effectiveness and parent assisted instruction in the third gradet. stanford, ca: center for educational research at stanford, stanford university. cook, t.d. & campbell, d.t. ( ). quasi-experimentation: _design and analysis issues for field settings. chicago:rand- mcnally. ericsson, k.a. & simon, h.a. ( u). verbal reports as data. review, , - . fletcher,j.m., satz, p., & scholes,r.j. ( ). developmen- tal chulges in the linguistic performance correlates of reading achievement. brain and language, , - . harste, j.c. (undated). portrait of'a new paradigm: reading comprehension research. in a.crismore (ed.), land- scapes:, a state-of-the-art assessment of reading compre- hension research - . final report, usde-c- - - . bloomington,in:indiana university language edu- cation departments. joel, c. ( ). the development and use of mediated word identification. reading research quarterly, , - . kuhn, t.s. ( ). the structure of scientific revolutions. chicago: university of chicago press. laberge, d. & samuels, s. ( ). toward a theory of automatic information processing in reading. cognitive psychology, , - . mcgee, l.m. ( ). awareness of text structure: effects on children's recall of expository text. readina research quarterly, , - u. medley, d.m. ( ). teacher competency and teacher effective- ness: a review of process:product research. washington, d.c.: aamerican association of colleges of teacher educa- tion. neisser, u. ( ). coanitive psychology. new york: appleton- century-crofts. nicholson,t. ( ). experts and novices: a study of reading in the high school classroom. reading research quarterly, . - . phillips, d.c. ( ). on describing a student's cognitive struc- ture. educational sysi., - . rowan, b., bossert, s.t. & dwyer, d.c. ( ). research on effective schools: a cautionary note.educational researcher, , - . rosenshine, b.v. ( ). instructional principles in direct in- struction. paper presented at the annual meeting of the american educational research association, toronto, march, . rupley, w.h., wise, b.s., & logan, j.w. ( ). research in effective teaching: an overview of its development. in j.v. hoffman (ed.), effective teaching of readino: research and practice. newark,del: international reading association. sannomiya, m. ( ). modality effects on text processing as a function of ability to comprehend. perceptual_ and motor skills - . schwartz, r.m. ( ). levels of processing: the strategic de- mands of reading cmprehension. reading research quar- terly, , - . segal, s.m., busse, t.v. & mansfield, r.s. ( ). the relationship of scientific creativity in the biological sciences to predoctoral accomplishments and experience. american educational research journal , - . simon, d.p. & simon, h.a. ( ). individual differences in solving physics problems. in r.siegler (ed.), children's thinking: what develops?. hillsdale,nj: lawrence erlbaum. taylor, b.m. ( ).children's memory for expository text after reading. reading research quarterly, , - . underwood, n.r. & zola, d. ( ). the span of letter recognition of good and poor readers. reading research quarterly, , - . van tassel-baska, j. ( ). effective curriculum and instructional models for talented students. gifted child quarterly, , - . weaver, c. ( ). parallels between new paradigms in science and in reading and literary theories: an essay review. research in the teaching of english, , - . poe.dvi genetic wavelet packets for speech recognition leandro d. vignolo∗, diego h. milone, hugo l. rufiner research center for signals, systems and computational intelligence, departamento de informática, facultad de ingenieŕıa y ciencias hı́dricas, universidad nacional del litoral, conicet, argentina abstract the most widely used speech representation is based on the mel-frequency cepstral coefficients, which incorporates biologically inspired characteristics into artificial recognizers. however, the recognition performance with these features can still be enhanced, specially in adverse conditions. recent ad- vances have been made with the introduction of wavelet based representations for different kinds of signals, which have shown to improve the classification performance. however, the problem of finding an adequate wavelet based representation for a particular problem is still an important challenge. in this work we propose a genetic algorithm to evolve a speech representation, based on a non-orthogonal wavelet decomposition, for phoneme classification. the results, obtained for a set of spanish phonemes, show that the proposed genetic algorithm is able to find a representation that improves speech recog- nition results. moreover, the optimized representation was evaluated in noise conditions. key words: phoneme classification, genetic algorithms, wrappers, wavelet packets ∗corresponding author. research center for signals, systems and computational intelligence, departamento de informática, facultad de ingenieŕıa y ciencias hı́dricas, universidad nacional del litoral, ciudad universitaria cc , ruta nacional no km . , te: + ( ) ext , fax: + ( ) , santa fe ( ), argentina. email address: ldvignolo@fich.unl.edu.ar (leandro d. vignolo) url: http://fich.unl.edu.ar/sinc (leandro d. vignolo) preprint submitted to expert systems with applications september , si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . . introduction automatic speech recognition systems need a pre-processing stage to make phoneme key-features more evident, in order to obtain significant im- provements in the classification results [ ]. this task was first addressed by signal processing techniques like filter-banks, linear prediction coding and cepstrum analysis [ ]. the most popular feature representation currently used for speech recognition is built from the mel-frequency cepstral coeffi- cients (mfcc) [ ], which are based on a linear model of voice production together with the codification on a psycho-acoustic scale. however, due to the degradation of recognition performance in the presence of additive noise, many advances have been conducted in the development of alternative fea- ture extraction approaches. in particular, techniques like perceptual linear prediction [ ] and relative spectra [ ] incorporate features based on the hu- man auditory system and provides some robustness in asr. also, significant progress has been made with the application of different artificial intelli- gence techniques in the field of speech processing [ ]. besides, the utilization of wavelet based analysis for speech feature extraction has recently been studied [ , , , ]. the multi-resolution analysis associated with discrete wavelet transform (dwt) can be implemented as a filter bank decomposition (or filter bank schemes) [ ]. wavelet packet transform (wpt) is a generalization of the dwt decomposition which offers a wider range of possibilities for signal rep- resentation in the time-scale plane [ ]. to obtain a representation based on this transform, usually, a particular orthogonal basis is selected among all the available basis. nevertheless, in phoneme classification applications there is not evidential benefit on working with orthogonal basis. moreover, it is known that the analysis performed at the level of the auditory cortex is highly redundant and, therefore, non-orthogonal [ ]. without this restric- tion the result of the full wpt decomposition is a highly redundant set of coefficients, from which a convenient representation for the problem in hand can be selected. many approaches addressing the optimization of wavelet decompositions for feature extraction have been proposed. for instance, in [ ] an automatic extraction of high quality features from continuous wavelet coefficients ac- cording to signal classification criteria was presented. in [ ], an approach based on the best basis wavelet packet entropy method was proposed for elec- troencephalogram classification. also, a method for the selection of wavelet si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . family and parameters was proposed for the phoneme recognition task [ ]. similarly, the use of wavelet based decompositions has also been proposed as a tool for the development of robust features for speaker recognition [ , ]. another interesting approach was proposed in [ ], in which a novel approach for generating the wavelet that maximizes the discrimination capability of ecg beats using particle swarm optimization. also, the use of evolutionary computation techniques in order to optimize over-complete decompositions for signal approximation was proposed in [ ]. furthermore, the use of a genetic algorithm to optimize wpt based features for pathology detection from speech was proposed in [ ], where an entropy criterion was minimized for the selection of the wavelet tree nodes. similar approaches propose the optimization of wavelet decomposition schemes using evolutionary compu- tation for denoising [ , ] and image compression [ ]. besides, different approaches have been proposed for the optimization of wavelet based repre- sentations using swarm intelligence [ , ]. many other studies also rely on evolutionary algorithms for feature selection [ , , ] and the optimization of speech representations [ , , ]. however, the flexibility provided by the full wpt decomposition has not yet been fully exploited in the search for a set of features to improve speech recognition results. when this search is not restricted to non-redundant representations, there is a large number of non-orthogonal dictionaries to be explored, leading to a hard combinatorial problem. here we propose a new approach to optimize over-complete decomposi- tions from a wpt dictionary, which consists in the use of a genetic algorithm (ga) for the selection of wavelet based features. in order to evaluate the so- lutions during the search, the ga uses a learning vector quantization (lvq) classifier. some preliminary results with this strategy were presented in [ ]. the methodology, referred to as genetic wavelet packets (gwp), relies on the benefits provided by evolutionary computation to find a better signal representation. this feature selection scheme, known as wrapper [ , ], is widely used as it allows to obtain the good solutions in comparison with other techniques [ ]. the organization of this paper is as follows. in section , brief descriptions of the properties of wpt and ga are presented. next, our wrapper method for the selection of the wpt components is described. the following section discusses the obtained recognition results for a set of spanish phonemes. finally, the general conclusions and future work are presented. si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . . materials and methods . . wavelet and wavelet packet transforms in contrast with sine and cosine bases, wavelet bases are simultaneously localized in time and frequency. this feature is particularly interesting in the case of signals which present both stationary and transient behaviors. wavelets can be defined, in a simplified manner, as a function of zero mean, unitary norm and centered in the neighborhood of t = [ ]: ψ(t) ∈ l (r); ∫ ∞ −∞ ψ(t)dt = ; ‖ψ(t)‖ = . ( ) a family of time-frequency atoms is obtained by scaling and translating the wavelet function: ψu,s(t) = √ s ψ ( t − u s ) , ( ) with u,s ∈ r. this way, the continuous wavelet transform of the signal x(t) is defined as the inner product with this family of atoms wx (u,s) = 〈x(t),ψu,s(t)〉 = +∞ ∫ −∞ x(t) ψ∗u,s (t) dt. ( ) the discrete dyadic wavelet transform (dwt) of x[n] ∈ rn is obtained by discretizing translation and scaling parameters in ( ), as u = m and s = j . a fast implementation of the dwt based multiresolution analysis exists [ ], which uses low-pass and high-pass filters to decompose a signal to detail (dj[n]) and approximation (aj [n]) coefficients. since the filter outputs contain half the frequency components of the original signal, both approximation and detail can be sub-sampled by two, maintaining the number of samples. this process is iteratively repeated for the approximation coefficients, increasing the frequency resolution on each decomposition step. as result a binary decomposition tree is obtained, where each level corresponds to a different scale j [ ] dj+ [m] = √ ∞ ∑ n=−∞ g[n − m]aj [n], ( ) aj+ [m] = √ ∞ ∑ n=−∞ h[n − m]aj[n], ( ) si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . here g[n] and h[n] are the impulse responses of the high-pass and low-pass filters associated with the wavelets and scaling functions, respectively. the wpt could be considered as an extension of the dwt which pro- vides more flexibility on frequency band selection. with the same reasoning above, details (high frequency components) can be decomposed as well as approximations (low frequency components). in a similar way to the dwt, the full wavelet packets decomposition tree is obtained by c p j+ [m] = √ ∞ ∑ n=−∞ g[n − m]cpj [n], ( ) c p+ j+ [m] = √ ∞ ∑ n=−∞ h[n − m]cpj [n], ( ) where j is the depth of the node and p indexes the nodes in the same depth, every c p j with p even is associated to approximations and every c p j with p odd is associated to details. the wavelet packet analysis allows to represent the information contained in a signal in a more flexible time-scale plane, by selecting different sub-trees from the full decomposition (figure ). for the selection of the best tree it is possible to make use of the knowledge about the characteristics of the signal and to obtain an efficient representation in the transform domain. for the case of signal compression the criteria is based on “entropy” measures, method named as best orthogonal basis [ ]. another possibility, closer to the classification problem, is to use the local discriminant basis algorithm, which provides an appropriate orthogonal basis for signal classification [ ]. these criteria are based on the assumption that an orthogonal basis is convenient. nevertheless, for the case in study there is not evidence on the convenience of a non-redundant representation. moreover, the redundancy often provides additional robustness for classification tasks in adverse conditions [ ]. be- cause of this, a method which explores a wider range of possibilities should be studied. . . genetic wavelet packets genetic algorithms are meta-heuristic optimization methods motivated by the process of natural evolution [ ]. a classic ga consists of three kinds of operators: selection, variation and replacement [ ]. selection mimics the natural advantage of the fittest individuals, giving them more chance to si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . figure : wavelet packets tree with six decomposition levels for a samples signals. figure : general scheme of the proposed wrapper method. reproduce. the purpose of the variation operators is to combine informa- tion from different individuals and also to maintain population diversity, by randomly modifying chromosomes. the number of individuals in the cur- rent population that are substituted by the offspring is determined by the replacement strategy. the information of a possible solution is coded in a chromosome and its fitness is measured by an objective function, which is specific to a given problem. parents, selected from the population, are mated to generate the offspring by means of the variation operators. the popula- tion is then replaced and the cycle is repeated until a desired termination criterion is reached. once the evolution is finished, the best individual in the population is taken as the solution for the problem [ ]. genetic algorithms are inherently parallel, and one can easily benefit from this to increase the computational speed [ ]. in this case, the objective function needs to evaluate the signal represen- tation suggested by a given chromosome, providing a measure of the class separability. therefore, the fitness function was defined as a phoneme clas- sifier, based on the optimized learning vector quantization (o-lvq) tech- si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . figure : frequency band integration scheme (half tree). nique [ ]. this classifier uses a set of reference vectors (codebook ) which are adapted using a set of training patterns in order to represent the distribu- tion of classes. o-lvq was chosen because it requires much less processing time than those used in state-of-the-art speech recognizers, based on hidden markov models (hmm). although, after the optimization, for the validation of the evolved representation an hmm based classifier was also used. for the evaluation of every individual in the population, the classifier is trained and tested on a phoneme corpus, and the recognition rate is used as the fitness value. the scheme of the proposed wrapper method is shown in figure . the ga uses roulette wheel selection method, the classic mutation and one-point crossover. also, an elitist replacement strategy was applied, which maintains the best individual to the next generation. the feature extraction scheme was designed for signals of samples length, this is ms frames at khz sampling frequency. and the iterative process of filtering and decimation was performed to obtain six decomposi- tion levels, obtaining a full wavelet packet tree consisting of coefficients. then, in order to reduce the dimensionality of the search space, the coeffi- cients inside each frequency band were “integrated” by groups. this means that each band was subdivided into groups, and an energy coefficient for each group was obtained by e s j = ∑ ∀ck∈g s j |ck| , ( ) where esj is the energy coefficient for integration group j in scale s, g s j , and ck is the k-th coefficient belonging to this group. figure illustrates the integration scheme for the first half of the wpt decomposition tree, while the other half is integrated in the same manner. each small square represents si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . table : integration scheme applied to the wpt tree for a sample signal, which reduces from wavelet coefficients to integration coefficients. level nodes integration groups per node wavelet coefficients per group integration coefficients per level a single component, in the first row (level ) this is a sample of the temporal signal and for the other rows (levels to ) each of them correspond to a single wavelet coefficient. white and gray zones delimit the different integration groups and the tree nodes in each decomposition level are indicated with thick line rectangles. table shows the number of components and coefficients in each integration group. this integration scheme was heuristically designed, considering the most relevant frequency bands in speech and their temporal resolutions. after band integration, a normalization was applied: if wp[k] is the k- th energy coefficient corresponding to the pattern p, then the normalized coefficient will be given by ŵp[k] = wp[k] max i {wi[k]} . ( ) a canonical evolution model with binary chromosomes was used, in which every individual represents a different selection of the wpt band-integrated coefficients. each gene in a chromosome represents one normalized coef- ficient, and its value indicates whether that coefficient should be used to parametrize the signals (figure ). once the data base is processed, each feature vector is composed by the normalized and band-integrated wpt co- efficients. these labeled patterns are used to train and test the o-lvq based classifier. when a particular individual is evaluated, each feature vector is reduced to the subset of coefficients indicated by the chromosome. the selection of individuals should be done considering the set of coeffi- cients represented by each chromosome. the chromosomes which codify the best signal representations, those which allow better classification results, should be assigned high probability. as the codification could be redundant and no restriction is imposed for coefficients combination, the ga initial- ization consists on a random settling of the genes in the chromosomes. all si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . figure : codification example with a genes chromosome and the corresponding wpt tree. dark boxes represent the used coefficients and the white boxes represent those that are discarded. algorithm : optimization for gwp. obtain full wpt for each phoneme example in the corpus using ( ) and ( ) apply the band-integration scheme to wpt coefficients using ( ) normalize the integrated coefficients for each pattern according to ( ) initialize the ga population evaluate population (algorithm ) repeat select parents (roulette wheel) create a new population from selected parents replace population evaluate population until stopping criteria is met the steps involved in the gwp feature selection strategy are summarized in algorithm , while the details for the population evaluation are shown in algorithm . . . phoneme corpus the speech data used for experimentation is a subset of the albayzin geographic corpus [ , ], named minigeo. this subset consists of ut- terances, spoken by twelve different speakers (six men and six women) which where between fifteen and fifty five years old. this speech data has been phonetically segmented using a speech recognition system based on hidden markov models [ ]. this way the temporal localization of every phoneme within each utterance was obtained. the extracted phonetic corpus was par- titioned in three groups, a training set and a testing set to be used during evolution, and a third data set which was used only for the validation of best si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . algorithm : evaluate population. for each individual in the population do re-parameterize training/test patterns according to the chromosome train the lvq based classifier on the training set test the lvq based classifier on the test set assign classification rate as the current individual’s fitness solutions after the feature selection process. the experiments included the phonemes /a/, /e/, /i/, /o/, /u/, /b/, /d/, /p/ and /t/ from spanish. the five vowels were included because of their obvious importance in the language, while the four occlusive phonemes have been included because of their similar characteristics, that make a set particularly difficult to distinguish [ ]. even though it can be suggested that all phonemes should be included, our hypothesis is that this strategy simplifies the task of the ga while still allows to find features useful in continuous asr. in order to avoid adding additional complexity to the search of the ga, every sample used in the optimization consisted in a single speech frame of ms length, which was extracted from the center of the phoneme utterance. . results and discussion . . genetic optimization of wavelet packets in [ ], various wavelet families have been tested in order to find which one is the most convenient for signal classification. in this work the tests included the most widely used families, among which we can mention meyer, daubechies, symmlets, coiflets y splines [ ]. as result, the th order coiflet family was chosen to be used on the following experiments. for the first experiment, a codebook of vectors ( per phoneme) was used within the lvq classifier and the initial learning rate was set on . . the classifier training was made in epochs with patterns and patterns were left for testing. for the ga, the population size was set on individuals, while crossover and mutation probabilities were set on . and . , respectively. the performance of the best solution found was . % of correct classifications. as the lvq codebook initialization has a random component, repeated evaluations for the same individual may result on different fitness values. si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . table : summary of obtained fitness and validation results. strategy convergence fitness validation (# generations) average std random lvq initialization . % . % . % fixed lvq initialization . % . % . % generational lvq initialization . % . % . % then, in order to obtain a good estimation of the performance for the best individual, after the evolution, it was evaluated ten times with a validation data set in this process we used training patterns and test pat- terns which were not used during the evolution. as result, an average of . % correct classifications with a standard deviation of . % was ob- tained. also, in order to analyze the effect of the random lvq initialization on the evolution, two different alternatives were considered. in the first case, the randomness was eliminated from the codebook initialization. in this case, the ga was able to find an individual with a fitness (classification rate) of . %. with the validation procedure, described earlier, an average clas- sification rate of . % was obtained. even though an improvement was obtained, it is possible that the evolution was biased by this fixed codebook initialization. then, another strategy was considered, in which a fixed initial- ization sequence was used for each generation. in order to allow individuals evaluated with different conditions (initializations) to coexist in the same generation a generational gap of individuals was used, maintaining more information from one generation to the next. this means that, in addition to the best individual which is preserved by the elitist strategy, another individuals are chosen by the selection algorithm to be maintained to the next generation. in this case the best solution achieved . % correct clas- sifications, and an average classification rate of . % was obtained with the validation data. table summarizes these results, showing that this last strategy allowed to improve the generalization capability. table shows a confusion matrix obtained from validation results. as this matrix shows, the /t/ phoneme is mainly ( . %) classified as /p/. this error turn up because the experimental data was taken from the cen- tral part of the samples and the plosive phonemes, like /p/ and /t/, have their most particular attributes at the beginning (the phoneme plosion). a similar problem happens with phoneme /d/, and this might be solved for all plosive phonemes by considering their context (i.e. a number of precedent si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . table : confusion matrix obtained from the validation of the best ga solution, giving . % average classification rate . /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ . . . . . . . . . /e/ . . . . . . . . . /i/ . . . . . . . . . /o/ . . . . . . . . . /u/ . . . . . . . . . /b/ . . . . . . . . . /d/ . . . . . . . . . /p/ . . . . . . . . . /t/ . . . . . . . . . and posterior frames). . . comparative analysis in order to compare this results, the same classifiers were trained with other state-of-the-art speech features: the classic mfcc [ ], the alterna- tive cepstral features based on slaney’s filter-bank [ ], and a representation based on the standard dwt. it is worth pointing out that, as the speech data used in these experiments was sampled at khz, the original parameters of the filter-bank proposed by slaney were modified by following the reason- ing in [ ]. table shows the results obtained with the above mentioned representations and using a validation data set consisting of patterns, which were not used for the optimization. as it can be seen, the best aver- age classification rate was obtained with the gwp. this shows that by the genetic optimization of the full wpt decomposition it is possible to improve the classification results, in contrast to the classic cepstrum based represen- tations. on the other hand, it can be noticed that the phoneme /d/ seems to be the most difficult for all representations, except for dwt. even though the state-of-the-art speech recognizers use hmm for acoustic modeling, the significant improvements provided by gwp when using an lvq based classifier should not be disregarded. these results clearly show that a more efficient class separation is provided in the gwp features space. it should be taken into account that this straightforward classifier was used as objective function to guide evolution, and only the central frame was considered for each phoneme pattern during the optimization. si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . table : classification results obtained with an lvq based classifier. phoneme cepstral coefficients wavelets mfcc( ) slaney( ) dwt( ) gwp( ) /a/ . . . . /e/ . . . . /i/ . . . . /o/ . . . . /u/ . . . . /b/ . . . . /d/ . . . . /p/ . . . . /t/ . . . . average . . . . . . evaluation with hidden markov models in this work we have raised the hypothesis that, by using a simple classifier in the optimization, the class separability would be maximized and this could also be beneficial to a more sophisticated classifier, like hmm [ ]. in order to verify this, the performance of an hmm based classifier was evaluated for each of the representations in table . this classifier is based on a continuous hmm, using gaussian mixtures with diagonal co-variance matrices for the observation densities, as common in asr [ ]. we used a three state model with mixtures of four gaussians, constructed with the tools provided in the hmm toolkit (htk) [ ]. these tools use the baum-welch algorithm [ ] to train the hmm parameters, and the viterbi algorithm [ ] to search for the most likely state sequence. this classifier was evaluated in a ten-fold cross-validation process with random partitions, each of which consisted of and patterns for training and testing, respectively. it is important to point out that, because of the nature of hmm, in the evaluation of this classifier all the successive frames composing a phoneme were used. while, for the lvq based classifier, only the central frame was used for a particular phoneme. for the features based on dwt the training of the hmm classifier could not converge, which is mainly because the gaussian mixtures are not able to adequately model the probability densities of these coefficients [ ]. an- other problem for training the models with dwt coefficients is due to the high dimensionality of this representation. then, a post-processing based on principal component analysis (pca) [ ] was applied in order to obtain a rep- resentation of lower dimensionality, with probability densities more similar to si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . table : classification results obtained with an hmm based classifier. phoneme cepstral coefficients wavelets mfcc( ) slaney( ) dwt+pca( ) gwp( ) /a/ . . . . /e/ . . . . /i/ . . . . /o/ . . . . /u/ . . . . /b/ . . . . /d/ . . . . /p/ . . . . /t/ . . . . average . . . . gaussians. the best result for dwt+pca was obtained when preserving the % of the variance, giving a representation of dimensions. even though our optimized representation is also based on wavelets, and the same prob- lem could be expected, no post-processing was necessary for gwp. thus, we can assume that the band integration, besides reducing dimension, produced coefficients with probability densities more appropriate to gaussian mixture modeling. table shows the phoneme classification results obtained with hmm, comparing gwp and other state-of-the-art speech representations. the op- timized representation is the same from table , which was evolved using an lvq-based classifier. it can be noticed that the best results are those ob- tained by means of the optimized representation and mfcc, similar to the case of the validation with lvq (table ). even though the fitness was mea- sured by means of an lvq based classifier in the optimization, the evolved representation provided satisfactory results when using hmm. this means that the optimized representation captures information which is relevant for the discrimination, regardless of the type of classifier. moreover, using this low-cost classifier, we have successfully saved significant computational time in the optimization. it should be taken into account that if we had used an hmm based classifier and, therefore, considered all the possible frames within a phoneme example, the evaluation of each individual would have taken approximately ten times more. it is also important to note that the proposed gwp representation, which yielded the best classification result when using an lvq based classifier, pro- vides relatively lower performance with hmm. this is because, as explained si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . before, the probability densities of the coefficients provided in wavelet based representations are not entirely suitable for gaussian mixture models [ ]. then, different alternatives remains to be explored besides band integration and pca post-processing, in order to obtain coefficients more suitable to gaussian mixture models. also, it should be considered that the dimension- alities of the wavelet based representations are much higher than those of the cepstral representations, which makes the training of the classifier more difficult. despite the previous considerations, the results from tables and show that the proposed method is useful for the optimization of wavelet based rep- resentations. moreover, the results obtained with the gwp features shown that using this evolutionary methodology it is possible to improve the per- formance of the classical representations in asr. . . evaluation in noise conditions in order to evaluate the robustness of the optimized representation, white noise was added to the original utterances. the tests were made at several noise levels, and the mismatch training and test (mmtt) condition was considered, which means that the classifiers were trained with clean signals and tested with noisy signals. in general, the input of an asr system would consist on speech signals with different snr to those in the training set. for this reason, the evaluation of the recognition performance in mmtt conditions is more realistic than the case where the snr is the same in both training and test sets. each test consisted in a ten-fold cross-validation process with random partitions, which consisted of and patterns for training and testing, respectively. the process of training and testing was repeated for these ten partitions and results were averaged, ensuring that the resultant accuracy would not be biased because of a particular data partition. figure shows the average results, as well as the estimated standard deviations, showing that the gwp improves the mfcc in all cases. at and db snr the dwt+pca representation gives better results, however, for most of the noise conditions its performance is noticeably worse than gwp. it can be noticed that, on average, the results given by gwp are significantly better when compared to the other representations. in order to evaluate the statistical significance of these results, we estima- ted the probability that gwp is better than each of the reference representa- tions for the given tasks. to perform this test, statistical independence of the si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . db db db db db db dbclean snr c la ss ifi ca tio n r a te [ % ] gwp dwt+pca mfcc slaney figure : classification results obtained with an hmm based classifier evaluated in mmtt conditions. classification errors for each phoneme has been assumed, and the binomial distribution of the errors was approximated by means of a gaussian distri- bution. table shows the statistical significance of the classification rates obtained with the hmm based classifier: results improved by gwp with statistical significance higher than % are indicated with the superscript △. it can be noticed that most of the classification improvements obtained with the proposed optimized representation are statistically significant. for instance, the probability that gwp performs better than the cepstral coef- ficients obtained with slaney’s filterbank is higher than . for all snrs. table shows more detailed information about the classification perfor- mance comparing mfcc and gwp, for the case of db snr in mmtt conditions. in these confusion matrices, the rows correspond to the actual phoneme and the columns to the predicted phoneme, while the percentages of correct classification are on the diagonal. these matrices show coincidences between the phonemes which are most confused with mfcc and the ones that are confused with gwp. for example, in both cases phoneme /t/ was confused with /p/ and vice versa, which is reasonable as these two plosive si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . table : statistical significance of classification results obtained with an hmm based classifier evaluated in mmtt conditions. superscript △ indicates that the statistical significance of the improvement of gwp is higher than %. snr cepstral coefficients wavelets mfcc( ) slaney( ) dwt+pca( ) gwp( ) clean . . △ . △ . db . △ . △ . △ . db . . △ . △ . db . △ . △ . △ . db . △ . △ . △ . db . △ . △ . . db . △ . △ . . db . △ . △ . . table : confusion matrix obtained from the validations with mfcc and gwp in mmtt conditions and db snr. mfcc gwp /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ . . . . . . . . . . . . . . . . . . /e/ . . . . . . . . . . . . . . . . . . /i/ . . . . . . . . . . . . . . . . . . /o/ . . . . . . . . . . . . . . . . . . /u/ . . . . . . . . . . . . . . . . . . /b/ . . . . . . . . . . . . . . . . . . /d/ . . . . . . . . . . . . . . . . . . /p/ . . . . . . . . . . . . . . . . . . /t/ . . . . . . . . . . . . . . . . . . average: . % average: . % consonants share many spectral features. in a similar way vowels /o/ and /u/, which are close in the formants map, are quite confused in both cases. similarly, table shows the confusion matrices obtained in the classification with mfcc and gwp in the case of db snr and mmtt conditions. it can be noticed that the vowels /a/ and /u/ are mostly misclassified with mfcc, but they are significantly better distinguished with gwp. the op- timized features also introduced an important improvement for the vowel /i/, which is confused with the phoneme /d/ when using mfcc. it is also interesting to note that the phonemes which have their classification rates most affected by noise when using mfcc and gwp do not match. though, when the noise level is increased, the number of confusions between phonemes /t/ and /d/ increases for both mfcc and gwp. similarly, the number of confusions between phonemes /t/ and /p/ is also increased in both cases. another interesting remark is that, for both representations, phoneme /t/ is si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . table : confusion matrix obtained from the validations with mfcc and gwp in mmtt conditions and db snr. mfcc gwp /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ . . . . . . . . . . . . . . . . . . /e/ . . . . . . . . . . . . . . . . . . /i/ . . . . . . . . . . . . . . . . . . /o/ . . . . . . . . . . . . . . . . . . /u/ . . . . . . . . . . . . . . . . . . /b/ . . . . . . . . . . . . . . . . . . /d/ . . . . . . . . . . . . . . . . . . /p/ . . . . . . . . . . . . . . . . . . /t/ . . . . . . . . . . . . . . . . . . average: . % average: . % better classified when the snr is db than when it is db. these results show that the classification performance of the classical representation can be improved. this suggest that, by means of this gwp methodology, ad- ditional robustness against white noise can be provided to a state-of-the-art asr system. in order to perform a qualitative analysis of the optimized representation, the tiling of the time-frequency plane was constructed from the selected de- composition using the criteria proposed in [ ]. the result is shown in fig- ure , where each decomposition level is depicted separately for an easier interpretation. each ellipse represents a selected group from the integration scheme (table ), then their widths and time localizations are determined by the time-frequency atoms corresponding to the integration group (figure ). therefore, each element in the tiling represents a time-frequency atom that was obtained by adding the original wavelet atoms, according to the integration scheme. this explains why the atoms for levels and are the same time width, as the number of coefficients integrated in the groups for level are twice the number of coefficients in the groups for level (figure ). the same explanation applies for the width of the atoms in levels and . note that the whole time-frequency tiling is obtained by the superposi- tion of these six sub-figures, yielding great overlapping between atoms from different decomposition levels. a first observation is that the optimization of the wpt-based decomposition led to a highly redundant non-orthogonal representation, which has been able to exploit this redundancy in order to increase robustness against additive noise. however, the optimized repre- sentation uses only % of the total of the coefficients obtained from the integration of the whole wpt-tree. this also suggest that it could be pos- si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . sible to achieve still further redundant and robust representations. it is also interesting to note that there are some selected atoms concentrated at the center of the time axis, which could be related to how the phonemes were sampled from the speech corpus, as only the frames extracted from the center of each utterance were considered for the optimization. also, there are some atoms concentrated at the side parts of the time axis, which could be related to the plosive phonemes. . conclusion and future work a wrapper optimization strategy has been proposed, taking advantage of the benefits provided by evolutionary computation techniques, in order to carry on the search for an advantageous wavelet-based speech representa- tion. the results, obtained in the classification of a group of nine phonemes from spanish, shown that the optimized representation provides important improvements in comparison to the classical features. this suggests that the task of a classifier is simplified when using this optimized representation, due to a better class separation in the features space. therefore, the proposed strategy provides an alternative feature set for speech signals, which allows to improve the classification results in the presence of noise. in future work we would design more specific genetic operators, so that more information about the problem in hand could be incorporated to the search. pursuing the objective of finding a representation more suitable for hmm with gaussian mixture modeling, one interesting idea is to incorpo- rate some measure about the gaussianity of the probability densities of the gwp coefficients to the fitness function. besides, we would study different alternatives to the proposed band integration scheme and the use of tempo- ral information, in regard to successive speech frames, like first and second derivatives. in order to obtain a representation which allows to improve the results in a continuous speech recognition system, future experiments will include more phonemes in the data-sets used for the optimization. also, the robustness of the representation will be evaluated in comparison with different state-of- the-art robust representations, considering different noise types. references [ ] l. rabiner, b. juang, fundamentals of speech recognition, prentice hall, nj, . si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . time [msec] f re q u e n cy [ k h z] selected atoms for level time [msec] f re q u e n cy [ k h z] selected atoms for level time [msec] f re q u e n cy [ k h z] selected atoms for level time [msec] f re q u e n cy [ k h z] selected atoms for level time [msec] f re q u e n cy [ k h z] selected atoms for level time [msec] f re q u e n cy [ k h z] selected atoms for level figure : tiling of the time-frequency plane obtained for the optimized decomposition. for a better visualization, each decomposition level was schematized separately. si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . [ ] l. rabiner, r. schafer, digital processing of speech signals, prentice hall, nj, . [ ] s. v. davis, p. mermelstein, comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences, ieee transactions on acoustics, speech and signal processing ( ) – . [ ] h. hermansky, perceptual linear predictive (plp) analysis of speech, j acoust soc am ( ) ( ) – . doi: . / . . [ ] h. hermansky, n. morgan, rasta processing of speech, ieee trans speech audio process ( ) – . doi: . / . . [ ] a. hassanien, g. schaefer, a. darwish, computational intelligence in speech and audio processing: recent advances, in: x.-z. gao, a. gaspar-cunha, m. kppen, g. schaefer, j. wang (eds.), soft com- puting in industrial applications, vol. of advances in intelligent and soft computing, springer berlin / heidelberg, , pp. – , . / - - - - - . [ ] n. nehe, r. holambe, dwt and lpc based feature extraction methods for isolated word recognition, eurasip journal on audio, speech, and music processing ( ) ( ) . doi: . / - - - . url http://asmp.eurasipjournals.com/content/ / / [ ] s. patil, m. dixit, speaker independent speech recognition for diseased patients using wavelet, journal of the institution of engineers (india): series b ( ) – , . /s - - - . url http://dx.doi.org/ . /s - - - [ ] e. avci, z. h. akpolat, speech recognition using a wavelet packet adap- tive network based fuzzy inference system, expert systems with appli- cations ( ) ( ) – . doi: . /j.eswa. . . . [ ] j.-d. wu, b.-f. lin, speaker identification using discrete wavelet packet transform technique with irregular decomposition, expert systems with applications ( , part ) ( ) – . [ ] m. vetterli, c. herley, wavelets and filter banks: theory and design, ieee trans. signal proc. ( ) ( ) – . si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . [ ] n. hess-nielsen, m. v. wickerhouser, wavelets and time-frequency analisys, proceedings of the ieee ( ) ( ) – . [ ] r. munkong, b.-h. juang, auditory perception and cogni- tion, signal processing magazine, ieee ( ) ( ) – . doi: . /msp. . . [ ] s. ray, a. chan, automatic feature extraction from wavelet coefficients using genetic algorithms, in: proceedings of the ieee signal pro- cessing society workshop, neural networks for signal processing xi, , pp. – . [ ] d. wang, d. miao, c. xie, best basis-based wavelet packet entropy feature extraction and hierarchical eeg classification for epileptic detec- tion, expert systems with applications ( ) ( ) – . doi: . /j.eswa. . . . [ ] h. rufiner, j. goddard, a method of wavelet selection in phoneme recog- nition, in: circuits and systems, . proceedings of the th midwest symposium on, vol. , , pp. – vol. . [ ] p. kumar, m. chandra, hybrid of wavelet and mfcc features for speaker verification, in: information and communication technolo- gies (wict), world congress on, , pp. – . doi: . /wict. . . [ ] v. tiwari, j. singhai, wavelet based noise robust features for speaker recognition, international journal of signal processing ( ) ( ) – . [ ] a. daamouche, l. hamami, n. alajlan, f. melgani, a wavelet optimiza- tion approach for ecg signal classification, biomedical signal processing and control. in press. doi: . /j.bspc. . . . [ ] a. r. ferreira da silva, approximations with evolutionary pursuit, sig- nal processing ( ) ( ) – . [ ] r. behroozmand, f. almasganj, optimal selection of wavelet- packet-based features using genetic algorithm in pathological assess- ment of patients’ speech signal with unilateral vocal fold paraly- sis, computers in biology and medicine ( ) ( ) – . doi: . /j.compbiomed. . . . si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . [ ] a. r. ferreira da silva, wavelet denoising with evolutionary al- gorithms, digital signal processing ( ) ( ) – . doi: . /j.dsp. . . . [ ] e.-s. el-dahshan, genetic algorithm and wavelet hybrid scheme for ecg signal denoising, telecommunication systems ( ) – , . /s - - - . [ ] r. salvador, f. moreno, t. riesgo, l. sekanina, evolutionary ap- proach to improve wavelet transforms for image compression in em- bedded systems, eurasip journal on advances in signal processing ( ). doi: . / / . [ ] w. zhao, c. e. davis, swarm intelligence based wavelet coefficient feature selection for mass spectral classification: an application to proteomics data, analytica chimica acta ( ) ( ) – . doi: . /j.aca. . . . [ ] a. daamouche, f. melgani, swarm intelligence approach to wavelet design for hyperspectral image classification, geoscience and remote sensing letters, ieee ( ) ( ) – . doi: . /lgrs. . . [ ] w. pedrycz, s. s. ahmad, evolutionary feature selection via structure retention, expert systems with applications ( ) ( ) – . doi: . /j.eswa. . . . [ ] s. chatterjee, a. bhattacherjee, genetic algorithms for feature selection of image analysis-based quality monitoring model: an application to an iron mine, engineering applications of artificial intelligence ( ) ( ) – . doi: . /j.engappai. . . . [ ] y.-x. li, s. kwong, q.-h. he, j. he, j.-c. yang, genetic algorithm based simultaneous optimization of feature subsets and hidden markov model parameters for discrimination between speech and non-speech events, international journal of speech technology ( ) – , . /s - - - . [ ] l. d. vignolo, h. l. rufiner, d. h. milone, j. c. goddard, evolution- ary splines for cepstral filterbank optimization in phoneme classifica- si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . tion, eurasip journal on advances in signal processing volume , doi: . / / . [ ] l. d. vignolo, h. l. rufiner, d. h. milone, j. c. goddard, evolutionary cepstral coefficients, applied soft computing ( ) ( ) – . doi: . /j.asoc. . . . [ ] l. vignolo, h. rufiner, d. milone, j. goddard, genetic optimization of cepstrum filterbank for phoneme classification, in: proceedings of the second international conference on bio-inspired systems and signal processing (biosignals ), insticc press, porto (portugal), , pp. – . [ ] l. vignolo, d. milone, h. rufiner, e. albornoz, parallel implementation for wavelet dictionary optimization applied to pattern recognition, in: proceedings of the th argentine symposium on computing technology, mendoza, argentina, . [ ] s. durbha, r. king, n. younan, wrapper-based feature subset selection for rapid image information mining, geoscience and remote sensing letters, ieee ( ) ( ) – . doi: . /lgrs. . . [ ] h.-h. hsu, c.-w. hsieh, m.-d. lu, hybrid feature selection by com- bining filters and wrappers, expert systems with applications ( ) ( ) – . doi: . /j.eswa. . . . [ ] r. kohavi, g. h. john, wrappers for feature subset selection, artificial intelligence ( - ) ( ) – . doi: . /s x. [ ] s. mallat, a wavelet tour of signal processing, rd edition, academic press, . [ ] s. mallat, a theory of multiresolution of signal decomposition: the wavelet representation, ieee trans. pattern anal. machine intell. ( ) ( ) – . [ ] r. coifman, m. v. wickerhauser, entropy-based algorithms for best basis selection, ieee transactions on information theory ( ) ( ) – . si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . [ ] n. saito, local feature extraction and its applications using a library of bases, ph.d. thesis, yale university, new haven, usa, director-ronald r. coifman ( ). [ ] s. n. sivanandam, s. n. deepa, introduction to genetic algorithms, springer london, limited, . [ ] h. youssef, s. m. sait, h. adiche, evolutionary algorithms, simulated annealing and tabu search: a comparative study, engineering applica- tions of artificial intelligence ( ) ( ) – . doi: . /s - ( ) - . [ ] z. michalewicz, genetic algorithms + data structures = evolution programs, springer-verlag, . [ ] t. kohonen, improved versions of learning vector quantization, in: proc. of the int. joint conf. on neural networks, san diego, , pp. – . [ ] a. echeverŕıa, j. tejedor, d. wang, an evolutionary confidence measure for spotting words in speech recognition, in: y. demazeau, f. dignum, j. corchado, j. bajo, r. corchuelo, e. corchado, f. fernández-riverola, v. julián, p. pawlewski, a. campbell (eds.), trends in practical ap- plications of agents and multiagent systems, vol. of advances in intelligent and soft computing, springer berlin / heidelberg, , pp. – . [ ] a. moreno, d. poch, a. bonafonte, e. lleida, j. llisterri, j. marino, c. nadeu, albayzin speech database design of the phonetic corpus, tech. rep., universitat politecnica de catalunya (upc), dpto. dtsc ( ). [ ] s. young, g. evermann, d. kershaw, g. moore, j. odell, d. ollason, v. valtchev, p. woodland, the htk book, htk version . , cambridge university ( ). url http://htk.eng.cam.ac.uk [ ] a. quilis, tratado de fonoloǵıa y fonética españolas, biblioteca románica hispánica, editorial gredos, madrid, . si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . [ ] h. rufiner, comparación entre análisis onditas y fourier aplicados al reconocimiento automático del habla, master’s thesis, universidad autónoma metropolitana, iztapalapa ( ). [ ] i. daubechies, ten lectures on wavelets, society for industrial and applied mathematics, philadelphia, pa, usa, . [ ] m. slaney, auditory toolbox, version , technical report - , interval research corporation, apple computer inc. ( ). [ ] t. ganchev, n. fakotakis, g. kokkinakis, comparative evaluation of various mfcc implementations on the speaker verification task, in: proceedings of the specom- , , pp. – . [ ] x. d. huang, y. ariki, m. a. jack, hidden markov models for speech recognition, edinburgh university press, . [ ] k. demuynck, j. duchateau, d. van compernolle, p. wambacq, im- proved feature decorrelation for hmm-based speech recognition, in: proceedings of the th international conference on spoken language processing (icslp ), sydney, australia, . [ ] f. jelinek, statistical methods for speech recognition, mit press, cam- brige, masachussets, . [ ] d. h. milone, l. e. d. persia, m. e. torres, denoising and recognition using hidden markov models with observation distributions modeled by hidden markov trees, pattern recognition ( ) ( ) – . doi: . /j.patcog. . . . [ ] c. m. bishop, pattern recognition and machine learning, st edition, springer, . [ ] m. lewicki, efficient coding of natural sounds, nature neuroscience ( ) ( ) – . si nc (i ) r es ea rc h c en te r fo r s ig na ls , s ys te m s an d c om pu ta ti on al i nt el li ge nc e (f ic h. un l. ed u. ar /s in c) l . d . v ig no lo , d . h . m il on e & h . l . r uf in er ; "g en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" e xp er t s ys te m s w it h a pp li ca ti on s, . a rule-based expert system for music perception behavior research methods, instruments, & computers / , ( ), - a rule-based expert system for music perception jacqueline a. jones, benjamin o. miller, and don l. scarborough brooklyn college of the city university of new york, brooklyn, new york we describe an expert system written in pascal to simulate part of lerdahl and jackendoffs ( ) theory of music perception. this program, a production system based on a hearsay ii ar- chitecture (nii, ), illustrates a number of techniques, including program modularity, com- plex data structures, and simulated parallelism using operating system concepts. people perceive structure in the music they hear. con- sider the simple song "row, row, row your boat." asked to sing this melody, most people would, without thinking about it, stress certain syllables, breaking the melody up into sections corresponding to the following lines: row row row your boat gent ly down the stream mer ri ly mer ri ly mer ri ly mer ri ly life is but a dream, in which italicized syllables are stressed, capitalized syl- lables more than lowercase. the stresses are not random or idiosyncratic. the song divides into rhythmically simi- lar halves, each consisting of two lines, with two stressed notes per line, equally spaced in time. this pattern of stresses is the metric structure. the division into lines is the grouping structure, also an important part of music perception. listeners also perceive the tonality of the piece-the key it is in. after listening to a little bit of "row, row," a listener knows what chord to expect to end the piece. these are the kinds of intuitions that lerdahl and jackendoff ( ) formalized in a genera- tive theory of tonal music (gttm). the theory does this by means of four stages of analysis, each of which is em- bodied in a set of rules. this theory is, to our knowledge, the most comprehensive formal theory of music percep- tion published to date. we are building a computer simulation of gttm to test the theory. the basic framework of the simulation is in place and currently consists of about , lines of code. turbo pascal was chosen for initial development of our simulation because of the many strengths of the language (see jones & harrow, ; kaplan, ) and because it is a language that we know well. there are, however, problems with using pascal for a large project with several programmers. the authors are listed alphabetically. all contributedequally to this work. this researchwas supportedin part by a national sciencefoun- dation graduate fellowshipawarded to miller, and by a psc-cuny research award to jones. we thank hollis scarborough for help in preparing the figures. requestsfor reprints should be sent to jacque- line a. jones, department of computer andinformation science, brook- lyn college, brooklyn, ny , or to don scarborough, department of psychology, brooklyn college, brooklyn, ny . first, standard pascal (and dialects such as version . of turbo pascal that we have been using) provides limited support for modular development oflarge programs. all subprocedures are nested within the main procedure, and there is no separate compilation of subprocedures. thus, if a line of code is modified, the entire program must be recompiled. although compilation is fast with turbo pascal, the size of our program makes this time- consuming. more important, standard pascal does not allow a sub- program to define static variables-that is, variables that are allocated permanent memory locations for the life of the program. thus it is often necessary to use global variables-that is, variables declared in the main procedure-for this purpose. however, with global vari- ables, modifying a piece of code anywhere in the pro- gram may affect other aspects of the program in unantic- ipated ways. also, name conflicts arise when several programmers are defining global variables for different subprograms. to minimize such problems, we adopted several strate- gies for program development. first, we borrowed a mod- ula approach (wirth, ) to implementing the sys- tem. each module of the program is split into two separate source files: a definition file (with a def file extension) and an implementation file (with an imp file extension). the def file contains all the global variables whose values must be retained between calls to that module; these global variables take the place of static variables. the imp file contains the actual implementation code for the mod- ule. to minimize name conflicts, we adopted a conven- tion of prefixing all true global type definitions and global variables with an underbar. the names of global variables that mimic static variables-that is, variables that exist only for the use of the module in which they are defined- are preceded by the initials of the name of the module in which they are defined. a small main source file defines the overall program structure. this file contains global type definitions and variable declarations that must be shared by all parts of the program. the main file then specifies the def files that are to be "included" in this file in a special step dur- ing compilation. collectively, then, the main ftle and the def files specify the entire global environment of the pro- gram. the main file next specifies the imp files as an copyright psychonomic society, inc. jones, miller, and scarborough additional set of "include" files to be brought into the main file during compilation. finally, the main file con- tains a short main procedure that calls an initialization subroutine and then passes control to the scheduling rou- tine described below. a rule-based expert system although we considered implementing gtim as a neu- ral network (e.g., grossberg, ; rumelhart & mcclel- land, ; schneider, ), and hope to do so eventu- ally, there are several problems with the connectionist approach that led us to reject this approach for now. first, simulation of connectionist networks requires powerful computational resources for anything but the smallest net- works. second, the high degree of interaction between system components in a neural network makes modular development difficult. third, many properties of these systems are not well understood. fourth, many aspects of gtim are not well specified or articulated. to attempt to implement a complex and incompletely specified rule- based theory (gtim) in terms of a complex, interactive, nonsymbolic neural network is to risk chaos. thus, our first goal is to see whether the theory can be adequately implemented at the symbolic level at which it is cast. however, as noted below, some parts of the simulation are easily cast into a connectionist framework. we chose to develop this system as a rule-based expert system because ( ) gtim is expressed in terms of rules; ( ) the technology for rule-based systems is developing rapidly, and there are many available techniques to call upon; and ( ) a rule-based approach supports implemen- tation of a modular system. although gttm is expressed as a set of rules, many of the rules cannot be translated directly into simula- tion rules because, in general, the rules are not statements about how a piece of music should be analyzed, but rather are constraints that such an analysis should satisfy. a big hurdle in implementing gttm lies in discovering psychologically plausible procedures that can produce analyses satisfying such constraints. we want the simu- lation to be plausible as a real-time implementation of the psychological processes underlying music perception. to this end, we chose to implement the theory as a rule-based production system (newell & simon, ; winston, ), using a blackboard type of architecture (nii, ). blackboard systems have been popular in psychologi- cal modeling (e.g., mcclelland & rumelhart, ). a blackboard system has three basic components: ( ) a global data structure called the blackboard; ( ) a set of knowledge sources (kss), each of which encapsulates knowledge about some aspect of the problem; and ( ) a control structure that monitors and controls which kss are active. our current implementation of gtim in terms of these components is described below. arcidtecture of the system the blackboard the blackboard stores all the input data about a problem as well as information accumulated as the analysis pro- ceeds. a blackboard is generally divided into several levels, each representing a different type of information. in our implementation, the lowest level of the blackboard contains data about the individual notes in the music. other levels of the blackboard represent information about the analysis of the piece, including levels for grouping infor- mation, metric information, and tonality information, and levels based on higher level combinations of this infor- mation. in our implementation, information within each of these levels is stored in linked lists. pointers between levels of the blackboard provide appropriate linking and synchronization between levels. our blackboard data structure had to be suitable for representing musical input. this representation had to be compact, because musical notation is very dense, contain- ing thousands of notes, each supplying many different kinds of information. for this project, we assumed that the basic elements of a listener's experience include such information as pitch, note duration, dynamic level, and timbre. thus, it seemed reasonable to build the simula- tion assuming this level of knowledge as a starting point, an assumption also made by lerdahl and jackendoff ( ). input to the simulation program is a symbolic representation of such information, which forms the lowest level of the blackboard and represents what the listener hears. we needed to represent the time continuum of the music as well as the multiple events taking place at each time, with some notes continuing to sound as others begin and others end. we used a matrix in which each row represents a different instrument (or voice) and the sequence of notes it plays, and each column represents the notes played by all instruments at a specific time. a simple array representation would not be practical, given the limited memory of most personal computers, because it would be a "sparse" matrix (tenenbaum & augenstein, ), with far more storage occupied by blank information than by actual information. suppose we wanted to represent piano music. at any moment, up to notes may be played (more, if we consider sustained notes and cases where a-finger plays notes), so we might use an array with rows, representing possible notes at any time, and with columns representing successive notes in time. however, usually fewer than notes are actually played at any time. an example of this, taken from schubert, is illustrated schematically in figure la. the passage begins with an -note chord held for two time intervals, followed by a sequence of single notes. this general pattern is repeated eight times in the music. if we used an array of rows and columns to store the music, then in each occurrence of this rhythmic pattern an expert system for music perception we would have an array with cells, but would be empty. to avoid this problem, we used a sparse matrix im- plemented as a linked list. a linked list is a dynamic data structure that contains only those pieces of information that are necessary at any given time. an element of a linked list, known as a node, is a record that can contain several different kinds of data at the same time. an array exists in storage from the moment it is declared, whether it is used or not, but nodes in a linked list are allocated only as needed. the nodes are linked by pointers. pointers-graphically represented as arrows-point to places where data are stored. one node can point to the next node or to the previous node, allowing us to create a list. figure b shows the same schubert example as a sparse matrix using linked lists. here, the top line represents nodes that simply mark off successive events in the music. each box that is linked to an event is itself a node that stores information about an actual note. in this represen- (a) ~o - - ~- --- ---- --- i o o o o r-----f--+_~r__+_-._+_-_+_--l--+--_+_--_t_-+_~;-_+_-_+_-_+_-- o iii ii) ' z ! c.. ci c 'il .q --t--f---i--.-+----+-- ---- i-------+--+--f--.-+--l lime ~ (b) time -~ events i c . ci c 'il .q ~.... o z figure . (a) example of a "sparse" array with many empty cells. (b) a linked-list representation of (a). jones, miller, and scarborough tation, each occurrence of the schubert pattern requires nodes, compared with entries in the array im- plementation. a further advantage of the linked-list representation is that it requires no prior commitment as to the number of notes that can be represented at any time or the number of notes per measure. to simplify computation, the blackboard is a single global data structure that integrates all four facets of the simulation: the score, the metric structure, the grouping structure, and the tonality analysis. the "backbone" of the blackboard data structure is called the notochord. this consists of a doubly linked list of nodes corresponding to successive events in the score. each of these nodes, called event nodes, contains only one piece of informa- tion about the piece: the offset (time) from the beginning of the previous event to the beginning of the current event. an event node also stores pointers connecting it to nodes containing actual note information, as well as to the group- ing and metric representations. note that the event nodes themselves contain no score information. this is reserved for the note nodes, each of which is a record of all the information about a single note in a score. in addition to this information about each note, a note node contains a set of pointers that link it to other notes and to the notochord. the information in note of "row, row" is shown in figure in simplified schematic form. (note that there is no next voice because this is monophonic input. assume that the soprano sings the melody.) the knowledge sources the gttm rules represent sources of knowledge about how a piece should be analyzed. where possible, we im- plemented each gttm rule as a separate ks. a ks can be thought of as an "if-then" production rule; that is, a ks is a rule that specifies an action (or consequent) that applies whenever certain conditions (the antecedents or event "if' parts) occur. an example of one of the simpler gttm rules (grouping preference rule d) is as follows: if ntn nj is a -note sequence, and n, and n, differ in duration, and n, and n have equal durations, and n, and have equal durations, then there is evidence of a grouping boundary be- tween n, and nj' in the gttm simulation, such a rule is implemented as a ks module containing two parts: ( ) a pattern match- ing procedure that looks at the blackboard for the speci- fied antecedent pattern; and ( ) a consequent procedure that makes appropriate changes to the blackboard. to create a uniform interface between different kss, we adopted a limited form of "object oriented program- ming" (jacky & kalet, ). each ks is treated as an "object" that controls certain data structures, and that contains a set of methods for operating on these data struc- tures. the ks operates on the data structure when it is sent a message specifying an operation to perform. we designed our kss so that each responds to a limited set of messages (e.g., initialize, evaluate, execute), thus facilitating a modular design that makes it fairly easy to add and replace kss. the control structure the production rules (the kss) operate under control of a scheduling procedure that monitors the activities of the kss and the state of the blackboard. the scheduling mechanism reflects, in part, the fact that the simulation must be run on a serial von neumann type of computer, where only a single ks can be active at anyone time. however, the scheduling mechanism can also represent changes in attention and focus as the analysis of the piece proceeds. furthermore, when appropriately implemented, the scheduler can simulate parallel concurrent activity of node pre vi ous note for this voice ( ni ) / voice = soprano pitch = c in rd octave duration = dotted quarter loudness = mf articulation = normal offset from p r e v. note a next voice ( ni ) / next note (note ) figure . schematic representation of the first note node for "row, row, row your boat." an expert system for music perception many kss. the key to this aspect of our simulation lies in the window that specifies the portion of the blackboard that is "visible." a ks can examine only the portion of the blackboard within the window. the window represents the capacity of short-term memory, which applies to music as well as to verbal material (dowling & harwood, ). in contrast, ler- dahl and jackendoff( ) offered analyses that assumed unlimited memory. clearly, though, the perception of notes currently being heard is affected by how much a listener retains of the previous events. therefore, our simulationis organized around the concept of a temporal window, so that in running the simulation we can vary parametrically the amount of prior input and analysis available at any point. initially, the program reads a file of information about a piece of music and builds the notochord data structure with the linked note lists to represent the piece. an initial portion of the notochord is then marked off by the win- dow. the portion of the piece within the window represents what a listener first hears. this might be any- thing from note to the entire piece, although currently we are working with a window size of to notes. the scheduling procedure maintains a task table, which is an array of records, one record per ks. the array en- try for a ks provides the scheduler with information about the status of the ks, such as whether it is currently run- nable and its priority. a ks can run if the conditions it requires are currently met. the scheduler selects a ks on the basis of which kss are runnable and the priority blackboard assigned to each. it then copies the task table entry for that ks to a global variable, where it is accessible to the ks, and then sends the ks a message specifying the ac- tion to be performed (e.g., modify blackboard). the ks can modify entries in its task table entry to tell the scheduler about changes to its status and/or the actions it has performed. an outline of the system is shown in figure , which also shows the message manager mod- ule that controls all output messages sent by the scheduler and the kss. this general framework for controlling execution lends itself to several scheduling algorithms. in round-robin scheduling, the schedulersimply goes down the list of en- tries in the task table, giving each runnable ks a chance to execute within the current window. then the window is advanced (the new window may overlap the old and may exclude earlier events), and the scheduler calls each ks so the ks may evaluate its applicability to the new window. the schedulerthen calls the runnable kss again. a more complex algorithm schedules kss on the basis of priority. in this method, the scheduler maintains a list of runnable kss in order of priority and always calls the ks with the highest priority. this scheduling procedure simulates parallel execution of the kss in much the same way that multiuser operat- ing systems such as unix produce apparent parallel ex- ecutionof severaltasks: the scheduler, whichkeeps track of all the tasks in the system, selects a task and allocates a small amount of time to it, and then selects another task for execution whenthe first one has usedits allocated time, to flies, printer, console window ksl ks i ks' i( ) ksj param~ters i ksn ks task table --@ inactive knowledge urcetr figure . control structure for the simulation. ks = knowledge structure. jones, miller, and scarborough finished, or blocked itself because of some other limita- tion (deitel, ). in our simulation, the scheduler does not directly control the amount of time that a ks has for execution. rather, the size of the window determines how much data a ks has to work with, and this in turn deter- mines how long it will execute before running out of in- put data. with a small window, the effect is that all kss appear to be executing concurrently and no ks can get very much ahead of or behind any other ks. scheduling permits more complex control if priorities are changed dynamically during analysis. changing pri- orities is one way in which the system can incorporate a top-down component in the analysis: on the basis of earlier events in the piece, the system can learn which kss are probably most appropriate, and this can be reflected in the priorities. as we develop the simulation, we will explore more complicated scheduling procedures. for example, the order in which several kss are allowed to modify the blackboard can have important effects if the blackboard information is "nonmonotonic" (rich, ), that is, if blackboard information does not simply accumulate, but can be removed or revised. in this con- text, we may explore other variations in blackboard ar- chitectures (e.g., hayes-roth, ). grouping analysis grouping preference rules and subrules identify possi- ble grouping boundaries in the score on the basis of such qualities as scale distance between notes, duration of time between notes, length of notes, and articulation of notes. our prototype has kss so far for two of the seven group- ing preference rules, which comprise about nine subrules. each subrule routine determines whether it finds a group boundary. if a subrule determines that it has found a boundary between two notes, a grouping node is created and added to the list of boundaries produced by that rule. different grouping preference rules identify different boundaries, with varying amounts of evidence in their favor. this state of affairs must eventually give way to a single well-formed hierarchy of boundaries based on gpr d-note length gpr c - artlculatlon gpr a - pltdl gpr b - attaok point gpr a - blur/relt notochord the strength of the evidence. this distillation will be car- ried out by higher level rules operating on the data struc- ture(s) built by the grouping rules. each grouping ks creates a separate linked list of nodes running parallel to the notochord, one node for each can- didate boundary it has found. each grouping node is linked to the appropriate event in the notochord, as well as to the next grouping node created by the same ks, and to any other nodes produced by other grouping rules that have also identified a boundary at the same event. link- ing all nodes supporting a boundary at a particular event makes it easy for a higher level rule to evaluate all the evidence for a particular boundary. this is illustrated in figure , which shows the grouping analysis for "row, row." the boundary markers, attached to the note fol- lowing the boundary, show results that replicate what we obtained by applying the rules by hand. the appropriate set of weights applied to the rules will select the conjunc- tion of candidate boundaries between notes and (af- ter "stream") as the most important. this boundary di- vides the piece into two half-songs. the boundaries between notes and (after "boat") and between notes and (after the last "merrily") will be selected as the second most important. as noted in the introduction, these boundaries also correspond to our intuitive feeling that the music divides into lines at these points. metric analysis the metric rules of gttm cannot be translated directly into algorithms. that is, whereas the grouping rules them- selves constitute parsing algorithms, the metric rules specify only what the metric hierarchy ought to look like when the analysis is finished. for this reason, we carry out the metric analysis in a single pass with a collection of routines that embody a single algorithm. the output of this algorithm conforms to the requirements ofgttm, but the algorithm itself is not taken directly from the gttm rules. instead, our approach is an adaptation of the grid theory of povel ( ; povel & essens, ). briefly, grid theory works by trying to fit different-sized row. row, row your boat. gen-tly down thestream-- m..- ~ ) mer- rf- iy. mer- rf- iy. mtr- rf- iy, ute ii but q dreclm. figure . grouping analysis of "row, row, row your boat." gpr = grouping preference rule. an expert system for music perception j j. j. ne half-song song ,-. j. j. j jj -rj j'j jj. j ' mmmmj jj jj................................................. figure . metric analysis of "row, row, row your boat." the number of dots indicates the relative rhythmic salience of the notes. grids of equally spaced marks (grid ticks) to the musical events and then determining the degree of stress, or good- ness of fit, between each of the grids and the events. we have modified the procedures of grid theory in several ways. for instance, only those grids suggested by the in- tervals in the music are considered. each grid that fits corresponds to a metric level in the gttm hierarchy. our use of the grid stress is not to identify the one best grid, but rather to eliminate from the hierarchy those levels that least fit the music, that is, levels not heard by a listener. the representation of metric structure is analogous to that outlined above for the score. metric nodes contain information about the metric hierarchy at each beat, as well as pointers to other items and to the notochord. here again, the linked-list structure saves space. if a metric hi- erarchy were represented by a two-dimensional array it would be quite sparse. by allocating a node for each beat rather than for each beat-by-level combination, the linked list of metric nodes wastes no space. the metric analysis of "row, row" is shown in figure . tonality analysis listeners can identify the key of a piece, and this can affect both meter and grouping. for example, strong beats in the metric structure generally fallon the important pitches of the key, as is clear in the "row, row" exam- ple, in which the accents generally coincide with the notes c, e, and g. on a piano keyboard, there are black and white keys in each octave, representing the chromatic scale. in anyone piece of music, some of these notes are more important than others, reflecting the key in which the piece is written. thus, "row, row," which is in the key of c major, contains notes, but only (c, d, e, f, and g) of the possible chromatic notes occur. c occurs eight times, e occurs five times, and g occurs five times. thus, the notes c, e, and g account for of the notes, reflecting the fact that c, e, and g form the c major chord, which is the fundamental (tonic) chord for the key of c major. lerdahl and jackendoff ( ) did not discuss how listeners extract the tonality of a piece, but the example just given suggests a simple algorithm: if we count the frequency of the notes of the chromatic scale, the most common notes will generally identify the key of the piece, as simon ( ) has shown. bharucha ( ) adapted this general idea to a network model, and we have followed a similar approach. in our implementation, each note of the chromatic scale is represented by a pitch node. these pitch nodes are connected to chordnodes; each chord node receives input from just three pitch nodes. finally, sets of three chord nodes (representing the tonic, sulxlominant, and dominant chords of a particular key) are connected to key nodes. as the analysis of a piece progresses, pitch nodes are activated to the degree that their notes occur in the input. these pitch nodes in tum activate the chord nodes, which in tum activate key nodes. the activation level of each node is also affected by a decay parameter, so the activity of a node dies down if input to that node is not sustained, thus defining a different sort of temporal window. the most active key node at any point in the piece defines the perceived key at that point. this tonality extraction mechanism is implemented as a single ks that is triggered by the appearance of new notes in the window. the tonality ks posts the current best hypotheses about the tonality on the blackboard by creating tonality nodes, with each new input event lead- ing to the creation of a new tonality node when this ks is run. discussion our research program addresses basic questions about music perception, but it also illustrates a number of programming techniques that should be of interest to psy- chologists working in many areas, especially in the study of speech perception, reading, and other types of tem- poral pattern perception. first, organizing a simulation in the manner of a multitasking operating system allows the simulation of parallel processing within the context of a highly modular program structure. this permits easy experimentation with a minimum of reprogramming. sec- ond, the use of linked lists permits a data structure that efficiently stores information about temporal patterns in which the number of potential events far exceeds the num- ber of actual events. in effect, linked lists take the sparse- ness out of a sparse matrix. third, an input window is essential for simulating processes in which stimulus pat- terns exceed the span of the psychological present, be- cause a program that has access to all the data is not a reasonable psychological model. finally, the program- ming language must be equal to the task. we chose a powerful, widely known language that supports the com- plex data structures and sophisticated programming tech- niques that are essential in projects of this kind. jones, miller, and scarborough references bharucha, j. j. ( ). musact: a connectionist model ofmusi- cal harmony. in program of the ninthannualconference of the cog- nitive science society (pp. - ). hillsdale, nj: erlbaum. deitel, h. m. ( ). an introduction to operating systems. reading, ma: addison-wesley. dowuxo, w. j., &; harwood, d. l. ( ). music cognition. orlando, fl: academic press. grossberg, s. ( ). how does the brain build a cognitive code? psy- chological review, , i- . hayes-roth, b. ( ). a blackboard architecture for control. artifi- cial intelligence, , - . jacky, i., &; kalet, i. ( ). an object-oriented programming dis- cipline for standardpascal. communications of theacm, , - . lones, i. a., &; harrow, k. ( ). problem solving usingturbo pas- cal. englewood cliffs, ni: prentice-hall. kaplan, h. ( ). design decisionsin a pascal-basedoperant system. behavior research methods, instruments, & computers, , - . lerdahl, f., &; jackendoff, r. ( ). a generative theory of tonal music. cambridge, ma: mit press. mcclelland, i., &; rumelhart, d. ( ). an interactive model of context effects in letter perception: part i. an account of basic find- ings. psychological review, , - . newell, a., &; simon, h. a. ( ). human problemsolving. engle- wood cliffs, ni: prentice-hall. nil, h. ( ). blackboardsystems: the blackboard model of problem solving and the evolution of blackboard architectures. artificial in- telligence magazine, ( ), - . povel, d.-i. ( ). a theoretical framework for rhythm perception. psychalogical research, , - . povel, d.-i., &; essens, p. ( ). perception of temporal patterns. music perception, , - . rich, e. ( ). artificialintelligence. new york: mcgraw-hill. rumelhart, d., &; mcclelland, i. ( ). parallel distributed processing: explorations in the microstructure of cognition. cam- bridge, ma: mit press. schneider, w. ( ). connectionism: is it a paradigm shift for psy- chology?behavior research methods, instruments, & computers, , - . simon, h. a. ( ). perception du pattern musical par auditeur. sciences de l'art, v- , - . tenenbaum, a., &; augenstein, m. ( ). data structures using pas- cal. englewood cliffs, ni: prentice-hall. winston, p. ( ). artificialintelligence ( nd ed.). reading, ma: addison-wesley. wirth, n. ( ). programming in modula-z ( rd ed.). new york: springer-verlag. medicine expert system dynamic bayesian network and ontology based medicine expert system dynamic bayesian network and ontology based octavian arsene ⇑, ioan dumitrache, ioana mihu laboratory of intelligent systems, ‘‘politehnica’’ university of bucharest, spl. independentei , bucharest, romania a r t i c l e i n f o keywords: expert system ontology bayesian network a b s t r a c t the paper proposes an application framework to be used for medicine assisted diagnosis based on ontol- ogy and bayesian network (dbno). there are two goals: ( ) to separate the domain knowledge from the probabilistic information and ( ) to create an intuitive user interface. the framework architecture has three layers: knowledge, uncertainty model and user interface. the contributions of the domain experts are decoupled, the ontology builder will create the domain concepts and relationships focusing on the domain knowledge only. the uncertainty model is bayesian network and the probabilities of the vari- ables states are stored in a profile repository. the diagnostician will use the user interface feeded with the domain ontology and one uncertainty profile. the application was tested on a sample medicine model for the diagnose of heart disease. � elsevier ltd. all rights reserved. . introduction the diagnosis can be defined as the process of identifying a set of hypotheses that model the problem domain and finding that one with highest probability of matching the real world state. in med- ical diagnosis, the uncertainty arises from the inability to evaluate the degree of truth of a hypothesis due to unreliable and incom- plete information or inconsistent knowledge. the ontology and bayesian network (bn) methodologies have been chosen to address knowledge management and uncertainty. the ontology enables the representation of a domain knowledge in a machine understandable form. it can represent the organiza- tional structure of a large complex domains, but the inability to deal with the uncertainty can be a drawback for its application. one disadvantage of the bn is representation of complex struc- tured domains point of view. the ontology and bn can comple- ment themselves in order to overcome the each other disadvantages, thus an ontology-driven uncertainty model can be created. the main goal of the paper is to propose an application frame- work as a collaborative expert system for creating, developing and maintaining a general model for medicine assisted diagnosis (fig. ). from user point of view there will be three roles based on their competencies: concepts and relations definition, connect probabilities to states of the concepts, setting evidences in order to assist the diagnosis. each role is assigned to one of the triangle’ sides and thus depicting three operational layers. the automated connection between all three layers increases the efficiency of the entire process. the proposed model is implemented as a soft- ware using protege as an ontology framework, netica api (application programming interface) as a bayesian api and java technology as a development platform. the mapping between domain knowledge and uncertainty model is based on the fact that each concept defined into ontology is part of the bn as a variable. the diagnostician will use an intuitive graphical user interface for changing evidences of the bn variables and based on a thresh- old the application will depict a chart having the most significant nodes. the final chart will assist the medicine diagnostician to identify the significant factors for a particular case. in order to offer more flexibility the domain knowledge and probabilities are already build and ready to be used as a medicine ontology and respectively uncertainty profile. there is no need to be re-created each time during diagnosis phase. the diagnostician can choose between more than one uncertainty profiles for a medicine ontol- ogy. the relation between ontology and uncertainty profile is one to many (in case there are several sources for probabilities tables for the same ontology). this approach speed up the entire diagno- sis process and allows the asynchronous update of the ontology and uncertainty profiles by the domain experts in order to increase the degree of accuracy of the information. a similar model, ontobayes, was proposed in yi ( ). the major difference between ontobayes and this proposed model resides in separation between domain knowledge and quantitative component of bn in order to decouple the ontology from the uncertainty probabilities. bayesowl (zhongli, ) and pr-owl (cesar, costa, laskey, & laskey, ) are others probabilistic ontology approaches facilitating ontology mapping in the semantic web. some of their limitations refer to: two-valued variables only - /$ - see front matter � elsevier ltd. all rights reserved. doi: . /j.eswa. . . ⇑ corresponding author. e-mail address: octavianarsene@gmail.com (o. arsene). expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/ . /j.eswa. . . mailto:octavianarsene@gmail.com http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa allowed in bayesowl, the ontology model is too complex in pr-owl. an automated bn construction based on an ontology model was proposed in devitt, danev, and matusikova ( ) to be applied in the telecommunication network management domain. it’s more an algorithm model than an application. similar approach was proposed in mihu and arsene ( ) for monitoring and diagnosis of an it infrastructure, all layers (gathering evi- dences, ontology model and presentation layer) were implemented as software agents. section briefly introduces the ontology concept, the knowl- edge representation and the implementation of the medicine ontology using protege application (horridge, knublauch, rector, stevens, & wroe, ). section describes bayesian network concept, the uncertainty representation and the implementation module within the proposed software application. section covers the description of proposed software application. section pre- sents the results of the tests using heart disease data presented in ghosh and valtorta ( ) and ghosh and valtorta ( ). . knowledge representation: ontology the ontology may take a variety of forms, but necessarily it will include a vocabulary of terms and some specification of their mean- ing. the ontology is virtually always the manifestation of a shared understanding of a domain that is agreed between a numbers of parties. such agreement facilitates accurate and effective commu- nication of meaning, which in turn leads to other benefits such as inter-operability, reuse and sharing (uschold, ). the main steps of the methodology for developing ontologies are the follow- ing: ( ) identify purpose and scope, ( ) building the ontology. . . purpose and scope the proposed ontology addresses the medicine domain. it will be built and used by doctors. the purpose on the proposed ontol- ogy is to be knowledge representation of the medicine domain integrated into an application. the doctors will use it in order to define an accurate diagnosis based on the cumulated experience captured into ontology and probabilities tables. . . building the ontology the first step in building ontology is the identification of the key concepts. in the fig. are depicted the key concepts defined within the medicine ontology based on the medical documents and discussions with medical experts. the class is the formal way to represent a concept, e.g. the medicinenode is the root concept of the medicine classes. the two main classes are proposed here: cause and effect. the cause subclasses are considered: environment and heredity. the effect subclasses are: disease, symptom, sign and test. the second step is the identification of the relationships be- tween key concepts that were defined before. we distinguished two kinds of relations between proposed classes (see fig. ): � any cause will have at least one effect, implemented by cause has effect (the inverse relation is: effect of cause); � any effect can generate another effect, implemented by effect gen- erates effect (the inverse relation is: effect generatedby effect). this seems to be odd but let’s take these examples: disease gen- erates symptom or sign generates test. in the fig. is presented the implementation of the relations between concepts in protege. both relations are inverse type based. this kind of relation will be used in bidirectional inference necessary to graphical represen- tation of the bayesian network. different ontology languages provide different facilities. the most recent development in standard ontology languages is owl, endorsed by the world wide web consortium (w c) to promote the semantic web vision. the protege framework is going to be used for creating, maintaining the medicine ontology. protege is a free, open source ontology editor and knowledge-base framework. the protege-owl editor is an extension of protege that supports the web ontology language (owl). the owl ontology may include descriptions of classes, properties and their instances. the owl-dl language is used due to its expressivity and computational effi- ciency. the real life entities from the medicine domain are encoded as classes instantiations within an ontology (called individuals). each individuals has a clear identity that makes them different than others (even they have common attributes). the individuals used in the proposed application are based on the heart disease model presented in ghosh and valtorta ( ), they are depicted in the fig. . the domain expert (a senior doctor) will create the individuals according to his expertise and experience. in the fig. highbloodpresure individual is defined: fig. . architecture of the proposed concept. fig. . hierarchy of the medicine ontology classes defined in protege. fig. . medicine ontology – relations between classes. o. arsene et al. / expert systems with applications ( ) – https://isiarticles.com/article/ human action recognition with sparse classification and multiple-view learning this document is published in: expert systems ( ). ( ), - . doi: http://dx.doi.org/ . /exsy. © wiley publishing ltd. http://e-archivo.uc m.es/ http://dx.doi.org/ . /exsy. human action recognition with sparse classification and multiple-view learning rodrigo cilla, miguel a. patricio, antonio berlanga and josé m. molina applied artificial intelligence group, universidad carlos iii de madrid, madrid, spain e-mail: rodri.cilla@gmail.com;rcilla@inf.uc m.es abstract: employing multiple camera viewpoints in the recognition of human actions increases performance. this paper presents a feature fusion approach to efficiently combine d observations extracted from different camera viewpoints. multiple-view dimensionality reduction is employed to learn a common parameterization of d action descriptors computed for each one of the available viewpoints. canonical correlation analysis and their variants are employed to obtain such parameterizations. a sparse sequence classifier based on l regularization is proposed to avoid the problem of having to choose the proper number of dimensions of the common parameterization. the proposed system is employed in the classification of the inria xmas motion acquisition sequences (ixmas) data set with successful results. keywords: human action recognition, multiple view learning, l regularization . introduction the recognition of human actions has received increasing attention by the computer vision community during the last two decades (lavee et al., ; weinland et al., ). the objective is to build computational models to study how humans move and behave. a wide range of applications such as ambient intelligence (chaaraoui et al., ), video surveillance (foresti et al., ) or human–computer interaction (ren & xu, ) have benefitted from enhancements made in the field. whereas the first systems built for human action recognition were limited to a single camera (cedras & shah, ), advances made in visual sensor networks have motivated the inclusion of multiple cameras to enhance recognition robustness (cilla et al., ). this way, wider scenes can be covered, and systems can deal with occlusions caused by walls and furniture that complicate the recognition from a single camera view. additionally, the perception of the motion from different viewpoints provides supplementary information that facilitates the recognition. a current trend in human action recognition is the efficient combination of observations grabbed from different camera viewpoints. most of the existing approaches are based on recovering a d model of the target of study exploiting multiple-view geometry. visual hulls (weinland et al., ) or d skeletons (parameswaran & chellappa, ) are examples of these models. however, they have some drawbacks: ( ) the need of accurate camera calibration parameters to recover the d model and ( ) large network bandwidth requirements, as many raw data should be sent to a central node where the reconstruction is performed. all these facts make current d reconstruction methods not appropriate for their implementation in visual sensor networks. experimental evidence has shown that common d high- dimensional cues employed in human action recognition, such as silhouettes or optical flow, might be parameterized in low-dimensional manifolds where most of their variance is preserved (blackburn & ribeiro, ; wang & suter, ). at the same time, d descriptors recovered from multiple camera viewpoints, such as visual hulls or skeletons, are also parameterizable in low-dimensional manifolds (turaga et al., ; peng et al., b). both kinds of models contain information about the same real- world phenomena but at different depths, and both share the property of being parameterizable in low-dimensional manifolds preserving most of their variance. a question arising from this fact is if it would be possible to recover a low-dimensional manifold parameterization from multiple d descriptors with a similar performance to that recovered from d descriptors. a design decision when learning low-dimensional manifold parameterizations is the number of dimensions that the learned low-dimensional space should have. there are different proposals to automatically adjust it. variance preservance ratios (jolliffe & myilibrary, ) or automatic relevance determination priors (nounou et al., ) are some popular choices. they adjust manifold dimensionality to preserve the maximum variance employing a reasonable number of parameters. however, these methods do not ensure a good performance when employing the obtained low-dimensional parameterization in a subsequent task, such as action class prediction. in practice, experimental cross-validation is employed to select the right number of dimensions. it has a high computational cost, as many set- ups have to be evaluated to find the best one. an alternative is to select the right number of dimensions in the subsequent task incorporating feature selection to the model with a sparse prior (tibshirani, ). this paper presents a multi-camera human action recognition system taking into account the considerations presented earlier. a feature fusion algorithm is proposed to learn a common manifold representation of feature descriptors extracted from each camera viewpoint. the manifold is learned with a large number of dimensions. a sparse sequence classifier is proposed to select relevant features from the manifold parameterization. this way, the problem of choosing the right number of dimensions for the manifold vanishes. . . purpose and contributions the purpose and contributions of this paper are summarized as follows: • the aim of the paper is to provide a method to classify human actions using descriptors computed from different camera views. • a feature fusion approach is proposed to combine action descriptors extracted from each camera view. canonical correlation analysis (cca) and kernel cca (kcca) are employed to obtain a common manifold param- eterization of the motion descriptors. • to avoid the problem of selecting the manifold dimen- sionality, a sparse hidden conditional random field (shcrf) is introduced for action sequence classification. sparsification is achieved by introducing an l penalty to the objective function optimized during training. • an efficient online learning method is employed to train shcrfs, reducing training time. • the proposed system is evaluated in the recognition of ixmas data set. experimental evidence shows that the proposed method has a performance similar to state-of- the-art proposals employing d models. . . paper organization this paper is organized as follows: section reviews relevant work on human action recognition with a special focus on methods employing multiple camera views. section introduces feature fusion algorithms to recover common manifold parameterizations for the action descriptors extracted from each one of the camera views. section introduces an shcrf model employed to predict human actions while selecting relevant features from the manifold parameterization. experimental results are presented and discussed in section . finally, section summarizes the contributions of this work. . related work . . human action recognition multiple works have been developed to bridge the semantic gap from pixel intensity values in image sequences to descriptions of human actions performed in them. each work defines different steps to solve the correspondence, but, in general, the process might be split in two steps: ( ) feature extraction and representation and ( ) action class prediction. the former step deals with the extraction and efficient encoding of features to describe motions of interest. multiple features might be extracted for motion modelling. parametric models have been fitted to the targets, and temporal moments of the recovered parameter values have been employed for recognition (ribeiro & santos-victor, ). recovering model parameters is a difficult task, so these approaches have been deprecated in favour of visual features. appearance descriptors, such as silhouettes or skeletons, describe how the target looks like (bobick & davis, ). local motion descriptors, such as optical flow, describe the apparent motion of the pixels, providing a very strong cue for action recognition (efros et al., ). the main disadvantage of these methods is their lack of robustness towards partial occlusions of the target. local feature descriptors have been proposed to overcome this limitation, encoding temporal and spatial variations in the neighbourhood of pixels with spatio-temporal saliency properties (laptev, ). the latter step deals with the transformation of features into semantic descriptions. sequence models capture temporal correlations among feature values and employ them to select the appropriate action label. it is possible to apply exemplar-based models with different distances measures to select action labels (efros et al., ; blackburn & ribeiro, ; wang & suter, ), but most of the systems employ probabilistic graphical models. the generative hidden markov model (hmm) is the de facto algorithm for human action recognition (rabiner, ; piccardi & perez, ). however, discriminative graphical models have shown a better performance in action class prediction. the hcrf (quattoni et al., ) has outperformed hmms in multiple action recognition tasks. however, there are many open problems related to the usage of hcrf. this work deals with model and feature selection for the hcrf. . . multi-camera approaches to human action recognition dasarathy’s ( ) input–output data fusion model provides a categorization framework for data fusion systems according to the level of abstraction where system inputs and outputs are defined (data, features and decisions). here, it is employed to organize relevant works to human action recognition from multiple camera viewpoints. data-based levels of the framework are not considered, as the information employed in human action recognition is only defined at the feature and decision levels. diverse methods have been defined at the feature-in feature-out data fusion level to perform human action recognition from multiple camera viewpoints. it is possible to divide the works at this level into three different categories: ( ) projection of d features to d; ( ) feature fusion in a subspace; and ( ) selection of the best view. different d representations might be obtained from projecting d features to d. a popular approach is to project d silhouettes to d to obtain a visual hull representing d appearance (gkalelis et al., ; peng et al., a; pehlivan & duygulu, ). visual hull reconstruction requires accurate silhouette segmentation from the different views. recent works have proposed to project optical flow to d (holte & chakraborty, ) or to project local interest points (holte et al., ). other works recover d star skeletons from d skeleton feature correspondences (chen et al., ). action sketch correspondence across multiple views has been also proposed (yan et al., ). the main drawback of d projection approaches is the need of accurate camera calibration parameters. other methods compute d features for each view and combine them by employing some simple scheme. averaging of multiple features representing pose, global and local motion has been proposed, improving accuracy when compared with other alternatives (maatta et al., ). a joint bag-of-words histogram might be built with local feature descriptors extracted from each one of the camera views (wu et al., ), but a better performance is reported when other fusion strategies are employed. projections maximizing cross-covariance have been learned to combine r -transform derivatives extracted from each camera view (karthikeyan et al., ). two-level linear discriminant analysis has been employed to learn silhouette projections maximizing action class separability (iosifidis et al., ). all these methods provide more flexible solutions for the combination of the features obtained from multiple cameras. however, experimental evidence shows a performance lower than that reported by d projection methods. the last class of methods is based on the computation of a quality measure for each camera view, to perform the recognition employing only the data from the best view. an estimation of the orientation of human with respect to the camera (shen et al., ) or the different properties of the silhouette (määttä & aghajan, ) have been employed to obtain a quality measure. it has been proposed to select the camera with the highest number of detections (wu et al., ) when methods based on interest point location are employed. there are different quality measures for studying saliency, concavity or variations in silhouette stacks (rudoy & zelnik-manor, ). the main drawback of these approaches is that they do not exploit comple- mentary and redundant information obtained from multiple camera views. methods defined at the feature-in decision-out level encode existing correlations between feature descriptors extracted from each camera view and action labels. the concatenation of input features is the most straightforward procedure to perform data fusion in this way (määttä & aghajan, ; wu et al., ). the fused hmm (wang et al., ) proposes modelling correlations among observations coupling the values of the hidden state chains of parallel hmms defined for each camera view. histo- grams of local features are fused, rotating the ordering of the inputs to account for the variations in orientation (srivastava et al., a). the main drawback of these works is their lack of flexibility, assuming that camera configuration remains unchanged between the train and test steps. a procedure to align camera views when the configuration changes from the train to test steps is defined in ramagiri et al. ( ), but it requires to know relative camera placement. decision-in decision-out is the highest level of abstraction. action prediction is performed for each camera view to later combine the results obtained. majority voting is the most common method for the fusion of decisions (määttä & aghajan, ). a weighted voting strategy is proposed in zhu et al. ( ), correcting each vote according to the value of the observed feature. errors produced at each camera view might be incorporated into the voting procedure as proposed by cilla et al. ( ). . feature fusion for human action recognition this section presents feature extraction and feature fusion methods employed in the proposed systems. the d human motion descriptor employed is first introduced. then, cca, a multiple-view dimensionality reduction method employed to find a common manifold param- eterization of the motion descriptors, is described. this section finishes with a presentation of kcca, a non-linear extension to cca. . . feature extraction section has shown that it is possible to employ different visual features in the recognition of human actions. this work employs the action descriptor proposed by tran et al. ( ). it combines motion and appearance infor- mation. it has been selected because it has shown a high experimental performance in predicting the data set that will be employed for system evaluation. it extracted normalizing the bounding box of the observed human to a square box preserving aspect ratio. shape and optical flow are computed from the normalized box. vertical and horizontal planes of the optical flow are split and blurred with a median filter. thus, each box has three channels: silhouette, vertical flow and horizontal flow. the box is divided into four tiles, and a radial -bin histogram is computed from each tile and each channel. the obtained histograms are concatenated to obtain a -d vector. a principal component analysis (pca) reduction of the surrounding past, present and future vectors is appended to generate a descriptor of dtran = dimensions. readers are referred to tran et al. ( ) for more details. . . canonical correlation analysis the objective of cca (hardoon et al., ) is to find a pair of linear projections maximizing the correlation in the projected space between a pair of multivariate random variables. given the zero-mean random variables in the input space x and x with dimensions d and d , cca finds a pair of linear transformations w , w such that one component within each set of transformed variables is correlated with a single component in the other set. the correlation between the corresponding components is called canonical correlation, and there can be at most d = min(d , d ) canonical correlations. the first canonical correlation is defined as r ¼ max w ;w wt x �wt x � �ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi k wt x k � � k wt x k � �q ( ) ¼ max w ;w wt x x t � � w ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi wt x x t � � w w t x x t � � w q ( ) where x x t � � , x x t � � and x x t � � are estimated as eΣ , eΣ andeΣ , respectively, that is, the different minors of the empirical covariance matrix eΣ ¼ eΣ eΣ eΣ eΣ ! of a set of training data x = (x , x ). the remaining canonical correlation directions are orthogonal to w and w , respectively. they are solutions of the generalized eigenvalue problem: eΣ eΣ eΣ eΣ ! w w � � ¼ þ rð Þ eΣ eΣ ! w w � � the standard cca model is defined for only two random variables x and x . bach and jordan ( ) generalize cca to m random variables. the generalized eigenvalue problem to solve is defined as follows: eΣ ⋯ eΣ m ⋮ ⋮eΣm ⋯ eΣmm b@ ca w ⋮ wm b@ ca ¼ l eΣ ⋯ ⋮ ⋮ ⋯ eΣmm b@ ca w ⋮ wm b@ ca where eΣ ⋯ eΣ m ⋮ ⋮eΣm ⋯ eΣmm b@ ca denotes the empirical covariance matrix of a set of training data x = (x , . . ., xm) . . kernel canonical correlation analysis the main limitation of the cca model is given by the linearity of the projections obtained. kernel methods (burges, ) provide a procedure to transform linear algorithms based on inner products of the input data to non-linear algorithms, mapping the input data x to a high- dimensional feature space f(x): f : x ¼ x ; . . . ; xnð Þ ! f xð Þ ¼ f xð Þ; . . . ; fn xð Þð Þ n < nð Þ the linear algorithm (cca in this case) is applied in the transformed feature space. the mapping from the input to feature spaces is not explicitly made. instead, inner products performed by the linear algorithm in the input space are replaced by inner products in the feature space. inner products in the feature space are computed by means of kernel functions. a kernel is a function k such that for all x, z x, k x; zð Þ ¼ < f xð Þ�f zð Þ > ( ) in cca, data are introduced into the algorithm through the empirical covariance matrix eΣ . if the input data x = [x x ] is centred, the minors of eΣ are computed as follows: eΣ ¼ xt x ( ) eΣ ¼ xt x ( ) projection directions wx, wy can be replaced as the projection of the data into directions a and a : w ¼ xt a ( ) w ¼ xt a ( ) equation ( ) can then be rewritten as r ¼ max a ; a at x x t x x t a ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi at x x t x x ta � at x xt x xt a q ( ) let k ¼ x xt and k ¼ x xt be the gram matrices computed from the input data. substituting into equation ( ), r ¼ max a ; a at k k a ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi at k a � at k a q ( ) with this transform, the input data are now introduced into the algorithm through the gram matrices k and k instead of the empirical covariance matrix eΣ. if the gram matrices are obtained using a non-linear kernel function, the cca algorithm now is non-linear. the generalized eigenproblem to solve in order to obtain the projection directions is k k k k � � a a � � ¼ r k k ! a a � � the eigenproblem in equation ( ) has a trivial solution where all the values of r equal to one. that solution is not useful, and it can be avoided with a regularized version of the problem (bach & jordan, ): k k k k � � a a � � ¼ r k þ nk i � � k þ nk i � � bbb@ ccca a a � � where n is the number of samples in the training set, k is a small regularization constant and i is the n � n identity matrix. finally, the generalization of kcca to m input variables is given by where km corresponds to the gram matrix generated from the samples of the input variable m. the kernel function employed in this paper is the gaussian radial basis kernel: k x; zð Þ ¼ e� s x � zk k ( ) where s is a parameter controlling the bandwidth of the gaussian. . sparse sequence classification this section presents the sequence classification algorithm employed to test the performance of the feature fusion method introduced in the previous section. a sparse version of the hcrf model is developed to select relevant features from the result of the feature fusion algorithm. the hcrf model is introduced in the first term, to later present the proposed sparse extension, and the optimization algorithm employed to recover optimal model parameters. . . hidden conditional random fields the hcrf (quattoni et al., ) extends the crf (lafferty et al., ) by introducing hidden state variables into the model. an hcrf is an undirected graphical model composed of three different sets of nodes, as shown in figure . the node y represents the class label for an input sequence. x = x , . . ., xt is the set of nodes corresponding to the temporal observations in the input sequence. h = h , . . ., ht is the set of hidden variables modelling the relationship between the observations xi and the class label y and the temporal evolution of the sequence. the conditional probability of a sequence label y and a set of hidden part assignments h given a sequence of observations x is defined using the hammersley–clifford theorem of markov random fields: p y; hjx; yð Þ ¼ e Ψ y;h;x;yð Þx y x h eΨ y;h;x;yð Þ ( ) where y is the vector of model parameters. hcrfs belong to the general class of log-linear models. the conditional probability of the class label y given the observation sequence x is obtained by marginalizing over all the possible value assignments to hidden parts h: figure : graphical model representation of the hidden conditional random field. k þ nk i � � k k . . . k km k k k þ nk i � � . . . k km ⋮ ⋮ ⋱ ⋮ kmk kmk . . . km þ nk i � � bbbbbbbbbbbb@ cccccccccccca a a ⋮ am bbbbb@ ccccca ¼ l k þ nk i � � . . . k þ nk i � � . . . ⋮ ⋮ ⋱ ⋮ . . . km þ nk i � � bbbbbbbbbbbb@ cccccccccccca a a ⋮ am bbbbb@ ccccca p yjx; yð Þ ¼ x h eΨ y;h;x;yð Þx y x h eΨ y ;h;x;yð Þ ( ) the potential Ψ(y, h, x; y) is a linear function of the input variables: Ψ y; h; x; yð Þ ¼ x i f xið Þ � y hið Þ þ x i y y; hið Þ þ x j; kð Þ e y y; hj; hk � � ( ) the first term, parameterized by y(hi), measures the compatibility of the observation at instant xi with the assignment to the hidden variable hi. the second term measures the compatibility of the hidden part hi with the class label and is parameterized by y(y, hi). finally, the third term models sequence dynamics, measuring the compatibility of adjacent hidden parts hi and hj with the class y. values for model parameters y are estimated from training samples {xi, yi} to maximize the l -regularized conditional likelihood function of the model: l yð Þ ¼ xn i ¼ logp yijxi; yð Þ þ s jjyjj ( ) where the parameter s controls the amount of penalization induced by the l norm of the model parameters. different convex optimization techniques have been proposed to find the optimal parameters y* maximizing the conditional likelihood function in equation ( ). among them, limited memory - broyden–fletcher–goldfarb–shanno (l-bfgs) (liu & nocedal, ) and conjugate gradient (nocedal & wright, ) are the most popular. inference of the posterior probability distribution in equation ( ) and auxiliary probability distributions needed for the estimation of the conditional likelihood gradient in equation ( ) is made, using belief propagation as proposed in quattoni et al. ( ). . . l -regularized hidden conditional random fields the l -regularized training of the hcrf in equation ( ) produces solutions y * where all the components have a small value. this l regularization approach has some drawbacks from the computational learning perspective: • there is no feature selection during training. irrelevant features at input sequences are given a non-zero weight caused by the nature of the gradient of the l norm, producing model overfitting. • no model selection is performed. occam’s razor principle of machine learning stands that the best model is the one with a lower complexity best adapting to training data. a consequence of not fulfilling this requirement is overfitting. similarly to the previous case, parameters y(y, hi) and y(y, hj, hk) corresponding to unnecessary hidden parts obtain a non-zero weight in the hcrf result. in practice, the right number of hidden parts is selected by means of cross-validation, requiring to test multiple configurations to find the best one. a possible way to overcome these limitations is to replace the l penalty term in equation ( ) by an l penalty term. this way, the objective function to minimize for parameter estimation has the form: l yð Þ ¼ xn i ¼ logp yijxi; yð Þ þ s jjyj j ( ) the l norm causes components in y to have a zero value if they are not needed. if a zero value is given to a parameter y(hi), then the corresponding input feature is not taken into account, producing a feature selection effect. if the zero value is given to a parameter y(y,hi) or y(y,hj,hk), then model selection is performed, reducing the complexity of the model. although the l regularization has been employed to estimate the optimal parameters of log-linear models such as crfs (lafferty et al., ), this is, to the best of our knowledge, the first time that it has been employed to train hcrfs. unfortunately, the l norm has the property of non- smoothness at , and conventional convex optimization techniques are no longer valid to recover optimal models parameters. different approaches have been proposed for the optimization of l -regularized log-linear models. a first approach is to reparameterize the model to transform the optimization problem into a smooth one (vail et al., ). model parameters are split in a pair of vectors y = y+ � y�, such as y+ > and y� > . standard convex optimization techniques are then applied to find the optimal parameters y*. unfortunately, the convergence speed of the method is very slow requiring many iterations until an optimal solution is found. an alternative to this method is to perform stochastic gradient descent (tsuruoka et al., ). stochastic gradient descent updates model parameters after presenting each sample. although the obtained solutions are worse than those obtained using conventional methods, in practice, they are good enough to evaluate the performance of the feature fusion approaches proposed in the previous section. the algorithm performs as follows: y k þ i ¼ @ki þ �k @l j; yk � � @yki ( ) yk þ i ¼ max ; y k þ i � uk þ qk � i � � @ a if yk þ i < min ; y k þ i þ uk � qk � i � � @ a if yk þ i > >>>>>>>< >>>>>>>: ( ) where qki is the total l penalty that yi has received up to the point: qki ¼ xk t ¼ yt þ i � y t þ i ( ) �k controls the learning rate and at each iteration is decreased by an exponential decay: �k ¼ � a k n ( ) where � and a are constants and n is the number of training samples. . experiments . . experimental set-up algorithms presented in previous sections are going to be evaluated in the prediction of ixmas data set (weinland et al., ). it contains actions performed by different actors at least three times each. the actions are recorded from five different viewpoints. the actions are ( ) check watch; ( ) cross arms; ( ) scratch head; ( ) sit down; ( ) get up; ( ) turn around; ( ) walk; ( ) wave; ( ) punch; ( ) kick; and ( ) pick up. a sample frame of the data set observed from the five viewpoints is presented in figure . the protocol to evaluate system performance is the leave- one-actor-out cross-validation. the data set is divided into different sets according to the actor performing each sample. the system is trained using all sets except one used for testing. the procedure is repeated until every actor has been used for testing. the cca and kcca are configured to provide projection directions corresponding to the highest canonical correlations. it should be emphasized again that a high value is given to this parameter as the shcrf will select the relevant projections during training. kcca is employed with a radial basis kernel of width s = . . the shcrf has been set up with |h| = hidden parts. a value s = . is given to the sparsity penalty term. the training algorithm is run for iterations with a batch size of samples. the exponential decay of the learning rate � is configured with an initial value � = . and a decay rate a = . . a monte carlo scheme is employed to evaluate the action prediction performance as the objective function employed to train the shcrf is non-convex, producing different solutions depending on the initial value y . the accuracy obtained for each test set is averaged over different runs of the shcrf training algorithm to obtain a real estimate of the algorithm performance. finally, to measure the improvement produced by feature fusion in the predictive performance, a baseline model is going to be employed. the descriptors computed for cameras and are independently projected using pca to retain the most significant dimensions. the shcrf is respectively evaluated using this projected sequences with the same procedure presented earlier. . . results and discussion . . . feature fusion algorithms to visualize the effect of the feature fusion algorithms, the first three most significant features obtained by them and the pca baseline are shown in figures (a)– (f). projections have been obtained with the data of actors – . it can be observed that fused features have a stronger class structure than features obtained by the pca baselines, although they do not seem to be separated in any case. this is something normal in action recognition domains, as different action sequences usually share common frames, being their temporal evolution the real discriminative factor for action class prediction. another phenomenon observed in the plots of the linear models (figures (a), (b), (c) and (e)) is that the main cause of class variation is produced by the direction and amount of movement. it can be seen that the variation of the actions sit down, get up and pick up, involving a vertical movement, seems to have a stronger structure than the others. however, in the features from the non-linear models presented in figures (d) and (f), that behaviour is not present. instead, class structure is appreciated for most of the actions. it seems that non-linear algorithms model other hidden factor of variations distinct from the movement direction. . . . action classification table shows the accuracy obtained by the shcrf predicting class labels for the ixmas data set. for each evaluation fold, the worst, median and best results have been extracted. results show that predicting action classes with fused features has an accuracy higher than predicting them with only one camera, although for cca the worst case is worse than the classification with pca . (a) camera (b) camera (c) camera (d) camera (e) camera figure : the kick action in the ixmas data set from the five available views. other interesting fact is that cca and kcca have almost the same accuracy when employing the data from the five camera viewpoints, whereas in the two cameras scenario, kcca performs better than cca. the feature fusion algorithms employed here extract common information in the data from each camera, so this saturation phenomenon seems reasonable. a non-linear algorithm extracts common information better than the linear algorithm, but as the number of data sources increases, the gap is reduced until they have a similar performance. to better understand the behaviour of the presented algorithms, figure presents the box plots of accuracies obtained evaluating each actor following a monte carlo scheme. it should be noted that the feature fusion accuracy does not increase in all the cases reported. the most striking is when evaluating actor (figure (b)). the disastrous performance of camera makes the fusion perform worse in all the cases than when using only camera . the cause for this phenomenon is that the fusion algorithms assume that every data source has a similar quality, not weighting them according to some quality metric. the phenomenon is produced for other actors too (figures (h) and (i)), although there is not a big difference in the results using a single camera. that might be produced because the sources have complementary information that are not well represented in the transformed space. (a) camera pca (b) camera pca (c) cca cameras - (d) kcca cameras - (e) cca cameras - - - - (f ) kcca cameras - - - - figure : projections of the first three components obtained with the different dimensionality reduction algorithms. cca, canonical correlation analysis; kcca, kernel cca. table : worst, average and best case accuracy estimations of the different methods obtained using monte carlo leave-one-actor-out cross-validation pca pca cca kcca cca kcca minimum . . . . . . median . . . . . . maximum . . . . . . finally, table compares the result of the presented proposal to others. the best case result performs as good as d proposals using visual hulls, whereas the average case improves results from decision-in decision-out. . conclusions this work has proposed the usage of multiple-view learning as a feature fusion method for human action recognition from multiple cameras. it has been shown that the usage of the information shared by rich d motion descriptors computed from multiple camera viewpoints improves the predictive accuracy. the usage of an l -regularized sequence classifier has avoided the manual choice of the number of dimensions of the projected space. testing multiple configurations to find the best dimension has been avoided, saving computational time. the usefulness of the proposed systems has been shown by predicting ixmas data set with an accuracy similar to that reported by state-of-the-art methods. references bach, f.r. and m.i. jordan ( ) kernel independent component analysis, journal of machine learning research, , – . blackburn, j. and e. ribeiro ( ) human motion recognition using isomap and dynamic time warping, in proceedings of the nd conference on human motion: understanding, modeling, capture and animation, rio de janeiro, brazil: springer-verlag, – . bobick, a.f. and j.w. davis ( ) the recognition of human movement using temporal templates, ieee transactions on pattern analysis and machine intelligence, , – . burges, c.j.c. ( ) advances in kernel methods: support vector learning, the mit press. cedras, c. and m. shah ( ) motion-based recognition: a survey, image and vision computing, , – . chaaraoui, a.a., p. climent-pÉrez and f. flÓrez-revuelta ( ) a review on vision techniques applied to human behaviour analysis for ambient-assisted living, expert systems with applications, ( ), – . chen, d., p.c. chou c.b. fookes and, s. sridharan ( ) multi- view human pose estimation using modified five-point skeleton model. in international conference on signal processing and communication systems , – dec , gold coast, australia. cilla, r., m.a. patricio, a. berlanga and j.m. molina ( ) a probabilistic, discriminative and distributed system for the recognition of human actions from multiple views, neurocomputing, ( ), – . dasarathy, b.v. ( ) sensor fusion potential exploitation- innovative architectures and illustrative applications, proceedings of the ieee, , – . efros, a.a., a.c. berg, g. mori and j. malik ( ) recognizing action at a distance. ieee international conference on computer vision, , – . foresti, g.l., c. micheloni and l. snidaro ( ) event classification for automatic visual-based surveillance of parking lots, in pattern recognition, . icpr . proceedings of the th international conference on, vol. , – , ieee. gkalelis, n., h. kim, a. hilton, n. nikolaidis and i. pitas (november ) the i dpost multi-view and d human action/interaction database. conference for visual media production, – . hardoon, d.r., s. szedmak and j. shawe-taylor ( ) canonical correlation analysis: an overview with application to learning methods, neural computation, , – . holte, m.b. b. chakraborty, j. gonzalez and t. b. moeslund ( ) a local d motion descriptor for multi-view human action recognition from d spatio-temporal interest points, selected topics in signal processing, ieee journal of, ( ), – . holte, m.b., t.b. moeslund, n. nikolaidis and i. pitas (may ) d human action recognition for multi-view camera systems. international conference on d imaging, modeling, processing, visualization and transmission, – . iosifidis, a., a. tefas, n. nikolaidis and i. pitas ( ) multi- view human movement recognition based on fuzzy distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (a) actor (b) actor (c) actor (d) actor (e) actor (f ) actor (g) actor (h) actor (i) actor (j) actor figure : results for the different evaluation actors. a is the pca from camera . b is the pca from camera . c is the canonical correlation analysis (cca) fusion of cameras and . d is the kernel cca (kcca) fusion of cameras and . e is the cca fusion of the five cameras. f is the kcca fusion of the five cameras. table : comparison of the accuracy of our method to others method accuracy type srivastava et al. ( b) . decision-in decision-out our average . d feature-in d feature-out weinland et al. ( ) . d feature-in d feature-out our best . d feature-in d feature-out peng et al. ( b) . d feature-in d feature-out and linear discriminant analysis, computer vision and image understanding, , – . jolliffe, i.t. ( ) principal component analysis. springer verlag. karthikeyan, s., u. gaur, b.s. manjunath and s. grafton (november ) probabilistic subspace-based learning of shape dynamics modes for multi-view action recognition. ieee international conference on computer vision workshops (iccv workshops), – . lafferty, j., a. mccallum and f. pereira ( ) conditional random fields: probabilistic models for segmenting and labeling sequence data, in icml ’ proceedings of the eighteenth international conference on machine learning, – . laptev, i. ( ) on space-time interest points, international journal of computer vision, , – . lavee, g., e. rivlin and m. rudzsky ( ) understanding video events: a survey of methods for automatic interpretation of semantic occurrences in video, ieee transactions on systems, man and cybernetics - part c: applications and reviews, , – . liu, d.c. and j. nocedal ( ) on the limited memory bfgs method for large scale optimization, mathematical programming, , – . mÄÄttÄ, t. and h. aghajan ( ) on efficient use of multi-view data for activity recognition, – . mÄÄttÄ, t., a. hÄrmÄ and h. aghajan ( ) on efficient use of multi-view data for activity recognition, in proceedings of the fourth acm/ieee international conference on distributed smart cameras, (pp. – ), acm. nocedal, j. and s.j. wright ( ) numerical optimization, springer verlag. nounou, m.n., b.r. bakshi, p.k. goel and x. shen ( ) bayesian principal component analysis. journal of chemometrics, , – . parameswaran, v. and r. chellappa ( ) view invariance for human action recognition, international journal of computer vision, , – . pehlivan, s. and p. duygulu ( ) a new pose-based representation for recognizing actions from multiple cameras, computer vision and image understanding, , – . peng, b., g. qian and s. rajko (august a) view-invariant full-body gesture recognition via multilinear analysis of voxel data, third acm/ieee international conference on distributed smart cameras (icdsc), – . peng, b., g. qian and s. rajko (september b) view-invariant full-body gesture recognition via multilinear analysis of voxel data, third acm/ieee conference on distributed smart cameras. piccardi, m. and o. perez ( ) hidden markov models with kernel density estimation of emission probabilities and their use in activity recognition. ieee conference on computer vision and pattern recognition (cvpr’ ), – . quattoni, a., s. wang, l.-p. morency, m. collins and t. darrell ( ) hidden conditional random fields, pattern analysis and machine intelligence, ieee transactions on, , – . rabiner, l.r. ( ) a tutorial on hidden markov models and selected applications in speech recognition, proceedings of the ieee, , – . ramagiri, s., r. kavi and v. kulathumani (august ) real- time multi-view human action recognition using a wireless camera network, fifth acm/ieee international conference on distributed smart cameras, – . ren, h. and g. xu ( ) human action recognition in smart classroom, in automatic face and gesture recognition, . proceedings of the fifth ieee international conference on, – , ieee. ribeiro, p.c. and j. santos-victor ( ) human activity recognition from video: modeling, feature selection and classi- fication architecture, in international workshop on human activity recognition and modeling (harem). rudoy, d. and l. zelnik-manor ( ) viewpoint selection for human actions, international journal of computer vision, , – . shen, c., c. zhang and s. fels ( ) a multi-camera surveillance system that estimates quality-of-view measure- ment. ieee international conference on image proce- ssing, iii– –iii– . srivastava, g., h. iwaki, j. park and a.c. kak (august a) distributed and lightweight multi-camera human activity classification, third acm/ieee international conference on distributed smart cameras (icdsc), – . srivastava, c., h. iwaki, j. park and a.c. kak (september b) distributed and lightweight multi-camera human activity classification, in third acm/ieee conference on distributed smart cameras, – . tibshirani, r. ( ) regression shrinkage and selection via the lasso. journal of the royal statistical society. series b (methodological), ( ), – . tran, d., a. sorokin and d. forsyth ( ) human activity recognition with metric learning, in proceedings of the th european conference on computer vision: part i, springer berlin heidelberg: springer-verlag, . tsuruoka, y., j. tsujii and s. ananiadou ( ) stochastic gradient descent training for l -regularized log-linear models with cumulative penalty, in proceedings of the joint conference of the th annual meeting of the acl and the th international joint conference on natural language processing of the afnlp: volume -volume , association for computational linguistics, – . turaga, p., a. veeraraghavan and r. chellappa ( ) statistical analysis on stiefel and grassmann manifolds with applications in computer vision, in computer vision and pattern recognition, . cvpr . ieee conference on, – , ieee. vail, d.l., j.d. lafferty and m.m. veloso ( ) feature selection in conditional random fields for activity recognition, in intelligent robots and systems, . iros . ieee/rsj international conference on, – , ieee. wang, l. and d. suter ( ) visual learning and recognition of sequential data manifolds with applications to human movement analysis, computer vision and image understanding, , – . wang, y., k. huang and t. tan ( ) multi-view gymnastic activity recognition with fused hmm, in computer vision- accv (pp. – ). berlin heidelberg: springer. weinland, d., r. ronfard and e. boyer ( ) free viewpoint action recognition using motion history volumes, computer vision and image understanding, , – . weinland, d., r. ronfard and e. boyer ( ) a survey of vision-based methods for action representation, segmentation and recognition, computer vision and image understanding, , – . wu, c., a.h. khalili and h. aghajan ( ) multiview activity recognition in smart homes with spatio-temporal features, proceedings of the fourth acm/ieee international conference on distributed smart cameras - icdsc’ , . yan, p., s.m. khan and m. shah (june ) learning d action feature models for arbitrary view action recognition. ieee conference on computer vision and pattern recognition, – . zhu, f., l. shao and m. lin ( ). multi-view action recognition using local similarity random forests and sensor fusion. pattern recognition letters, ( ), – . the authors rodrigo cilla rodrigo cilla received his bs, ms and phd degrees in computer science from the universidad carlos iii de madrid in , and , respectively. his research interest includes, among others, human action recognition, multiple target tracking, high-dimensional data analysis, data fusion and biological image processing. miguel a. patricio miguel a. patricio received his bs and ms degrees in computer science and his phd degree in artificial intelligence from the universidad politécnica de madrid in , and , respectively. he has held an administrative position at the computer science department of universidad politécnica de madrid since . he is currently an associate professor at the escuela politécnica superior of the universidad carlos iii de madrid and a research fellow of the applied artificial intelligence group (giaa). he has carried out a number of research projects and consulting activities in the areas of automatic visual inspection systems, texture recognition, neural networks and industrial applications. antonio berlanga antonio berlanga received the ba degree in physics from the universidad autónoma de madrid, spain, and the phd degree in computer engineering from the universidad carlos iii de madrid in and , respectively. he is currently an associate professor with the department of computer science, universidad carlos iii de madrid. his current research interests include evolutionary computation for multi-objective optimization, machine learning and data mining. josé m. molina josé m. molina is a full professor at the universidad carlos iii de madrid. he joined the computer science department of the universidad carlos iii de madrid in . currently, he leads the applied artificial intelligence group (giaa). his current research focuses on the application of soft computing techniques (neural network, evolutionary computation, fuzzy logic and multi-agent systems) to radar data processing, air traffic management and e-commerce. he is the author of more than journal papers and conference papers. he received his bs and phd degrees in telecommunications engineering from the universidad politécnica de madrid in and , respectively. variable neighborhood search heuristics for a test assembly design problem jordi pereiraa,b,∗, mariona vilàb adepartamento de ingenieŕıa industrial, universidad católica del norte, av. angamos , antofagasta, chile bescola universitària d’enginyeria tècnica industrial de barcelona, universitat politècnica de catalunya. c. comte urgell, st. floor, , barcelona, spain abstract test assembly design problems appear in the areas of psychology and educa- tion, among others. the goal of these problems is to construct one or multiple tests to evaluate some criteria. this paper studies a recent formulation of the problem known as the one-dimensional minimax bin-packing problem with bin size constraints (minimax bsc). in the minimax bsc, items are initially divided into groups and multiple tests need to be constructed using a single item from each group, while minimizing the difference among the tests. we first show that the problem is np-hard, which remained an open question. second, we propose three different local search neighborhoods derived from the exact resolution of special cases of the problem, and combine them into a variable neighborhood search (vns) metaheuristic. finally, we test the proposed al- gorithm using real-life-based instances. the results show that the algorithm is able to obtain optimal or near-optimal solutions for instances with item pools with up to . items. consequently, the algorithm is a viable option to de- sign large-scale tests, as well as to provide tests for online small-sized situations such as those found in e-learning platforms. keywords: bin packing, test assembly design, test splitting, variable neighborhood search. . introduction test assembly design studies the problem of selecting the optimal subset of items (questions) among those available to define one or several tests (question- naires) subject to specific constraints and objectives. these problems appear in multiple areas, specifically in test design (van der linder, ) both for edu- ∗corresponding author email addresses: jorgepereira@ucn.cl, jorge.pereira@upc.edu (jordi pereira), mariona.vila.bonilla@upc.edu (mariona vilà) preprint submitted to expert systems with applications january , cation (porter et al., ; sun et al., ) and psychology studies (veldkamp, ). based on the specific objective of the test, as well as the characteristics of the questions and questionnaires, different problems and formulations may appear. one of these formulations arises when considering the problem as a packing problem, specifically as a one dimensional bin packing problem (bpp) with additional constraints (see dyckhoff ( ) for a classification of packing problems, and wäscher et al. ( ) for an update of the previous classification). in the bpp formulation, the questions and questionnaires are interpreted as the items and the bins of the bpp, respectively, and the weight of the items corresponds to some metric (e.g., the difficulty of the question) that needs to be considered during the test design. the objective of the problem is to find disjoint subsets of all the items so that the sum of weights (i.e., the total difficulty) of each subset is as evenly matched as possible. in this work we consider the formulation introduced in brusco et al. ( ). this formulation is known as the one-dimensional minimax bin-packing problem with bin size constraints (minimax bsc). to solve the problem, the authors propose a mixed zero-one integer linear programming model that is then solved using the cplex commercial software. for large real-life problems, the authors propose a simulated annealing (sa) algorithm that outperforms the quality of the solutions provided by cplex. while brusco et al. ( ) do not address the complexity of the problem, the analysis of the results showed that all of the algorithms proposed in the paper were able to optimally solve large problems with questionnaires, leading the authors to hypothesize that the special case of questionnaires may be solvable in polynomial time (brusco et al., , p. ). furthermore, the quality of the solutions provided by the sa deteriorates as the number of questionnaires in- creases, which may identify possible improvements if different solution methods were used. . . contributions of this work in this paper, we address the complexity of the minimax bsc and demon- strate that the minimax bsc is np-hard when two questionnaires are to be constructed and strongly np-hard when more than two questionnaires are re- quired. we also show that the case with two sets of questions can be optimally solved in polynomial time and that the case with two questionnaires can be efficiently solved as a subset sum knapsack problem (kellerer et al., ). based on these special solvable cases, we derive a new constructive heuristic and three local search neighborhoods. these methods are then combined in a variable neighborhood search (vns) metaheuristic (mladenović & hansen, ) to provide a combined approach to solve the problem. vns is a popu- lar metaheuristic approach (mladenović & hansen, ; hansen et al., ) based on the use and exploration of different local neighborhoods of an incum- bent solution and a mechanism to escape from local optima by restarting the search on random neighbor solutions. both mechanisms provide a method to systematically explore the solution space. the results of the method clearly outperform those provided by previous algorithms, such as the sa method proposed in brusco et al. ( ). more specifically, the proposed method is able to routinely reach a bpp-based lower bound in instances with a large number of questions per questionnaire (i.e., over ten questions per questionnaire) with characteristics of real-life problems, and still manages to obtain small deviations from the theoretical bound in other cases. . . paper outline the remainder of the paper is structured as follows. in section , we study the problem, describe its mathematical formulation, address the issue of com- plexity and identify special cases that can be efficiently solved. we also outline some of the work in the literature devoted to test design and other related prob- lems, such as the bpp. in section , we describe the proposed local searches and a constructive heuristic. we also propose their combination in a vns-based method. in section , we describe the results of a computational experiment to assess the quality of the proposed algorithm. finally, in section , we provide conclusions of the present work and identify some future areas of research. . problem description . . literature review test assembly design was early identified as a combinatorial optimization problem. according to van der linden (van der linden, ), an initial work in the field (birnbaum, ) proposed a three step process to design and con- struct tests: ( ) identification of the goal of the test under construction; ( ) identification of a target function in accordance with the goal; and ( ) selection of items based on the target function and the fulfillment of some constraints. these steps clearly mimic the formulation and resolution of any other combina- torial optimization problem. these three steps are preceded by a preliminary one in which the item pool is constructed. item pool construction requires the estimation of the information function of the items, a step that is performed using and item response theory (irt), see veldkamp ( ) for a basic description and a critique of irt, and lu ( ) for a description of an estimation method. the information function is then discretized into specific ability levels of interest (such as the pass/fail ability level for an examination test) that provide coefficients for the objective and/or the constraints of the combinatorial optimization model. the present paper only considers the item selection process. even when we limit our attention to the selection step, the literature of the field is very broad. we can divide previous work into three groups: ( ) offline single test construction, in which the objective is to design a single test, which is com- pletely constructed before its use; ( ) online single test construction, in which the objective is to design a single test, which is sequentially constructed using the information provided by the answers to previous items; and ( ) multiple test construction, in which multiple tests are constructed together in order to evaluate examinees with different, but equivalent, test forms. note that a method to solve any of these problems can be used to solve other problems after some modifications, and consequently we offer a review of recent procedures for each of these three methods. offline single test construction is known in the literature as static test gen- eration (nguyen & fong, ), or automated test assembly (veldkamp, ). a general reference on the area (van der linder, ) provides a general in- teger programming formulation for the problem. the proposed objective is to maximize the minimum amount of discrimination index on any ability level of interest, while satisfying some constraints on the different requirements of the test. veldkamp ( ) considers the model in van der linder ( ) under ro- bustness considerations on coefficients provided during item pool construction. the effect of using a robust formulation is then verified by constructing a test using a items pool, and comparing the solutions provided by both models. the robust solution is shown to be less sensitive to errors in the evaluation of the information function. nguyen & fong ( ) identifies the problem as a multidimensional knap- sack problem (kellerer et al., ), in which the test tries to maximize the discrimination degree of the test (how good the question is at recognizing user proficiency) under constraints on number of questions, time to perform the test, average difficulty and number of questions per topic. the model is then solved using a branch and cut based procedure and the solutions are compared to previous methods found in the literature on instances with up to . items. online single test construction corresponds to the construction of the test along with its resolution and it is usually known in the literature as cat (com- puterized adaptive test). due to the variable nature of the problem, the methods are usually constructive and item selection makes use of greedy rules. further- more, cats try to take into account that multiple tests are to be generated and, thus some randomization is introduced to avoid excessive use of a subset of questions. in he et al. ( ), four different item-selection methods are evaluated. these methods take into account multiple particularities, such as side con- straints, content specifications, time requirements, or item formats, among oth- ers. the heuristics are compared in terms of their use of items (how good are they providing multiple different tests) and their accuracy to measure the information function. edmonds & armstrong ( ) considers a hybrid between offline and online single test construction, denoted as multiple stage adaptive test (mst). the hybrid divides items into groups, each belonging to a different stage of the test. when an item is to be selected, the method randomly takes an item from the corresponding group by taking into account the stage and the level shown by the examinee on previous items. the definition of each group is modeled and solved using a commercial solver (cplex) for a . items pool. while irt is the predominant method to evaluate the information provided by the items, alternatives exist. in smits & finkelman ( ) a health-oriented alternative is put forward. the objective is now to minimize the time required to administer a diagnosis. consequently, the method minimizes the number of questions required to identify the expected total score of the test. the method applies an ordinal regression on the previous answers in order to stop admin- istering questions once a required degree of precision on the final result of the test has been reached. the applicability of the model is then tested on a real data set provided by a screener for depression. multiple test construction, or parallel test design, is used when multiple in- terchangeable tests need to be constructed. an example in which multiple tests are needed is the evaluation of different candidates at different time frames, or the assignment of different tests to different students in order to avoid cheating. a basic multiple test construction method builds tests sequentially until the desired number of tests have been generated. chen et al. ( ) proposes one of these sequential methods. the method first divides the item pool into groups (referred as cells) containing questions with similar characteristics. then, a specified number of items from each group are randomly selected in order to generate different tests in each execution. chen ( ) provides an improved method that incorporates a content-balancing element similar to those found on cat test construction methods. both works use a real items pool to test their proposed method. chang & shiu ( ) proposes a method to construct multiple tests based on clonalg, an algorithm based on the clonal selection principle in a biological immune system. the authors report results with a . items pool used to construct parallel tests. hwang et al. ( ) proposes a tabu search approach that generates multiple tests subject to time constraints and identical number of questions per test. the method was tested in simulated . items pools, which represents the largest sized item pool used in the literature. the previous references try to focus on constructing a predefined number of tests while minimizing deviations among the tests and/or maximizing the dis- crimination power of each test. another possible objective aims at maximizing the utilization of the item pool, by constructing the maximum number of tests under some constraints on the repeated use of each item. songmuang & ueno ( ) consider a mixed problem in which both the number of constructed tests and quality of each individual test are considered. the problem is then solved using a bee algorithm. the computation is per- formed using parallelization and multiple tests are obtained for item pools with up to . items. ishii et al. ( ) considers the maximization of the number of tests using a maximum clique formulation. their work builds upon the study by belov & armstrong ( ), in which the problem is assimilated to a set packing problem and solved using a constraint programming approach. the proposal of ishii et al. ( ) constructs a graph in which a maximal subset of tests that share none or few items (the clique) is sought. the method first constructs multiple tests with the required conditions like the discrimination power and number of items used (see also belov, ). the tests define the vertices of the graph and edges among vertices are included if the number of shared items is deemed small. once the graph has been built, a classical method to ascertain the maximum clique is used. a different approach, which also links test assembly design to a classical combinatorial optimization problem was proposed in brusco et al. ( ). in this case, the problem is formulated as a constrained bin packing problem (bpp) and it is named the one-dimensional minimax bin-packing problem with bin size constraints (minimax bsc). the authors consider that the item pool has been previously divided into groups of items with similar characteristics, as seen in van der linden & boekkooi-timminga ( ), chen et al. ( ) or chen ( ), and then they consider the generation of multiple tests each containing an item from each group. as pointed out by van der linden & boekkooi-timminga ( ) and brusco et al. ( ), the division of items into groups has multiple advantages as it can be used to obtain tests with similar composition as well as to enforce constraints which may be difficult to specify. the objective of the problem is then to obtain tests with similar discrimination index on the ability level of interest. the authors solve the problem using a commercial software (cplex) and a simulated annealing approach on item pools with up to . items. in this paper, we extend the work of brusco et al. ( ) by considering its complexity and by providing a new efficient solution method for the min- imax bsc. the computational experiment shows that we are able to solve instances with up to . -item pools within less than a minute in a modern commodity computer, and the solutions obtained show very small optimality gaps when compared to a theoretical lower bound on the maximum discrimi- nation index. for smaller sized instances, the method provides solutions of the same quality within seconds. these results open the door to their use in on- line single test construction methods using the shadow test approach (he et al., ). the shadow test approach is a online single test construction method in which a complete test is constructed in each cat decision step and the final item used is selected from the constructed test. additionally, this work also verifies the applicability of the proposed method on different conditions to those proposed in brusco et al. ( ). according to brusco et al. ( ), items are grouped according to their discrimination indexes. we conduct an experiment in which items are randomly grouped (so we can consider that they have been grouped by subject topic or some other constraint). the results of the additional experiment show that the method is still capable of obtaining optimal or near-optimal solutions, highlighting the possible application of this method to other multiple test construction problems in which the number of tests is known in advance and the constraints can be adapted into the multiple initial groups method required by the minimax bsc formulation. as the problem under study is a special case of the bpp, we also provide a brief bibliography on the methods available for packing problems. we refer the reader to the classifications of packing problems provided in dyckhoff ( ); wäscher et al. ( ) and the references included in these papers for a more extensive and comprehensive literature review. we will only highlight that mul- tiple approaches have been proposed for the resolution of the one-dimensional case, both exact (i.e., usually based on column generation kallrath et al., ; brandão & pedroso, ) and heuristic, such as genetic algorithms (quiroz- castellanos et al., ), variable neighborhood search (fleszar & hindi, ; m’hallah et al., ) or hyperheuristics (lópez-camacho et al., ) methods. our proposal belongs to the second group and its more distinguishing fea- ture is the use of special neighborhood structures, rather than the common swap or exchange neighborhoods. the proposed neighborhood structures allow us to explore larger neighborhoods and to find the best move in these neighborhoods within very short running times. the neighborhood structures are then com- bined in a variable neighborhood search (vns) method, and allow us to obtain optimal or near-optimal solutions in very short running times. while the pro- posed method is not directly applicable to other bin packing problems, it offers an example on how to apply the resolution of special cases of the problem into the definition of effective heuristics. . . mathematical formulation, lower bounds and special cases the minimax bsc considers the problem of sorting a number of questions into questionnaires of equal length (i.e., each is composed of the same number of questions). the questions are initially divided into different homogeneous sets, and a solution corresponds to the assignment of one question from each set to each questionnaire so that all questions are assigned, and some measure of fairness among the resulting questionnaires is optimized (e.g., minimizing the maximum difference between the total difficulty of the questions assigned to two different questionnaires). these constraints (e.g., multiple sets and assigning a question of each set to each questionnaire) are particularly common in applications pertaining to the splitting of a set of items to form parallel tests of equal length. for example, in the research of van der linden & boekkooi-timminga ( ), a first stage divides the questions into homogenous sets. then, a second stage tries to find a solution in which questionnaires may only use one question from each set. while the first and second stage problem could be solved as a single problem, the first stage allows for the inclusion of additional constraints, (e.g., the grouping of questions with comparable difficulty or similar content) and avoids the selection of highly unbalanced solutions (e.g., one test is only composed of easy and difficult questions, while another contains a combination of moderately difficult questions). the formulation proposed in brusco et al. ( ) is described as follows: t disjoint sets ( ≤ t ≤ t) of questions exist, and each set is composed of b questions ( ≤ r ≤ b) with weights wrt ( ≤ r ≤ b, ≤ t ≤ t). these weights represent the discrimination index of the question (i.e., the percentage of recipients that are capable of correctly answering the question). the questions need to be grouped into b ( ≤ b ≤ b) groups (i.e., questionnaires) in such a way that each questionnaire contains exactly one item from each one of the sets; and the total weight of the questions assigned to a questionnaire identifies the figure : example of a minimax bsc instance and one of its possible solutions. the instance has five sets and three questions per set. the questions must be divided into three questionnaires, which are pictured in the solution. each question is named with a lowercase letter, and its weight wrt is indicated above (e.g., question c has a weight of ). in the solution pictured, questionnaires are identified with uppercase letters, and the total weight of the questionnaires can be calculated by summing the weights of all the questions assigned to each questionnaire (e.g., questionnaire a, which contains questions a, e, h, l and o, has a total weight of ). the value that needs to be optimized is the maximum total weight for any questionnaire. in the solution pictured, the objective value is , which corresponds to questionnaire b. difficulty of the questionnaire. the objective is then to obtain questionnaires with similar difficulty. an example of both an instance and a solution for this problem is detailed in figure . one of the possible methods to achieve the previous objective and the one used in the minimax bsc is to minimize the maximum difficulty of any ques- tionnaire; this is known as a minimax objective and it is similar to the ob- jective of some bin packing related problems, like the unrelated parallel ma- chine scheduling problem with maximum completion objective (dell’ amico & martello, ). equations ( ) to ( ) are a valid integer programming (ip) formulation for the problem, which uses binary decision variables xrbt to identify the assignment of different questions to the questionnaires. the variable is equal to when item r from set t is assigned to group b, and otherwise. finally, a real-valued variable z is used to identify the maximum weight among the questionnaires. [min]z ( ) subject to: z ≥ b∑ r= t∑ t= wrt ·xrbt, ≤ b ≤ b; ( ) b∑ r= xrbt = , ≤ b ≤ b, ≤ t ≤ t ; ( ) b∑ b= xrbt = , ≤ r ≤ b, ≤ t ≤ t ; ( ) xrbt ∈{ , }, ≤ r ≤ b, ≤ b ≤ b, ≤ t ≤ t ( ) equation ( ) corresponds to the objective function (i.e., the minimization of the weight of the questionnaire with the largest sum of weights). z is calculated in the constraint set ( ). equation set ( ) verifies that each questionnaire has one question from each set, while equation set ( ) checks that each question of each set is assigned to one questionnaire. finally, constraint set ( ) defines the domain of the variables. note that constraint set ( ) corresponds to the traditional bpp constraint set, forcing each item to be assigned to one bin, while constraint set ( ) corresponds to the set of additional constraints over the traditional bpp constraint set. while the mathematical formulation can be used to obtain the optimal solu- tion to small-sized problems and to obtain lower bounds on the optimal solution of larger-sized ones, alternative methods exist in some cases: . let w denote the sum of all of the weights of the questions (w =∑ i≤b≤b, ≤t≤t wrt). then, equation ( ) provides a valid lower bound for z. this lower bound is used in the computational experiments to verify the optimality of the solutions. zlb = dw/be. ( ) . when t= , the optimal solution can be obtained as follows: first, order the questions of one set in non-decreasing order of weights and the ques- tions of the other set in non-increasing order of weights; let (a ,a , ...,ab) and (a′ ,a ′ , ...,a ′ b) be such sequences. then, build a solution so that the b-th questionnaire ( ≤ b ≤ b) is composed by questions {ab,a′b}. claim . the previous algorithm provides an optimal solution. proof the claim is verified using exchange arguments. let us consider a solution constructed with the previous method, and consider the case when the sum of weights from the first questionnaire equals the objective function. then, to reduce the objective value, one of the questions in (a , ...,ab) must be assigned together with question a′ ; however, because the weight of a is smaller than or equal to the weights of the questions in (a , ...,ab), the objective value would not improve. hence, the current solution is optimal. now suppose that the sum of weights from the second questionnaire equals the value of the objective function. then, to reduce the objective value, a can only be assigned together with a′ , effectively reducing the sum of weights of the second questionnaire. however, the best possible assignment for a′ among the remaining questions is thus a and wa′ + wa ≥ wa + wa′ . hence, the exchange would increase the objective value. the previous exchange argument can be extended to the remaining question- naires. note that we must exchange the questions between a questionnaire and an earlier questionnaire, but this would not improve the global solution unless the newly exchanged question is exchanged with the question of an earlier ques- tionnaire. eventually, the question from the first or the second questionnaire is to be considered, and the previous arguments state that the exchange is not going to improve the solution, thus proving this claim. � a computer implementation to solve the minimax bsc with t = re- quires the ordering of two sets with complexity of o(b · logb). this is a clear improvement over the b! permutations that would need to be explored if a brute force method were used. . . complexity of the problem in this section, we first reduce the partition problem to the minimax bsc problem in which b = . then, we show that the -partition problem reduces to a general minimax bsc, ascertaining the theoretical difficulty of the problem. finally, we identify the relationship between the minimax bsc problem and the subset sum knapsack problem (kellerer et al., ), which will be used in the proposed vns as one of the neighborhood evaluation methods. theorem . the minimax bsc problem with b= is np-hard. proof. we use a reduction for the partition problem that is known to be np-complete (garey & johnson, , p. ). problem partition. given a finite set a and a size s(a) ∈ z+ for each a ∈ a, is there a subset a′ ⊂ a such that ∑ a∈a′ s(a) = ∑ a∈a′−a s(a)? to ease the explanation of the reduction procedure, we assume that a is an ordered set, and thus, the t-th element of a can be referred to as at. reduction: given an instance of the problem partition, the correspond- ing minimax bsc instance with b = is constructed as follows: let t be the cardinality of a (t = |a|), and let each set ( ≤ t ≤ t) be composed by two items with weights wt = at and wt = . the optimum objective value for this instance is equal to ∑ a∈a s(a)/ if and only if the answer to the original instance of the problem partition is “yes” � theorem . the general minimax bsc problem is strongly np-hard. proof. we use a reduction for the -partition problem that is known to be strongly np-complete (garey & johnson, , p. ). problem -partition. given a set a of ·m elements, a bound u ∈ z+, and a size s(a) ∈ z+ for each a ∈ a such that u/ < s(a) < u/ and such that∑ a∈a s(a) = m · u, can a be partitioned into m disjoint sets a ,a , ...,am such that, for ≤ i ≤ m, ∑ a∈ai s(a) = u? reduction: given an instance of problem -partition, the corresponding instance of minimax bsc is constructed as follows: b = m and t = · m. then, let wt = at, wtb = for ≤ b ≤ b and ≤ t ≤ t . the optimum objective value for this instance of the minimax bsc is equal to u if and only if the answer to the original instance of problem -partition is “yes” � the reduction of the minimax bsc with b = to the partition problem allows us to solve this special case using efficient methods found in the literature for the subset sum knapsack problem. while the subset sum knapsack problem is np-hard (garey & johnson, , p. ), the problem is known to be easily solvable in many practical circumstances. the subset sum knapsack problem can be defined as follows: let there be n ( ≤ i ≤ n) items, each with weight wi, and a maximum capacity w . the objective is to find the subset of items such that the total weight is maximized without surpassing the maximum capacity w . the mathematical formulation of the problem using a set of variables xi ∈ { , }, ( ≤ i ≤ n) corresponds to maximizing ∑ ≤i≤n wi ·xi subject to ∑ ≤i≤n wi ·xi ≤ w . in this case, the subset sum knapsack problem considers the optimal as- signment of questions to one of the questionnaires (i.e., implicitly assigning the remaining questions to the other questionnaire). let wi be the additional dif- ficulty associated to selecting the most difficult question in the set i, which implies wi = |wi −wi | for each i ( ≤ i ≤ t). let xi ∈{ , }, ≤ i ≤ t be a binary variable that equals if the most difficult question from set i is assigned to the questionnaire under construction and if the easy question from set i is assigned instead. finally, let w be b ∑ i≤i≤t |wi −wi |/ c. claim . the optimal solution to the previously defined subset sum knap- sack problem provides an optimal assignment of questions to questionnaires. proof. to prove this claim, first note that each question must belong to one questionnaire; then, the unselected question of each set defines the second questionnaire. let z and z be the total difficulty of each questionnaire, which can be obtained using ( ) and ( ): z = t∑ i= xi · |wi −wi | + t∑ i= min(wi ; wi ) ( ) z = t∑ i= ( −xi) · |wi −wi | + t∑ i= min(wi ; wi ) ( ) then, the ideal weight for the non-constant part of z is b ∑ i∈b |wi − wi |/ c, which would lead to a non-constant part of z equal to d ∑ i∈b |wi − wi |/ e. hence, if z decreases, z increases. as the subset sum knapsack problem maximizes z , and the maximum possible value for z is b ∑ i∈b |wi − wi |/ c, the optimal subset sum problem minimizes the non-constant part of ( ), effectively minimizing the minimax bsc problem.� note that during the previous proof, the definition of the problem imposes w.l.o.g. that z ≤ z . . description of the algorithm the variable neighborhood search (vns) approach is a metaheuristic used as a framework to develop heuristics for specific problems. the primary con- cepts of the vns approach are based on the following observations (hansen et al., ): ( ) local optimality with respect to a neighborhood does not im- ply local optimality in other neighborhood structures; ( ) an optimum is a local optimum for any neighborhood structure; and ( ) usually a local minimum for one neighborhood is similar to those in other neighborhoods. based on these observations, vns uses multiple local searches, each with a different neighborhood structure. the algorithm starts from an initial solution and applies different local searches until a locally optimal solution to each of these neighborhoods is found. this technique is usually known as a variable neighborhood descent (vnd). while vnd improves upon the use of traditional local search methods, it may still become trapped in a local (i.e., non-global) optimum. to escape from these optima and to diversify the search, an additional mechanism must be incorporated. this mechanism is known as a shake operator and introduces a random movement from the current solution to a neighbor solution in one of the neighborhood structures. the resulting algorithm is known as a general vns. this paper proposes a general vns for the minimax bsc; the remaining subsections describe the primary components of the implementation. subsection . describes a greedy constructive procedure, which is used to obtain an initial solution and also during the “shaking” procedure. subsection . describes the neighborhood structures used in this implementation and the methods used to reach the best neighbor for each neighborhood structure. subsection . documents the ”shaking” procedure, and subsection . details the conditions used to terminate the search. . . greedy initialization heuristic while the definition of a good initial solution is not an important aspect of an effective vns heuristic, the quality of the initial solution helps reduce the time required to reach the first local optimum. therefore, we propose the use of a fast o(t ·b · logb) heuristic, with an absolute performance guarantee equal to r = max ≤t≤t rt (rt = max ≤r≤bwrt − min ≤r≤bwrt). algorithm depicts the algorithm. algorithm . greedy heuristic. for each set ≤ b ≤ b do zb= end for for each set ≤ t ≤ t do order the questions of the set into a non-decreasing order of the weights. order the questionnaires into a non-increasing order of accumulated weights. for each t, ≤ t ≤ t do assign the t-th question to the t-th questionnaire. end for end for an example depicting the process of algorithm can be found in figure . note that the method is based on considering the questionnaires of a partial solution as single questions and finding the best assignment combining these figure : illustration of the constructive heuristic. the instance is illustrated in the first row of the figure. first, set is assigned to the questionnaires at random, as shown in the second row of the figure. then, set is assigned following the heuristic logic described as follows: the question with the lowest weight is assigned to the partially filled questionnaire with the highest weight; the question with the second lowest weight is assigned to the questionnaire with the second highest weight; etc. until all of the questions are assigned, as shown in the third row of the figure. afterwards, the process is repeated for the following sets until all questions from all sets have been assigned. the last row in the figure shows the final solution. questions with an additional set, which is equivalent to finding the optimal solution of a minimax bsc problem when t = (see subsection . ). to analyze the computational complexity of the algorithm, first note that the computationally expensive part of the algorithm is ordering the questions and questionnaires. this step is executed t times, and each ordering has o(b · logb) complexity. hence, the final computational complexity of the algorithm is o(t ·b · logb). claim . algorithm provides a solution with an absolute performance guarantee of r. proof. first, we show that the maximum difference between then accumu- lated weights of any two questionnaires is equal to r. to simplify the explana- tion, let d denote the maximum difference in the accumulated weights between two questionnaires. initially, d is set to . accordingly, after the first assignment of questions to questionnaires, r ≥ d holds. during subsequent assignments, d is bounded by max{d,rt} with rt equal to the range of the assigned set (if rt ≤ d, then d−rt is bounded by d, and if rt ≥ d, then rt − d is bounded by rt). because both d ≤ r and rt ≤ r hold, the difference between any two groups will always be smaller than r. second, note that for any valid solution, the minimum total weight of any questionnaire (min ≤b≤b{zb}) is a lower bound on the optimal solution. be- cause r is a bound on the difference between any two questionnaires obtained using algorithm , the difference between the solution provided by algorithm and a lower bound for the problem is also bounded by r. this construct proves the claim that algorithm provides a solution that has an absolute performance guarantee equal to r.� the solution offered by this constructive method depends on the order in which the sets are considered. a preliminary computational experiment showed that the experimental efficiency of algorithm is improved when the sets are assigned in non-increasing values of rt order. accordingly, the initial solution of the vns uses this ordering. . . neighborhood structures neighborhood structures are the most important part of the definition of a successful vns implementation. in this work, we propose three different neighborhood structures that try to improve the incumbent solution in different ways. the search in each of the neighborhoods is conducted by solving special cases of the problem (i.e., a minimax bsc with either t = or b = ) until the solution is locally optimal for all of the explored neighborhoods. we now proceed to define each of the three neighborhoods: neighborhood . this neighborhood considers the reassignment of the questions of one set to the questionnaires. this neighborhood corresponds to the exploration of all possible solutions that can be obtained by removing all the questions of a set from the incumbent solution and their reassignment to the questionnaires. for all of the sets, the following three steps are performed: ( ) remove the questions that correspond to that set from the questionnaires; ( ) explore the reassignment of the extracted questions to the partially filled questionnaires; and ( ) assign the questions to the questionnaires according to the least expensive solution obtained in step ( ). step ( ) can be time consuming, as all possible permutations of the assign- ments of questions to questionnaires need to be considered in the worst case. note that the weight of the questionnaires in the partial solution can be con- sidered as a unified contribution to the total weight of the final questionnaires. consequently, each partially filled questionnaire can be considered as a single question from a different set, and the optimal reassignment of the extracted questions can be obtained by solving the minimax bsc with t = (see subsection . ). this neighborhood is implemented in a first descent fashion, and as such, the ordering of sets may modify the final solution. therefore, to reduce the effect of a specific ordering, whenever neighborhood is explored, a random order of sets is generated and the sets are considered in accordance with the proposed order. algorithm depicts the final algorithm. algorithm . neighborhood structure search. define a random order of the t sets repeat until no further improvement is possible for each t ( ≤ t ≤ t) do remove the questions from ordered set t from the questionnaires solve the minimax bsc problem with t = . if the new solution improves upon the incumbent solution then update the incumbent solution end if end for end repeat an example of this neighborhood is depicted in figure . note that the neighborhood considered in brusco et al. ( ) can be seen as a special case of this neighborhood. their neighborhood studies the reassign- ment of two questions from the same set to their respective questionnaires, which is a limited version of neighborhood . therefore, the limited neighborhood is not used in the final vns implementation. neighborhood . this neighborhood considers the division of the solution into two partial solutions that are then optimally combined to possibly form an improved solution. let t ′ and t ′′ be a partition of the sets in t . for any given solution s, let s′ and s′′ be the partial assignment of the questions to questionnaires defined in s for the questions of subsets t ′ and t ′′, respectively. then, consider each questionnaire in s′ and s′′ as a single question (i.e., all questions assigned to a partial questionnaire are to be assigned together when the partial questionnaires are combined). based on the previous definition, neighborhood corresponds figure : example of the application of the neighborhood . the figure depicts an initial solution in the first row (taken from the example in figure ), from which the first neigh- bourhood is created. first, all of the questions that correspond to set are removed from the solution (second row in the figure). underneath the partial solution, the total weight of each questionnaire is depicted in a box (for example, the total weight for the partial questionnaire a is ). then, the questions corresponding to set are optimally reassigned (third row) providing an improved solution. to the exploration of all possible partitions of s into s′ and s′′, as well as all of their respective assignments combining s′ and s′′ to obtain a new solution. the previous neighborhood definition has two issues that must be addressed to obtain an efficient implementation: . the number of partitions in the previous definition is not bounded by a polynomial function to the number of sets in the instance, and thus only a limited number of partitions can be considered. the implementation uses the following technique; each time neighborhood is explored, we define a random order of the sets and we explore the t− ( ≤ t < t) partitions in which t ′ is composed of the t first sets of questions according to the random ordering, and t ′′ is composed of the remaining sets of questions. . once t ′ and t ′′ have been defined, the optimal combination of the partial questionnaires in s′ and s′′ can be solved as a minimax bsc with t = , which is solvable in polynomial time (see subsection . ). algorithm shows the final algorithm. algorithm . neighborhood structure search. define a random order of the t sets repeat until no further improvement is possible for each t ( ≤ t ≤ t) do obtain the minimax bsc problem with t = with partitions t ′ and t ′′ solve the minimax bsc problem with t = if the new solution improves upon the incumbent solution then update the incumbent solution end if end for end repeat an example of an improving move using neighborhood is depicted in figure . note that neighborhood is a special case of neighborhood if multiple orderings are considered. the previous statement holds if at least one of the considered orderings contains each set as its first or last studied set. then, the exploration of neighborhood using all of these orderings subsume neighborhood as the algorithm considers the reassignment of each individual set of questions in addition to the joint reassignment of sets of questions. neighborhood . this neighborhood explores the optimal reassignment of questions to two questionnaires. while neighborhood considers changes of different sets of questions and neighborhood considers changes of different groups of sets of questions, neigh- borhood studies the effect of changing the questions assigned to two ques- tionnaires. neighborhood is thus defined as the exploration of all possible solutions in which the reassignment of questions to exactly two questionnaires may vary. figure : example of the application of the neighborhood . the initial solution (first row, from the example in figure ) is divided by selecting two groups of sets (one is composed of sets and and is highlighted by boxing their respective questions; the other group corresponds to the remaining sets). the last row depicts the solution obtained after optimally solving the related minimax bsc problem with t = . the previous definition naturally leads to multiple possible partial solutions, each requiring the exploration of a combinatorial problem (i.e., all possible as- signments of questions to two questionnaires must be investigated). conse- quently, an efficient exploration of the neighborhood must be based on some observations that may lead to an effective search for improving solutions. the following observations are used in the proposed neighborhood search method: . to improve the solution, some questions assigned to the questionnaire with the largest sum of weights must be exchanged. . the optimal assignment of two questionnaires can be found by solving a subset sum knapsack problem (see section . ). combining both observations, we can establish that one of the questionnaires considered in the subset sum knapsack problem must be one of the question- naires in which constraint ( ) in the mathematical model is fulfilled in equality (i.e., the questionnaire with the largest weight, or one of them if several exist). as in the previous neighborhoods, it is important to define the order in which the questionnaires are inspected. while the objective in the previous cases was to avoid a bias towards some sets, we opt for an ordering in this case that maximizes the probability of finding an improved solution; specifically, the questionnaires are ordered in non-decreasing order of the sum of weights. then, for each questionnaire and in the order defined above, the algorithm solves a minimax bsc problem that combines the said and the last questionnaire in the list. algorithm illustrates this procedure. algorithm . neighborhood structure search. order the questionnaires in non-decreasing order to their respective weights for each r ( ≤ r ≤ b) do solve the minimax bsc problem using the r-th and b-th questionnaires if the new solution improves upon the incumbent solution then update the incumbent solution end if end for an example of neighborhood is depicted in figure . . . shaking procedure the shaking procedure randomly chooses exchanges defined by one of the considered neighborhoods. in this case, we opt for the application of a neigh- borhood structure similar to neighborhood with a variable parameter k that controls the distance from the current partial solution. for any given k ( ≤ k < t), the proposed procedure removes the questions associated to k distinct sets from the incumbent solution. once these questions are removed, the questions of each set are introduced in the solution using the selection rule specified by neighborhood structure , which considers the partial figure : example of the application of the neighborhood . the first row depicts an initial solution (from the example in figure ). a new instance of the problem is created by consid- ering the questions assigned to questionnaires a and b; questionnaire b is selected because it is the questionnaire with the highest weight. the new instance is depicted in the second row and corresponds to an instance with sets and questions per set. the third row depicts the optimal reassignment of questions to questionnaires a and b that, with questionnaire c (which remains unchanged) provide an improved solution. solution as a single set of questions and finds the optimal assignment of a new set of questions. note that the consecutive introduction of multiple sets of questions can also be achieved using algorithm ; thus, the shaking procedure can be described as the elimination of k sets from the solution and their reintroduction into the current solution using the constructive greedy heuristic; however, a discrepancy exists because the sets are reintroduced in random order rather than in the order proposed in subsection . . the shaking procedure highlights the relationship between the constructive procedure and the two first neighborhood structures, because all three structures are based on the optimal solution to a minimax bsc instance with t = . the values of parameter k are controlled as follows: the initial value is set to k = because k = in the reintroduction scheme would return the solution to the same local optimum and thus it does not modify the solution. whenever a local optimum for each of the three neighborhoods is reached, the shaking procedure is restarted, and k is increased by one. whenever k = t − , the shaking procedure generates a random solution which can be considered as restarting the search; k is also restarted in this circumstance (i.e., k is set equal to ). . . termination condition and outline of the complete algorithm the termination condition is linked to a run time limit or the verification of an optimal solution (i.e., when the incumbent solution is equal to the lower bound defined in subsection . ). the complete algorithm is depicted in algo- rithm . algorithm . vns method. find a heuristic solution using algorithm , and initialize k. init: repeat until termination condition is met search neighborhood structure until a locally optimal solution is found. search neighborhood structure until a locally optimal solution is found. if the solution was improved during the search, goto init. search neighborhood structure until a locally optimal solution is found. if the solution was improved during the search, goto init. end repeat note that the application of the neighborhoods in the predefined order was decided based on run time considerations: the less-time-consuming algorithm is the first to be applied, while the most-time-consuming is the last to be applied. . computational experiments the previous algorithm was implemented in c and compiled using gcc v. . . and run on computer with a . ghz intel core i processor and gb of ram running the mac os x . operating system. to solve the subset sum knapsack problems, the expknap implementation (pisinger, ) available at http://www.diku.dk/-pisinger/codes.html was used. in addition to the algorithms described in this paper, we also implemented the sa algorithm proposed in brusco et al. ( ). we did not report results for the ip formulation because brusco et al. ( ) have already verified that this formulation is inefficient for large size instances. the parameters of the sa were identical to those proposed in brusco et al. ( ). furthermore, if the sa terminates before the specified running time limit, the algorithm is restarted with a different randomly generated initial solution. this vns implementation is parameter-free, and thus no parameter tuning was required. both algorithms were run for at most s, which was measured as wall time. the two algorithms were initially tested on a set of instances constructed by brusco et al. ( ). the characteristics of both sets correspond to those found in previously simulated data found in the literature (veldkamp, ; van der linder, ; van der linden & boekkooi-timminga, ). instances with varying number of questions (e.g., , , and questions) were generated. these questions were divided into different num- ber of questionnaires (e.g., b = , , , , , , , , , , and questionnaires). for each number of questions, the abovementioned num- ber of questionnaires is only used if the number of questions is at least five times bigger than the number of questionnaires (i.e., each questionnaire will be formed by five or more questions). finally, the weights of the questions in each instance were generated using the following method, which was proposed in br- usco et al. ( ), and based on van der linden & boekkooi-timminga ( ). the method uses formula ( ) to compute the weight of each question based on two parameters πrt and ρrt, which represent the difficulty (i.e., percentage of examinees who obtained the correct answer) and the discrimination effect of the question, respectively. wrt = . ·πrt + . ·πrt · ( −πrt) ·ρrt ( ) to generate sets of instances with similar weights, the method initially ob- tains the difficulty π t and discrimination effect ρ t from the first question of each set t. these values are drawn from a uniform distribution with interval [. , . ] for the difficulty effect (π) and from the interval [. , . ] for the discrimination effect (ρ). for the remaining questions πrt and ρrt ( ≤ r ≤ b), the values are randomly drawn from uniform distributions with interval [π t − . ,π t + . ] and [ρ t − . ,ρ t + . ]. these weights are then used in combination with ( ) to generate the final weights. ten instances per combination of number of questions, questionnaires and distribution of weights were constructed to create a total of instances. note that while the weights were obtained from uniform distributions and thus are real values, the method works with integer-valued weights to avoid rounding issues. to achieve this objective, we decided to multiply the obtained values by a constant ( ) and round to the nearest integer. consequently, this algorithm is working using real values with six digits of precision. table provides a summary of the experimental results. for each number of questions and questionnaires, the table provides a comparison between the heuristic method depicted in subsection . , the vns method proposed in this paper and the sa algorithm proposed by brusco et al. ( ). for each al- gorithm, we provide the average absolute gap between the lower bound (lb), calculated using equation ( ), and the upper bound (ub) provided by the al- gorithm (that is, ub −lb). for the metaheuristic approaches, the number of optimal solutions (out of instances) and the running time required to reach the best solution (measured in wall time) are also provided. an analysis of the table shows that the gap between the solutions provided by both metaheuristics and the lower bound are small in all cases. furthermore, both methods are able to dramatically improve the initial solutions provided by the heuristic with a performance guarantee. furthermore, the vns was able to obtain the optimal solution for most of the instances, including all of the instances in which the ratio between the questions and questionnaires was higher than within negligible running times (i.e., within one second). hence, we can conclude that those instances in which the ratio of questions per questionnaire is larger than or equal to are to be considered easy for the proposed method. the running time increased for the instances in which the algorithm was not able to verify optimality, but were still reduced. note that table reports the running time to reach the best solution for all methods rather than the total computing time required. according to the termination condition, if the optimal solution is not verified, both algorithms continue until the time limit is reached. based on the results provided above, we test the algorithm with larger in- stances (e.g., with t= . and . , and different values of b ranging from to . ). similar results to those reported in table were obtained, and the vns method was able to optimally solve all instances with a ratio of questions to questionnaires larger than within a maximum of seconds of computing time. therefore, we limit this report to the results for those groups of instances in which the vns fails to verify the optimal solution for every instance with the given ratio of questions per questionnaire (“challenging” instances). table reports these results. the results from table show that the algorithm is still effective for larger- sized instances. furthermore, the number of verified optimal solutions increases due to the existence of more alternatives to assign the questions. consequently, we conclude that for the vns method, increasing the number of questions does not affect the quality of the solutions, while the running times increase in ac- cordance with the requirement of solving larger problems, specifically larger knapsack instances. the results provided in tables and indicate that the proposed vns offers high quality results, but do not identify which components of the algorithm are responsible for the good performance. therefore, an additional experiment table : results of the proposed algorithms. for each combination of characteristics, we provide the number of optimal solutions verified by each of the metaheuristic algorithms, the average absolute gap (i.e., the average difference between the solution and the lower bound, provided by equation ) for each algorithm and the average time required to reach the best solution by the metaheuristic methods; for the sa algorithm, we report the time required for replication that provided the solution under consideration. questions b heuristic vns vns vns sa sa sa gap opt gap t opt gap t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . table : results for larger instances. only those combinations of the number of questions and questionnaires that are deemed to be ”challenging” are reported. for each combination of characteristics, we provide the same results that in table . questions b heuristic vns vns vns sa sa sa gap opt gap t opt gap t . . . . . . . . . . . . . . . . geared towards evaluating the contribution of each of the specific components was conducted. the experiment considers different configurations of the vns algorithm in which some of the neighborhoods or the shaking procedure are not used. while specific neighborhood can be removed from the algorithm without modifying its global structure, it is necessary to include some kind of restarting procedure to escape from local optimality, continue the search and effectively use the allotted computing time. consequently, when the shaking procedure is not used, it is substituted by the greedy initialization procedure proposed in subsection . in which the order of the sets is randomly determined before executing algorithm . a total of different configurations of the vns algorithm were tested (all possible combinations of use/not use for each of the three neighborhoods and the shaking procedure, with the exception of the two combinations that do not use any of the three neighborhoods) on the large and “challenging instances (the instances reported in table ). each configuration was run for at most seconds of wall time, and three values were recorded: ( ) the running time required to reach the best solution; ( ) the absolute gap between the solution found by the configuration and the best known solution for the instance; and ( ) the absolute gap between the solution found by the configuration and the theoretical bin packing lower bound. furthermore, and in order to verify if the restarting procedure was efficient, we also run the procedure as a simple descent method with all of the neighborhood structures and recorded the same information. table summarizes the results obtained for the experiment. the examina- tion of table highlights the effectiveness of neighborhood (gaps are much smaller if neighborhood is used). note that the gap reduction is accompanied by larger computational times. the shaking procedure, as well as neighborhood and neighborhood do not affect the solution quality but they do seem to reduce the running time required to reach the best solution. also note that the gap difference among the vns configurations using neighborhood is minimal. if we compare the single descent method with the vns algorithm, the re- sults from table show that the vns is capable of providing improvements on the solution provided by the simple descent method, at the cost of larger computational times. please note that the gap reduction is small as the solu- table : results for different configurations of the proposed algorithms. for each combination of use/not use of each of three neighborhoods (columns n , n and n ) and shaking procedure (column “shake”), the average absolute gap between the solution found by the configuration and the theoretical bin packing lower bound (column gapbound); the average absolute gap between the solution found by the configuration and the best known solution for the instance (column gapbest); and the average running time required to reach the best solution (column t) are reported. additionally, the results of the simple descent algorithm (no restart used) are also reported (row with “shake” entry equal to “descent”). n n n shake gapbound gapbest t no no yes no . . . no no yes yes . . . no yes no no . . . no yes no yes . . . no yes yes no . . . no yes yes yes . . . yes no no no . . . yes no no yes . . . yes no yes no . . . yes no yes yes . . . yes yes no no . . . yes yes no yes . . . yes yes yes no . . . yes yes yes yes . . . yes yes yes decent . . . tions provided by the descent heuristic are already very good and show close to nonexistent optimality gaps. in addition to the previous qualitative analysis, we conducted an anova test on two response variables (gapbest and running time) and a manova test using both response variables (see christensen, , for a general reference on anova and manova). the tests reported differences among vns configura- tions, but an analysis of the residuals showed severe violations of the normality assumptions and heteroscedasticity issues. as a consequence, we chose to reex- amine the results using a non-parametric manova alternative. the permutational manova using distance matrices proposed in anderson ( ) was selected. the method does not require traditional assumptions; it is based on a geometric interpretation of the analysis of variance; and it uses permutations in order to obtain p-values. furthermore, the code is readily available on the r statistical package (oksanen et al., ). the test was run using four factors (the three neighborhood searches, and the shaking procedure) and a blocking factor (the instances) that tries to block the variability caused by specific characteristics of each instance. the results show that neighborhood structures and are significant with a p-value < . , and the shaking procedure is significant with a p-value < . . neighborhood structure is not significant, leading us to conclude that it could be removed from the algorithms without modifying its results (both in terms of running time and solution quality). if we combine the p-values with the analysis of table , we conclude that the different components of the proposed vns method are able to improve the solution obtained by the final algorithm (neighborhood structure ); to reduce the required running time to reach the aforementioned solution(neighborhood structure and shaking procedure); or do not significatively modify running time or solution quality. a final experiment was performed to evaluate the behavior of the proposed method if the sets were based on some other constraints (e.g., length, time re- quired to answer the question) and not on similar objective function weights. a new instance set in which the weights of the questions are uniformly distributed is considered, see for example nguyen & fong ( ). for these instances, each weight is randomly selected form a uniform distribution on the interval [. , . ] maintaining the remaining characteristics from the instances. the results, see table , show that the performance of the vns method is significantly better than its sa counterpart. hence the vns depends less on a grouping of questions by similar weights. the relation between the number of questions and questionnaires to ascertain the difficulty of the instance still holds. however, the instances are harder to solve for both metaheuristics, and larger absolute gaps appear for those instances in which optimality is not verified (even if they are still small). note that the heuristic constructive method is clearly affected by changes in structure and generates significantly worse initial solutions. as similar results in terms of optimally solved instances are obtained, we conclude that the distribution of weights among the questions in each set does not greatly affect the performance of the vns method. table : results for the instances generated according to a uniform distribution as in nguyen & fong ( ). for each combination of characteristics, we provide the number of optimal solutions verified by each algorithm; the average absolute gap, which is measured as the average difference between the solution and the lower bound using equation ; and the average time required to reach the best solution; for the sa algorithm, we report the time required for replication that provided the solution under consideration. questions b heuristic vns vns vns sa sa sa gap opt gap t opt gap t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . conclusions in this paper, we have investigated a bin packing-based formulation of the test assembly design problem known as the one-dimensional minimax bin- packing problem with bin size constraints (minmax bsc). we first show that the problem in np-hard when the number of questionnaires is and strongly np-hard if more questionnaires are considered. second, we develop a con- structive heuristic with an absolute performance guarantee, and we derive three neighborhood structures for the problem that we can explore by optimally solv- ing special cases of the problem. third, we combine these procedures into a variable neighborhood search in order to efficiently search the solution space. finally, the quality of the proposed procedures is tested in computational exper- iments on instances with up to . item pools, up to . questionnaires, and different discrimination effect structures (see brusco et al., ; nguyen & fong, ). the proposed method shows significant improvements over previous algorithms proposed in the literature (brusco et al., ), and the optimality of the solutions provided for instances with a . items pool and up to . tests has been verified. our complexity results follow the line of previous works (like belov & arm- strong, ; nguyen & fong, ; ishii et al., ) in which other formula- tions of the test assembly design problem were identified as classical strongly np-hard combinatorial optimization problems. our approach differs when we consider the resolution considerations, as belov & armstrong ( ); nguyen & fong ( ); ishii et al. ( ) apply classical algorithms of the identified optimization problem to the resolution of the test design problem, while our bin packing-based formulation significantly differs from the classical problem, thus forcing us to develop specially tailored procedures for its resolution. the proposed algorithm corresponds to a neighborhood exploration heuris- tic. if compared to previous neighborhood-based approaches for test assembly design (hwang et al., ; brusco et al., ), which are based on traditional swap or exchange neighborhoods, our neighborhoods are based on the optimal exploration of special structures. these neighborhoods are derived from theo- retical results for the minmax bsc, and successfully use the structure of the problem to reach high quality results within very reduced running times. the quality of the solutions provided by the method is demonstrated by our experimental results with instances larger than those previously considered in the literature (see hwang et al., ; nguyen & fong, ). please note that the results are not directly comparable as there are strong differences among different formulations for the test assembly design problem. nevertheless, we can make comparisons with a bin packing lower bound, which shows the high quality of the solutions found. the proposed method routinely finds the optimal solution when the ratio between items per questionnaire is above , and obtains very low gaps if the aforementioned ratio is or below. these results apply under varying conditions, such as when the items are grouped according to similar weights, as in brusco et al. ( ), or when the items are randomly grouped. results for randomly grouped items consider cases in which additional constraints are to be enforced (he et al., ; ishii et al., ; nguyen & fong, ), or cases in which a cell approach (chen et al., ; chen, ) is used to maintain similar characteristics not only among questionnaires, but also among the item composition of each questionnaire. in addition to the quality of the results, we would like to highlight the reduced running times required to solve the instances, which are below the seconds mark even for the largest item pools. these results lead us to conjecture that the proposed procedure could also be used for on-line test construction methods as in van der linden & veldkamp ( ) or he et al. ( ). the possibility of efficiently solving large-sized instances leads us to con- clude that any further research should address the inclusion of additional char- acteristics into the formulation. among these characteristics, we would like to highlight the joint evaluation of multiple ability levels of the recipients on a single unified test (van der linder, ; chang & shiu, ; nguyen & fong, ) or the inclusion of additional constraints, such as incompatibilities among questions of different sets or limiting the number of questions with specific at- tributes (hwang et al., ; yao, ; yao et al., ). these additional characteristics are usually present in many real situations and would impose challenging changes to both the bin-packing-based formulation as well as to the neighborhood structures proposed in this work. references anderson, m. j. ( ). a new method for non-parametric multivariate analysis of variance. austral ecology, , – . belov, d. i. ( ). uniform test assembly. psychometrika, , – . doi: . /s - - - . belov, d. i., & armstrong, r. d. ( ). a constraint programming ap- proach to extract the maximum number of non-overlapping test forms. computational optimization and applications, , – . doi: . / s - - -z. birnbaum, a. ( ). some latent trait models and their use in inferring an examinees ability. in f. lord, & m. novick (eds.), statistical theories of mental test scores chapter some laten. (pp. – ). reading, ma: addison- wesley. brandão, f., & pedroso, j. a. p. ( ). fast pattern-based algorithms for cut- ting stock. computers and operations research, , – . url: http://dx. doi.org/ . /j.cor. . . . doi: . /j.cor. . . . brusco, m. j., köhn, h. f., & steinley, d. ( ). exact and approximate meth- ods for a one-dimensional minimax bin-packing problem. annals of opera- tions research, , – . url: http://link.springer.com/ . / s - - - . doi: . /s - - - . http://dx.doi.org/ . /s - - - http://dx.doi.org/ . /s - - - http://dx.doi.org/ . /s - - - http://dx.doi.org/ . /s - - -z http://dx.doi.org/ . /s - - -z http://dx.doi.org/ . /s - - -z http://dx.doi.org/ . /j.cor. . . http://dx.doi.org/ . /j.cor. . . http://dx.doi.org/ . /j.cor. . . http://dx.doi.org/ . /j.cor. . . http://link.springer.com/ . /s - - - http://link.springer.com/ . /s - - - http://link.springer.com/ . /s - - - http://dx.doi.org/ . /s - - - chang, t. y., & ke, y. r. ( ). a personalized e-course composition based on a genetic algorithm with forcing legality in an adaptive learn- ing system. journal of network and computer applications, , – . url: http://dx.doi.org/ . /j.jnca. . . . doi: . /j. jnca. . . . chang, t.-y., & shiu, y.-f. ( ). simultaneously construct irt-based paral- lel tests based on an adapted clonalg algorithm. applied intelligence, , – . url: http://link.springer.com/ . /s - - -x. doi: . /s - - -x. chen, p.-h. ( ). a sampling and classification item selection ap- proach with content balancing. behavior research methods, . doi: . / s - - - . chen, p.-h., chang, h.-h., & wu, h. ( ). item selection for the devel- opment of parallel forms from an irt-based seed test using a sampling and classification approach. educational and psychological measurement, , – . doi: . / . christensen, r. ( ). plane asnwers to complex questions. the theory of linear models. ( th ed.). springer. dell’ amico, m., & martello, s. ( ). optimal scheduling of tasks on identical parallel processors. orsa journal of computing, , – . doi: . / ijoc. . . . dyckhoff, h. ( ). a typology of cutting and packing problems. eu- ropean journal of operational research, , – . url: http: //www.sciencedirect.com/science/article/pii/ k. doi: . / - ( ) -k. edmonds, j., & armstrong, r. ( ). a mixed integer programming model for multiple stage adaptive testing. european journal of operational re- search, , – . url: http://linkinghub.elsevier.com/retrieve/ pii/s . doi: . /j.ejor. . . . falkenauer, e. ( ). a hybrid grouping genetic algorithm for bin packing. journal of heuristics, , – . url: http://link.springer.com/ . / bf . doi: . /bf . fleszar, k., & hindi, k. s. ( ). new heuristics for one-dimensional bin- packing. computers & operation research, , – . doi: . / s - ( ) - . garey, m. r., & johnson, d. s. ( ). computers and intractability. a guide to the theory of np-completeness. new york: freeman. http://dx.doi.org/ . /j.jnca. . . http://dx.doi.org/ . /j.jnca. . . http://dx.doi.org/ . /j.jnca. . . http://dx.doi.org/ . /j.jnca. . . http://link.springer.com/ . /s - - -x http://dx.doi.org/ . /s - - -x http://dx.doi.org/ . /s - - - http://dx.doi.org/ . /s - - - http://dx.doi.org/ . /s - - - http://dx.doi.org/ . / http://dx.doi.org/ . /ijoc. . . http://dx.doi.org/ . /ijoc. . . http://dx.doi.org/ . /ijoc. . . http://www.sciencedirect.com/science/article/pii/ k http://www.sciencedirect.com/science/article/pii/ k http://www.sciencedirect.com/science/article/pii/ k http://dx.doi.org/ . / - ( ) -k http://linkinghub.elsevier.com/retrieve/pii/s http://linkinghub.elsevier.com/retrieve/pii/s http://linkinghub.elsevier.com/retrieve/pii/s http://dx.doi.org/ . /j.ejor. . . http://link.springer.com/ . /bf http://link.springer.com/ . /bf http://link.springer.com/ . /bf http://dx.doi.org/ . /bf http://dx.doi.org/ . /s - ( ) - http://dx.doi.org/ . /s - ( ) - http://dx.doi.org/ . /s - ( ) - hansen, p., mladenović, n., & moreno pérez, j. a. ( ). variable neighbour- hood search: methods and applications. annals of operations research, , – . url: http://link.springer.com/ . /s - - - . doi: . /s - - - . he, w., diao, q., & hauser, c. ( ). a comparison of four item-selection methods for severely constrained cats. educational and psychological mea- surement, , – . url: http://epm.sagepub.com/cgi/doi/ . / . doi: . / . hwang, g. j., chu, h. c., yin, p. y., & lin, j. y. ( ). an innovative parallel test sheet composition approach to meet multiple assessment criteria for national tests. computers and education, , – . doi: . / j.compedu. . . . ishii, t., songmuang, p., & ueno, m. ( ). maximum clique algorithm and its approximation for uniform test form assembly. ieee transactions on learning technologies, , – . url: http://ieeexplore.ieee.org/ articledetails.jsp?arnumber= . doi: . /tlt. . . kallrath, j., rebennack, s., kallrath, j., & kusche, r. ( ). solving real-world cutting stock-problems in the paper industry: mathematical ap- proaches, experience and challenges. european journal of operational re- search, , – . url: http://dx.doi.org/ . /j.ejor. . . . doi: . /j.ejor. . . . kellerer, h., pferschy, u., & pisinger, d. ( ). knapsack problems. springer. van der linden, w. j. ( ). optimal assembly of psychological and educational tests. applied psychological measurement, , – . url: http://apm.sagepub.com/content/ / / .abstract. doi: . / . van der linden, w. j., & boekkooi-timminga, e. ( ). a zero-one program- ming approach to guiliksen’s matched random subtests method. applied psychological measurement, , – . url: http://apm.sagepub.com/ content/ / / .abstract. doi: . / . van der linden, w. j., & veldkamp, b. p. ( ). constraining item exposure in computerized adaptive testing with shadow tests. journal of educational and behavioral statistics, , – . doi: . / . van der linder, w. j. ( ). linear models for optimal test design. springer. lópez-camacho, e., terashima-marin, h., ross, p., & ochoa, g. ( ). a unified hyper-heuristic framework for solving bin packing prob- lems. expert systems with applications, , – . url: http: //linkinghub.elsevier.com/retrieve/pii/s . doi: . /j.eswa. . . . http://link.springer.com/ . /s - - - http://dx.doi.org/ . /s - - - http://epm.sagepub.com/cgi/doi/ . / http://epm.sagepub.com/cgi/doi/ . / http://epm.sagepub.com/cgi/doi/ . / http://dx.doi.org/ . / http://dx.doi.org/ . /j.compedu. . . http://dx.doi.org/ . /j.compedu. . . http://dx.doi.org/ . /j.compedu. . . http://ieeexplore.ieee.org/articledetails.jsp?arnumber= http://ieeexplore.ieee.org/articledetails.jsp?arnumber= http://ieeexplore.ieee.org/articledetails.jsp?arnumber= http://dx.doi.org/ . /tlt. . http://dx.doi.org/ . /j.ejor. . . http://dx.doi.org/ . /j.ejor. . . http://dx.doi.org/ . /j.ejor. . . http://dx.doi.org/ . /j.ejor. . . http://apm.sagepub.com/content/ / / .abstract http://dx.doi.org/ . / http://dx.doi.org/ . / http://dx.doi.org/ . / http://apm.sagepub.com/content/ / / .abstract http://apm.sagepub.com/content/ / / .abstract http://apm.sagepub.com/content/ / / .abstract http://dx.doi.org/ . / http://dx.doi.org/ . / http://linkinghub.elsevier.com/retrieve/pii/s http://linkinghub.elsevier.com/retrieve/pii/s http://linkinghub.elsevier.com/retrieve/pii/s http://dx.doi.org/ . /j.eswa. . . http://dx.doi.org/ . /j.eswa. . . http://dx.doi.org/ . /j.eswa. . . lu, h.-y. ( ). application of optimal designs to item calibration. plos one, , – . doi: . /journal.pone. . m’hallah, r., alkandari, a., & mladenovic, n. ( ). packing unit spheres into the smallest sphere using vns and nlp. computers and operations re- search, , – . url: http://dx.doi.org/ . /j.cor. . . . doi: . /j.cor. . . . mladenović, n., & hansen, p. ( ). variable neighborhood search. computers & operations research, , – . url: http: //www.sciencedirect.com/science/article/pii/s . doi: . /s - ( ) - . nguyen, m. l., & fong, a. c. m. ( ). large-scale multiobjective static test generation for web-based testing with integer programming. ieee transac- tions on learning technologies, , – . url: http://www.computer.org/ csdl/trans/lt/ / /tlt .html. doi: . /tlt. . . oksanen, j., blanchet, f. g., kindt, r., legendre, p., minchin, p. r., o’hara, r. b., simpson, g. l., solymos, p., m., h., h., s., & wagner, h. ( ). vegan: community ecology package. url: http://cran.r-project.org/ package=vegan. pisinger, d. ( ). an expanding-core algorithm for the exact - knapsack problem. european journal of operational research, , – . url: http://www.sciencedirect.com/science/article/pii/ . doi: . / - ( ) - . porter, a., polikoff, m. s., barghaus, k. m., & yang, r. ( ). construct- ing aligned assessments using automated test construction. educational researcher, , – . url: http://edr.sagepub.com/content/ / / .abstract. doi: . / x . quiroz-castellanos, m., cruz-reyes, l., torres-jimenez, j., s, c. g., fraire, h. j., & alvim, a. c. f. ( ). a grouping genetic algorithm with controlled gene transmission for the bin packing problem. computers & operation re- search, , – . url: http://dx.doi.org/ . /j.cor. . . . doi: . /j.cor. . . . smits, n., & finkelman, m. d. ( ). variable length testing using the ordi- nal regression model. statistics in medicine, , – . doi: . /sim. . songmuang, p., & ueno, m. ( ). bees algorithm for construction of multiple test forms in e-testing. ieee transactions on learning tech- nologies, , – . url: http://ieeexplore.ieee.org/lpdocs/epic / wrapper.htm?arnumber= . doi: . /tlt. . . http://dx.doi.org/ . /journal.pone. http://dx.doi.org/ . /j.cor. . . http://dx.doi.org/ . /j.cor. . . http://dx.doi.org/ . /j.cor. . . http://dx.doi.org/ . /j.cor. . . http://www.sciencedirect.com/science/article/pii/s http://www.sciencedirect.com/science/article/pii/s http://www.sciencedirect.com/science/article/pii/s http://dx.doi.org/ . /s - ( ) - http://www.computer.org/csdl/trans/lt/ / /tlt .html http://www.computer.org/csdl/trans/lt/ / /tlt .html http://www.computer.org/csdl/trans/lt/ / /tlt .html http://dx.doi.org/ . /tlt. . http://cran.r-project.org/package=vegan http://cran.r-project.org/package=vegan http://cran.r-project.org/package=vegan http://www.sciencedirect.com/science/article/pii/ http://www.sciencedirect.com/science/article/pii/ http://www.sciencedirect.com/science/article/pii/ http://dx.doi.org/ . / - ( ) - http://edr.sagepub.com/content/ / / .abstract http://edr.sagepub.com/content/ / / .abstract http://edr.sagepub.com/content/ / / .abstract http://dx.doi.org/ . / x http://dx.doi.org/ . /j.cor. . . http://dx.doi.org/ . /j.cor. . . http://dx.doi.org/ . /sim. http://dx.doi.org/ . /sim. http://dx.doi.org/ . /sim. http://ieeexplore.ieee.org/lpdocs/epic /wrapper.htm?arnumber= http://ieeexplore.ieee.org/lpdocs/epic /wrapper.htm?arnumber= http://ieeexplore.ieee.org/lpdocs/epic /wrapper.htm?arnumber= http://dx.doi.org/ . /tlt. . sun, k.-t., chen, y.-j., tsai, s.-y., & cheng, c.-f. ( ). creating irt-based parallel test forms using the genetic algorithm method. applied measure- ment in education, , – . url: http://www.tandfonline.com/doi/ abs/ . / . doi: . / . vanderbeck, f. ( ). computational study of a column generation algorithm for bin packing and cutting stock problems. mathematical programming, , – . doi: . /s . veldkamp, b. p. ( ). encyclopedia of social measurement. el- sevier. url: http://www.sciencedirect.com/science/article/pii/ b . doi: . /b - - - / - . veldkamp, b. p. ( ). application of robust optimization to auto- mated test assembly. annals of operations research, , – . url: http://link.springer.com/ . /s - - -y. doi: . /s - - -y. wäscher, g., hauß ner, h., & schumann, h. ( ). an improved typology of cutting and packing problems. european journal of operational research, , – . url: http://www.sciencedirect.com/science/article/ pii/s x. doi: . /j.ejor. . . . yao, l. ( ). multidimensional cat item selection methods for domain scores and composite scores with item exposure control and content con- straints. journal of educational measurement, , – . doi: . /jedm. . yao, l., pommerich, m., & segall, d. o. ( ). using multidimensional cat to administer a short , yet precise , screening test. applied psychological measurement, , – . doi: . / . http://www.tandfonline.com/doi/abs/ . / http://www.tandfonline.com/doi/abs/ . / http://www.tandfonline.com/doi/abs/ . / http://dx.doi.org/ . / http://dx.doi.org/ . /s http://www.sciencedirect.com/science/article/pii/b http://www.sciencedirect.com/science/article/pii/b http://www.sciencedirect.com/science/article/pii/b http://dx.doi.org/ . /b - - - / - http://link.springer.com/ . /s - - -y http://dx.doi.org/ . /s - - -y http://dx.doi.org/ . /s - - -y http://dx.doi.org/ . /s - - -y http://www.sciencedirect.com/science/article/pii/s x http://www.sciencedirect.com/science/article/pii/s x http://www.sciencedirect.com/science/article/pii/s x http://dx.doi.org/ . /j.ejor. . . http://dx.doi.org/ . /jedm. http://dx.doi.org/ . /jedm. http://dx.doi.org/ . /jedm. http://dx.doi.org/ . / introduction contributions of this work paper outline problem description literature review mathematical formulation, lower bounds and special cases complexity of the problem description of the algorithm greedy initialization heuristic neighborhood structures shaking procedure termination condition and outline of the complete algorithm computational experiments conclusions a survey on sentiment detection of reviews expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a a survey on sentiment detection of reviews huifeng tang, songbo tan *, xueqi cheng information security center, institute of computing technology, chinese academy of sciences, beijing , pr china a r t i c l e i n f o keywords: sentiment detection opinion extraction sentiment classification - /$ - see front matter � published by doi: . /j.eswa. . . * corresponding author. e-mail addresses: tanghuifeng@software.ict.ac.cn (h e.ict.ac.cn (s. tan), cxq@ict.ac.cn (x. cheng). a b s t r a c t the sentiment detection of texts has been witnessed a booming interest in recent years, due to the increased availability of online reviews in digital form and the ensuing need to organize them. till to now, there are mainly four different problems predominating in this research community, namely, sub- jectivity classification, word sentiment classification, document sentiment classification and opinion extraction. in fact, there are inherent relations between them. subjectivity classification can prevent the sentiment classifier from considering irrelevant or even potentially misleading text. document sen- timent classification and opinion extraction have often involved word sentiment classification tech- niques. this survey discusses related issues and main approaches to these problems. � published by elsevier ltd. . introduction today, very large amount of reviews are available on the web, as well as the weblogs are fast-growing in blogsphere. product re- views exist in a variety of forms on the web: sites dedicated to a specific type of product (such as digital camera), sites for newspa- pers and magazines that may feature reviews (like rolling stone or consumer reports), sites that couple reviews with commerce (like amazon), and sites that specialize in collecting professional or user reviews in a variety of areas (like rottentomates.com). less formal reviews are available on discussion boards and mailing list archives, as well as in usenet via google groups. users also com- ment on products in their personal web sites and blogs, which are then aggregated by sites such as blogstreet.com, allconsum- ing.net, and onfocus.com. the information mentioned above is a rich and useful source for marketing intelligence, social psychologists, and others interested in extracting and mining opinions, views, moods, and attitudes. for example, whether a product review is positive or negative; what are the moods among bloggers at that time; how the public reflect towards this political affair, etc. to achieve this goal, a core and essential job is to detect subjec- tive information contained in texts, include viewpoint, fancy, atti- tude, sensibility etc. this is so-called sentiment detection. a challenging aspect of this task seems to distinguish it from traditional topic-based detection (classification) is that while top- ics are often identifiable by keywords alone, sentiment can be ex- pressed in a much subtle manner. for example, the sentence ‘‘what a bad picture quality that digital camera has! . . . oh, this elsevier ltd. . tang), tansongbo@softwar- new type camera has a good picture, long battery life and beautiful appearance!” compares a negative experience of one product with a positive experience of another product. it is difficult to separate out the core assessment that should actually be correlated with the document. thus, sentiment seems to require more understand- ing than the usual topic-based classification. sentiment detection dates back to the late s (argamon, koppel, & avneri, ; kessler, nunberg, & schÄutze, ; sper- tus, ), but only in the early s did it become a major sub- field of the information management discipline (chaovalit & zhou, ; dimitrova, finn, kushmerick, & smyth, ; durbin, neal richter, & warner, ; efron, ; gamon, ; glance, hurst, & tomokiyo, ; grefenstette, qu, shanahan, & evans, ; hil- lard, ostendorf, & shriberg, ; inkpen, feiguina, & hirst, ; kobayashi, inui, & inui, ; liu, lieberman, & selker, ; rau- bern & muller-kogler, ; riloff and wiebe, ; subasic & huettner, ; tong, ; vegnaduzzo, ; wiebe & riloff, ; wilson, wiebe, & hoffmann, ). until the early s, the two main popular approaches to sentiment detection, espe- cially in the real-world applications, were based on machine learn- ing techniques and based on semantic analysis techniques. after that, the shallow nature language processing techniques were widely used in this area, especially in the document sentiment detection. current-day sentiment detection is thus a discipline at the crossroads of nlp and ir, and as such it shares a number of characteristics with other tasks such as information extraction and text-mining. although several international conferences have devoted spe- cial issues to this topic, such as acl, aaai, www, emnlp, cikm etc., there are no systematic treatments of the subject: there are neither textbooks nor journals entirely devoted to sentiment detection yet. mailto:tanghuifeng@software.ict.ac.cn mailto:tansongbo@software.ict.ac.cn mailto:tansongbo@software.ict.ac.cn mailto:cxq@ict.ac.cn http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa h. tang et al. / expert systems with applications ( ) – this paper first introduces the definitions of several problems that pertain to sentiment detection. then we present some appli- cations of sentiment detection. section discusses the subjectivity classification problem. section introduces semantic orientation method. the sixth section examines the effectiveness of applying machine learning techniques to document sentiment classification. the seventh section discusses opinion extraction problem. the eighth part talks about evaluation of sentiment detection. last sec- tion concludes with challenges and discussion of future work. . sentiment detection . . subjectivity classification subjectivity in natural language refers to aspects of language used to express opinions and evaluations (wiebe, ). subjectiv- ity classification is stated as follows: let s = {s , . . . , sn} be a set of sentences in document d. the problem of subjectivity classification is to distinguish sentences used to present opinions and other forms of subjectivity (subjective sentences set ss) from sentences used to objectively present factual information (objective sen- tences set so), where ss [ so = s. this task is especially relevant for news reporting and internet forums, in which opinions of various agents are expressed. . . sentiment classification sentiment classification includes two kinds of classification forms, i.e., binary sentiment classification and multi-class senti- ment classification. given a document set d = {d , . . . , dn}, and a pre-defined categories set c = {positive, negative}, binary senti- ment classification is to classify each di in d, with a label expressed in c. if we set c* = {strong positive, positive, neutral, negative, strong negative} and classify each di in d with a label in c*, the problem changes to multi-class sentiment classification. most prior work on learning to identify sentiment has focused on the binary distinction of positive vs. negative. but it is often helpful to have more information than this binary distinction pro- vides, especially if one is ranking items by recommendation or comparing several reviewers’ opinions. koppel and schler ( a, b) show that it is crucial to use neutral examples in learning polarity for a variety of reasons. learning from negative and posi- tive examples alone will not permit accurate classification of neu- tral examples. moreover, the use of neutral training examples in learning facilitates better distinction between positive and nega- tive examples. . applications of sentiment detection in this section, we will expound some rising applications of sen- timent detection. . . products comparison it is a common practice for online merchants to ask their cus- tomers to review the products that they have purchased. with more and more people using the web to express opinions, the number of reviews that a product receives grows rapidly. most of the researches about these reviews were focused on automatically classifying the products into ‘‘recommended” or ‘‘not recom- mended” (pang, lee, & vaithyanathan, ; ranjan das & chen, ; terveen, hill, amento, mcdonald, & creter, ). but every product has several features, in which maybe only part of them people are interested. moreover, a product has shortcomings in one aspect, probably has merits in another place (morinaga, yamanishi, tateishi, & fukushima, ; taboada, gillies, & mcfe- tridge, ). to analysis the online reviews and bring forward a visual man- ner to compare consumers’ opinions of different products, i.e., merely with a single glance the user can clearly see the advantages and weaknesses of each product in the minds of consumers. for a potential customer, he/she can see a visual side-by-side and fea- ture-by-feature comparison of consumer opinions on these prod- ucts, which helps him/her to decide which product to buy. for a product manufacturer, the comparison enables it to easily gather marketing intelligence and product benchmarking information. liu, hu, and cheng ( ) proposed a novel framework for ana- lyzing and comparing consumer opinions of competing products. a prototype system called opinion observer is implemented. to en- able the visualization, two tasks were performed: ( ) identifying product features that customers have expressed their opinions on, based on language pattern mining techniques. such features form the basis for the comparison. ( ) for each feature, identifying whether the opinion from each reviewer is positive or negative, if any. different users can visualize and compare opinions of different products using a user interface. the user simply chooses the prod- ucts that he/she wishes to compare and the system then retrieves the analyzed results of these products and displays them in the interface. . . opinion summarization the number of online reviews that a product receives grows rapidly, especially for some popular products. furthermore, many reviews are long and have only a few sentences containing opin- ions on the product. this makes it hard for a potential customer to read them to make an informed decision on whether to purchase the product. the large number of reviews also makes it hard for product manufacturers to keep track of customer opinions of their products because many merchant sites may sell their products, and the manufacturer may produce many kinds of products. opinion summarization (ku, lee, wu, & chen, ; philip et al., ) summarizes opinions of articles by telling sentiment polari- ties, degree and the correlated events. with opinion summariza- tion, a customer can easily see how the existing customers feel about a product, and the product manufacturer can get the reason why different stands people like it or what they complain about. hu and liu ( a, b) conduct a work like that: given a set of customer reviews of a particular product, the task involves three subtasks: ( ) identifying features of the product that customers have expressed their opinions on (called product features); ( ) for each feature, identifying review sentences that give positive or negative opinions; and ( ) producing a summary using the dis- covered information. ku, liang, and chen ( ) investigated both news and web blog articles. in their research, trec, ntcir and articles collected from web blogs serve as the information sources for opinion extraction. documents related to the issue of animal cloning are selected as the experimental materials. algorithms for opinion extraction at word, sentence and document level are proposed. the issue of relevant sentence selection is discussed, and then top- ical and opinionated information are summarized. opinion sum- marizations are visualized by representative sentences. finally, an opinionated curve showing supportive and non-supportive de- gree along the timeline is illustrated by an opinion tracking system. . . opinion reason mining in opinion analysis area, finding the polarity of opinions or aggregating and quantifying degree assessment of opinions h. tang et al. / expert systems with applications ( ) – scattered throughout web pages is not enough. we can do more critical part of in-depth opinion assessment, such as finding rea- sons in opinion-bearing texts. for example, in film reviews, infor- mation such as ‘‘found positive reviews and negative reviews” may not fully satisfy the information needs of different people. more useful information would be ‘‘this film is great for its novel originality” or ‘‘poor acting, which makes the film awful”. opinion reason mining tries to identify one of the critical ele- ments of online reviews to answer the question, ‘‘what are the rea- sons that the author of this review likes or dislikes the product?” to answer this question, we should extract not only sentences that contain opinion-bearing expressions, but also sentences with rea- sons why an author of a review writes the review (cardie, wiebe, wilson, & litman, ; clarke & terra, ; li & yamanishi, ; stoyanov, cardie, litman, & wiebe, ). kim and hovy ( ) proposed a method for detecting opinion- bearing expressions. in their subsequent work (kim & hovy, ), they collected a large set of hreview text, pros, consi triplets from epinions.com, which explicitly state pros and cons phrases in their respective categories by each review’s author along with the re- view text. their automatic labeling system first collects phrases in pro and con fields and then searches the main review text in or- der to collect sentences corresponding to those phrases. then the system annotates this sentence with the appropriate ‘‘pro” or ‘‘con” label. all remaining sentences with neither label are marked as ‘‘neither”. after labeling all the data, they use it to train their pro and con sentence recognition system. . . other applications thomas, pang, and lee ( ) try to determine from the tran- scripts of us congressional floor debates whether the speeches rep- resent support of or opposition to proposed legislation. mullen and malouf ( ) describe a statistical sentiment analysis method on political discussion group postings to judge whether there is oppos- ing political viewpoint to the original post. moreover, there are some potential applications of sentiment detection, such as online message sentiment filtering, e-mail sentiment classification, web- blog author’s attitude analysis, sentiment web search engine, etc. . subjectivity classification subjectivity classification is a task to investigate whether a par- agraph presents the opinion of its author or reports facts. in fact, most of the research showed there was very tight relation between subjectivity classification and document sentiment classification (pang & lee, ; wiebe, ; wiebe, bruce, & o’hara, ; wiebe, wilson, bruce, bell, & martin, ; yu & hatzivassiloglou, ). subjectivity classification can prevent the polarity classifier from considering irrelevant or even potentially misleading text. pang and lee ( ) find subjectivity detection can compress re- views into much shorter extracts that still retain polarity informa- tion at a level comparable to that of the full review. much of the research in automated opinion detection has been performed and proposed for discriminating between subjective and objective text at the document and sentence levels (bruce & wiebe, ; finn, kushmerick, & smyth, ; hatzivassiloglou & wiebe, ; wiebe, ; wiebe et al., ; wiebe et al., ; yu & hatzivassiloglou, ). in this section, we will discuss some approaches used to automatically assign one document as objective or subjective. . . similarity approach similarity approach to classifying sentences as opinions or facts explores the hypothesis that, within a given topic, opinion sen- tences will be more similar to other opinion sentences than to fac- tual sentences (yu & hatzivassiloglou, ). similarity approach measures sentence similarity based on shared words, phrases, and wordnet synsets (dagan, shaul, & markovitch, ; dagan, pereira, & lee, ; leacock & chodorow, ; miller & charles, ; resnik, ; zhang, xu, & callan, ). to measure the overall similarity of a sentence to the opinion or fact documents, we need to go through three steps. first, use ir method to acquire the documents that are on the same topic as the sentence in question. second, calculate its similarity scores with each sentence in those documents and make an average va- lue. third, assign the sentence to the category (opinion or fact) for which the average value is the highest. alternatively, for the frequency variant, we can use the similarity scores or count how many of them for each category, and then compare it with a prede- termined threshold. . . naive bayes classifier naive bayes classifier is a commonly used supervised machine learning algorithm. this approach presupposes all sentences in opinion or factual articles as opinion or fact sentences. naive bayes uses the sentences in opinion and fact documents as the examples of the two categories. the features include words, bigrams, and trigrams, as well as the part of speech in each sen- tence. in addition, the presence of semantically oriented (positive and negative) words in a sentence is an indicator that the sentence is subjective. therefore, it can include the counts of positive and negative words in the sentence, as well as counts of the polarities of sequences of semantically oriented words (e.g., ‘‘++” for two con- secutive positively oriented words). it also include the counts of parts of speech combined with polarity information (e.g., ‘‘jj+” for positive adjectives), as well as features encoding the polarity (if any) of the head verb, the main subject, and their immediate modifiers. generally speaking, naive bayes assigns a document dj (repre- sented by a vector d�j ) to the class ci that maximizes pðcijd � j Þ by applying bayes’ rule as follow, pðcijd � j Þ¼ pðciÞpðd � j jciÞ pðd�j Þ ð Þ where pðd�j Þ is the probability that a randomly picked document d has vector d�j as its representation, and p(c) is the probability that a randomly picked document belongs to class c. to estimate the term pðd�j jcÞ, naive bayes decomposes it by assuming all the features in d�j (represented by fi,i = to m) are con- ditionally independent, i.e., pðcijd � j Þ¼ pðciÞ qm i¼ pðfijciÞ � � pðd�j Þ ð Þ . . multiple naive bayes classifier the hypothesis of all sentences in opinion or factual articles as opinion or fact sentences is an approximation. to address this, multiple naive bayes classifier approach applies an algorithm using multiple classifiers, each relying on a different subset of fea- tures. the goal is to reduce the training set to the sentences that are most likely to be correctly labeled, thus boosting classification accuracy. given separate sets of features f , f , . . . , fm, it train separate na- ive bayes classifiers c , c , . . . , cm corresponding to each feature set. assuming as ground truth the information provided by the docu- ment labels and that all sentences inherit the status of their docu- ment as opinions or facts, it first train c on the entire training set, h. tang et al. / expert systems with applications ( ) – then use the resulting classifier to predict labels for the training set. the sentences that receive a label different from the assumed truth are then removed, and train c on the remaining sentences. this process is repeated iteratively until no more sentences can be removed. yu and hatzivassiloglou ( ) report results using five feature sets, starting from words alone and adding in bigrams, trigrams, part-of-speech, and polarity. . . cut-based classifier cut-based classifier approach put forward a hypothesis that, text spans (items) occurring near each other (within discourse boundaries) may share the same subjectivity status (pang & lee, ). based on this hypothesis, pang supplied his algorithm with pair-wise interaction information, e.g., to specify that two particu- lar sentences should ideally receive the same subjectivity label. this algorithm uses an efficient and intuitive graph-based formula- tion relying on finding minimum cuts. suppose there are n items x , x , . . . , xn to divide into two classes c and c , here access to two types of information: indj(xi): individual scores. it is the non-negative estimates of each xi’s preference for being in cj based on just the features of xi alone; assoc(xi, xk): association scores. it is the non-negative estimates of how important it is that xi and xk be in the same class. then, this problem changes to calculate the maximization of each item’s score for one class: its individual score for the class it is assigned to, minus its individual score for the other class, then minus associated items into different classes for penalization. thus, after some algebra, it arrives at the following optimization problem: assign the xi to c and c so as to minimize the partition cost:x x c ind ðxÞþ x x c ind ðxÞþ x xi c ;xk c assocðxi; xkÞ ð Þ this situation can be represented in the following manner. build an undirected graph g with vertices {v , . . . , vn, s, t}; the last two are, respectively, the source and sink. add n edges (s, vi), each with weight ind (xi), and n edges (vi, t), each with weight ind (xi). finally, add ðc nÞ edges (vi, vk), each with weight assoc(xi, xk). a cut (s, t) of g is a partition of its nodes into sets s = {s}us and t = {t}ut , where s r s , t r t . its cost cost(s, t) is the sum of the weights of all edges crossing from s to t. a minimum cut of g is one of minimum cost. then, finding solution of this problem is changed into looking for a minimum cut of g. . word sentiment classification the task on document sentiment classification has usually in- volved the manual or semi-manual construction of semantic orien- tation word lexicons (hatzivassiloglou & mckeown, ; hatzivassiloglou & wiebe, ; lin, ; pereira, tishby, & lee, ; riloff, wiebe, & wilson, ; turney & littman, ; wiebe, ), which built by word sentiment classification tech- niques. for instance, das and chen ( ) used a classifier on investor bulletin boards to see if apparently positive postings were correlated with stock price, in which several scoring methods were employed in conjunction with a manually crafted lexicon. classify- ing the semantic orientation of individual words or phrases, such as whether it is positive or negative or has different intensities, generally using a pre-selected set of seed words, sometimes using linguistic heuristics (for example, lin ( ) & pereira et al. ( ) used linguistic co-locations to group words with similar uses or meanings). some studies showed that restricting features to those adjec- tives for word sentiment classification would improve perfor- mance (andreevskaia & bergler, ; turney & littman, ; wiebe, ). however, more researches showed most of the adjectives and adverb, a small group of nouns and verbs possess semantic orientation (andreevskaia & bergler, ; esuli & sebas- tiani, ; gamon & aue, ; takamura, inui, & okumura, ; turney & littman, ). automatic methods of sentiment annotation at the word level can be grouped into two major categories: ( ) corpus-based ap- proaches and ( ) dictionary-based approaches. the first group in- cludes methods that rely on syntactic or co-occurrence patterns of words in large texts to determine their sentiment (e.g., hatzi- vassiloglou & mckeown, ; turney & littman, ; yu & hat- zivassiloglou, and others). the second group uses wordnet (http://wordnet.princeton.edu/) information, especially, synsets and hierarchies, to acquire sentiment-marked words (hu & liu, a; kim & hovy, ) or to measure the similarity between candidate words and sentiment-bearing words such as good and bad (kamps, marx, mokken, & de rijke, ). . . analysis by conjunctions between adjectives this method attempts to predict the orientation of subjective adjectives by analyzing pairs of adjectives (conjoined by and, or, but, either-or, or neither-nor) which are extracted from a large unlabelled document set. the underlying intuition is that the act of conjoining adjectives is subject to linguistic constraints on the orientation of the adjectives involved (e.g. and usually conjoins two adjectives of the same-orientation, while but conjoins two adjectives of opposite orientation). this is shown in the following three sentences (where the first two are perceived as correct and the third is perceived as incorrect) taken from hatzivassiloglou and mckeown ( ): ‘‘the tax proposal was simple and well received by the public”. ‘‘the tax proposal was simplistic but well received by the public”. ‘‘the tax proposal was simplistic and well received by the public”. to infer the orientation of adjectives from analysis of conjunc- tions, a supervised learning algorithm can be performed as follow- ing steps: . all conjunctions of adjectives are extracted from a set of documents. . train a log-linear regression classifier and then classify pairs of adjectives either as having the same or as having different ori- entation. the hypothesized same-orientation or different-orien- tation links between all pairs form a graph. . a clustering algorithm partitions the graph produced in step into two clusters. by using the intuition that positive adjectives tend to be used more frequently than negative ones, the cluster containing the terms of higher average frequency in the docu- ment set is deemed to contain the positive terms. the log-linear model offers an estimate of how good each pre- diction is, since it produces a value y between and , in which corresponds to same-orientation, and one minus the produced value y corresponds to dissimilarity. same- and different-orienta- tion links between adjectives form a graph. to partition the graph nodes into subsets of the same-orientation, the clustering algo- rithm calculates an objective function u scoring each possible par- tition p of the adjectives into two subgroups c and c as, uðpÞ¼ x i¼ jcij x x;y ci;x – y dðx; yÞ ! ð Þ where jcij is the cardinality of cluster i, and d(x, y) is the dissimilarity between adjectives x and y. http://wordnet.princeton.edu/ h. tang et al. / expert systems with applications ( ) – in general, because the model was unsupervised, it required an immense word corpus to function. . . analysis by lexical relations this method presents a strategy for inferring semantic orienta- tion from semantic association between words and phrases. it fol- lows a hypothesis that two words tend to be the same semantic orientation if they have strong semantic association. therefore, it focused on the use of lexical relations defined in wordnet to calcu- late the distance between adjectives. generally speaking, we can defined a graph on the adjectives contained in the intersection between a term set (for example, tl term set (turney & littman, )) and wordnet, adding a link between two adjectives whenever wordnet indicates the presence of a synonymy relation between them, and defining a distance measure using elementary notions from graph theory. in more de- tail, this approach can be realized as following steps: . construct relations at the level of words. the simplest approach here is just to collect all words in wordnet, and relate words that can be synonymous (i.e., they occurring in the same synset). . define a distance measure d(t , t ) between terms t and t on this graph, which amounts to the length of the shortest path that connects t and t (with d(t , t ) = + if t and t are not connected). . calculate the orientation of a term by its relative distance (kamps et al., ) from the two seed terms good and bad, i.e., soðtÞ¼ dðt; badÞ� dðt; goodÞ dðgood; badÞ ð Þ . get the result followed by this rules: the adjective t is deemed to belong to positive if so(t) > , and the absolute value of so(t) determines, as usual, the strength of this orientation (the con- stant denominator d(good, bad) is a normalization factor that constrains all values of so to belong to the [� , ] range). . . analysis by glosses the characteristic of this method lies in the fact that it exploits the glosses (i.e. textual definitions) that one term has in an online ‘‘glossary”, or dictionary. its basic assumption is that if a word is semantically oriented in one direction, then the words in its gloss tend to be oriented in the same direction (esuli & sebastiani, ; esuli & sebastiani, a, b). for instance, the glosses of good and excellent will both contain appreciative expressions; while the glosses of bad and awful will both contain derogative expressions. generally, this method can determine the orientation of a term based on the classification of its glosses. the process is composed of the following steps: . a seed set (sp, sn), representative of the two categories positive and negative, is provided as input. . search new terms to enrich sp and sn. use lexical relations (e.g. synonymy) with the terms contained in sp and sn from a thesau- rus, or online dictionary, to find these new terms, and then append them to sp or sn. . for each term ti in s p [ s n or in the test set (i.e. the set of terms to be classified), a textual representation of ti is generated by collating all the glosses of ti as found in a machine-readable dic- tionary. each such representation is converted into a vector by standard text indexing techniques. . a binary text classifier is trained on the terms in s p [ s n and then applied to the terms in the test set. . . analysis by both lexical relations and glosses this method determines sentiment of words and phrases both relies on lexical relations (synonymy, antonymy and hyponymy) and glosses provided in wordnet. andreevskaia and bergler ( ) proposed an algorithm named ‘‘step” (semantic tag extraction program). this algorithm starts with a small set of seed words of known sentiment value (positive or negative) and implements the following steps: . extend the small set of seed words by adding synonyms, ant- onyms and hyponyms of the seed words supplied in wordnet. this step brings on average a -fold increase in the size of the original list with the accuracy of the resulting list comparable to manual annotations. . go through all wordnet glosses, identifies the entries that con- tain in their definitions the sentiment-bearing words from the extended seed list, and adds these head words to the corre- sponding category – positive, negative or neutral. . disambiguate the glosses with part-of-speech tagger, and elim- inate errors of some words acquired in step and from the seed list. at this step, it also filters out all those words that have been assigned contradicting. in this algorithm, for each word we need compute a net overlap score by subtracting the total number of runs assigning this word a negative sentiment from the total of the runs that consider it posi- tive. in order to make the net overlap score measure usable in sen- timent tagging of texts and phrases, the absolute values of this score should be normalized and mapped onto a standard [ , ] interval. step accomplishes this normalization by using the value of the net overlap score as a parameter in the standard fuzzy mem- bership s-function (zadeh, ). this function maps the absolute values of the net overlap score onto the interval from to , where corresponds to the absence of membership in the category of sentiment (in this case, these will be the neutral words) and re- flects the highest degree of membership in this category. the func- tion can be defined as follows, sðu; a; b; cÞ¼ if u a u�ac�a � � if a u b � u�ac�a � � if b u c if u p c >>>>>>< >>>>>>: ð Þ where u is the net overlap score for the word and a, b, c are the three adjustable parameters: a is set to , c is set to and b, which represents a crossover point, is defined as b = (a + c)/ = . defined this way, the s-function assigns highest degree of membership (= ) to words that have the net overlap score u p . net overlap score can be used as a measure of the words degree of membership in the fuzzy category of sentiment: the core adjec- tives, which had the highest net overlap score, were identified most accurately both by step and by human annotators, while the words on the periphery of the category had the lowest scores and were associated with low rates of inter-annotator agreement. . . analysis by pointwise mutual information the general strategy of this method is to infer semantic orienta- tion from semantic association. the underlying assumption is that a phrase has a positive semantic orientation when it has good asso- ciations (e.g., ‘‘romantic ambience”) and a negative semantic orien- tation when it has bad associations (e.g., ‘‘horrific events”) (turney, ). h. tang et al. / expert systems with applications ( ) – the semantic orientation of a given word is calculated from the strength of its association with a set of positive words, minus the strength of its association with a set of negative words. more con- cretely, the strength of the semantic association between words can express by calculating their pointwise mutual information (pmi) value. so, it focuses on inferring the semantic orientation of a word from its statistical association with a set of positive and negative paradigm words. given a term t, and seed term sets (sp for positive set and sn for negative set), the t’s orientation value o(t) (where positive value means positive orientation, and higher absolute value means stronger orientation) is given by: oðtÞ¼ x ti sp pmiðt; tiÞ� x ti sn pmiðt; tiÞ ð Þ pointwise mutual information can be computed based on ir tech- niques. term frequencies and co-occurrence frequencies are mea- sured by querying a document set by means of a search engine with a ‘‘t” query, a ‘‘ti” query, and a ‘‘t near ti” query, and using the number of matching documents returned by the search engine as estimates of the probabilities needed for the computation of pmi. in the altavista search engine (http://www.altavista.com/), the near operator produces a match for a document when its oper- ands appear in the document at a maximum distance of ten terms, in either order. the paradigm words can be selected as following (turney & littman, ): sp ¼fgood; nice; excellent; positive; fortunate; correct; superiorg sn ¼fbad; nasty; poor; negative; unfortunate; wrong; inferiorg in addition, gamon and aue ( ) described an extension to the technique for the automatic identification and labeling of sen- timent terms described in turney and littman ( ). besides the basic assumption in turney and littman ( ), turney ( ), gamon and aue ( ) adds a second assumption, namely that sentiment terms of opposite orientation tend not to co-occur at the sentence level. this additional assumption allows them to identify sentiment-bearing terms more reliably to some extent. . . analysis by general inquirer general inquirer (gi) is a system which lists terms as well as dif- ferent senses for the terms. for each sense it provides a short def- inition as well as other information about the term. this includes tags that label the term as being positive, negative, a negation term, an overstatement, or an understatement. the labels are for each sense of a word. for example, there are two senses of the word fun as seen in table . one sense is a noun or adjective for enjoyment or enjoyable. the second sense is a verb that means to ridicule or tease, to make fun of. the first sense of the word is positive, while the second is negative. the entry also indicates that the first sense is more fre- quent than the second sense (estimated to occur % of the times while the second sense occurs only % of the times). in addition, the gi dictionary includes negations, intensifiers, and diminishers. table shows the gi entries of the words not, fantastic and barely which are examples of a negation, an overstatement and an understatement, respectively. the gi dictionary contains , positive senses and , neg- ative senses. kennedy and inkpen ( ) add more positive and negative senses from choose the right word (hayakawa, ) table gi entries for the word fun. fun (sense ) h lvd positiv pstv pleasur ex fun (sense ) h lvd negativ ngtv hostile c (hereafter ctrw). ctrw is a dictionary of synonyms, which lists nuances of lexical meaning. after adding them, they obtain positive senses and negative senses. there are overstate- ments and understatements in gi. when adding those from ctrw, they obtain overstatements and understatements. kennedy and inkpen ( ) used this approach to classify reviews based on the number of positive and negative terms they contain. they examined the effect of three types of valence shifters: negations, intensifiers, and diminishers. in: gi intensifiers are known as overstate- ments and diminishers are known as understatements. negations are used to reverse the semantic polarity of a particular term, while intensifiers and diminishers are used to increase and decrease, respec- tively, the degree to which a term is positive or negative. . document sentiment classification based on machine learning methods . . single domain there are many possible approaches to identifying the actual polarity of a document. here we talk about using supervised ma- chine learning techniques to identify the likelihood of reviews hav- ing ‘‘positive” or ‘‘negative” polarity with respect to previously hand-classified training data. the key problem of this method in- cludes two aspects: extracting feature and training classifier. . . . extracting feature starting with a raw document (a portion of a web page in test- ing and training, and a complete web page for mining), then strip out html tags and separate the document into sentences. these sentences are optionally run through a parser before being split into single-word tokens. a variety of transformations can then be applied to this ordered list of lists. this is called feature extraction. there are several methods for feature extraction: . lexical filtering lexical filtering can apply to reviews, on the accuracy of statis- tical classifiers trained on such filtered data. there are mainly two different kinds of lexical filters used (salvetti, lewis, & rei- chenbach, ): one based on hypernymy as provided by wordnet (budanitsky & hirst, ; curran, ; devitt & vogel, ; edmonds, ; edmonds & hirst, ; fellbaum, ; grefenstette, ; justeson & slava, ; kamps & maarten, ; rapp, ; riloff & shepherd, ; roark & charniak, ; taboada, ; thelen & riloff, ; valitutti, strapparava, & stock, ), the other based on part-of-speech (pos) tags (ait-mokhtar, chanod, & roux, ; brill, ; brill, ; brill, ; losee, ; ratnaparkhi, ; schmid, ; wiebe, wilson, & bell, ; wiebe, wilson, & cardie, ; wilks & stevenson, ). wordnet filter attempted to substitute synonyms in reviews by a set of likely synonyms and hypernymy generalization in wordnet, because it is uncommon to encounter repetitions of identical words in non-technical written text. pos filter considered that individual phrases and words vary in their contribution to opinion polarity. it may even be said that only some part of the meaning of a word contributes to opinion polarity. any portion that does not contribute to sentiment detection is noise. pos filters were developed to reduce these noises. prsv wlbpsyc wlbtot noun pfreq % noun–adj: enjoyment, enjoyable omform sv rsploss rsptot supv % idiom–verb: make fun (of) – to tease, parody http://www.altavista.com/ table gi entries for the words not, fantastic and barely. not h lvd negate notlw ly – adv: expresses negation notlw ly adv: expresses negation fantastic h lvd positiv pstv virtue ovrst eval posaff modif – virtue ovrst eval posaff modif barely h lvd undrst quan if ly – quan if ly h. tang et al. / expert systems with applications ( ) – . appraising adjective appraising adjective method (whitelaw, argamon, & garg, ; whitelaw, garg, & argamon, ) focuses on the extraction and analysis of adjectival appraisal groups headed by an appraising adjective (such as ‘‘beautiful” or ‘‘boring”) and optionally modified by a sequence of modifiers (such as ‘‘very”, ‘‘sort of”, or ‘‘not”). it made a more detailed semantic analysis of attitude expressions, in the form of a well-designed taxonomy of attitude types and other semantic properties. furthermore, it treats ‘‘atomic units” of such expressions not with individual words, but with appraisal groups: coherent groups of words that express together a particu- lar attitude, such as ‘‘extremely boring”, or ‘‘not really very good”. this method can be described as following steps: . build a lexicon using semi-automatic techniques, gathering and classifying adjectives and modifiers to categories in several taxonomies of appraisal attributes. . extract adjectival appraisal groups from texts and compute their attribute values according to this lexicon. . represent documents as vectors of relative frequency features using these groups. . train a support vector machine algorithm discriminating posi- tively from negatively oriented test documents. beineke, hastie, and vaithyanathan ( ) extend this proce- dure by extract a pair of derived features that are linearly com- bined to predict sentiment. this perspective allows to improving upon previous methods, primarily through two strategies: incorpo- rating additional derived features into the model and, where pos- sible, using labeled data to estimate their relative influence. matsumoto, takamura, and okumura ( ) used text-mining techniques to extract frequent word sub-sequences and depen- dency sub-trees from sentences in a document dataset and used them as features of support vector machines. . . . training classifier several classical text classifiers, such as k-nearest neighbor, winnow, naı̈ve bayes, maximum entropy and support vector ma- chine (svm), are used in machine learning method for document sentiment classification. pang et al. ( ) applied naı̈ve bayes, maximum entropy and support vector machine classification techniques to the identifica- tion of the polarity of movie reviews. their best result ( . % accu- racy) was obtained by using unigrams and support vector machine. in fact, most of the literatures showed that svm and naı̈ve bayes are perfect methods in single domain document sentiment classification (aue and gamon, ; beineke et al., ; kennedy and inkpen, ; lin et al., ; matsumoto et al., ; mullen and collier, ; pang and lee, ; pang et al., ; read et al., ; salvetti et al., ; whitelaw, garg, et al., ). . . multiple domains document sentiment classification is a very domain-specific problem: classifiers trained in one domain do not perform well in others charlotta ( ), nigam, mccallum, and thrun ( ), lafferty, mccallum, and pereira ( ), avrim and chawla ( ). the main factor is that the polarity of sentiment words may change with the domain too. for instance, the word ‘‘small” in house review is negative (e.g. ‘‘the bedroom is very small”), while in cell-phone review is positive (e.g. ‘‘the nokia n is so small as to be put in any pockets”). tan, wu, tang, and cheng ( ) attempted to tackle domain- transfer problem by combining old-domain labeled examples with new domain unlabeled ones. their basic idea is to use old-domain- trained classifier (‘‘old classifier” for brevity) to label top n most informative unlabeled examples in new domain and learn a new classifier based on these selected examples (n is a pre-defined number indicating how many examples in new domain shall be picked out as informative ones). detailed algorithm for proposed scheme is: . train a base classifier using labeled data in old-domain. . label some informative unlabeled ones in new domain. . train a new classifier based on these selected examples. . classify examples in new domain using new classifier. the experimental results demonstrate their proposed scheme can boost the accuracy of the base sentiment classifier on new domain. aue and gamon ( ) surveyed four different approaches to customizing a sentiment classification system to a new target do- main with a small amount of labeled data. read et al. ( ) pro- posed a novel source of training data based on the language used in conjunction with emoticons in usenet newsgroups. then they used emoticons-labeled data to train a classifier, which has the po- tential to reduce dependent of domain, topic and time. . opinion extraction analysis of favorable and unfavorable opinions is a task requir- ing high intelligence and deep understanding of the textual con- text, drawing on common sense and domain knowledge as well as linguistic knowledge. the interpretation of opinions can be debatable even for humans. for example, when we tried to deter- mine if each specific document was on balance favorable or unfa- vorable toward a subject after reading an entire group of such documents, we often found it difficult to reach a consensus, even for very small groups of evaluators. many researchers think it is too coarse to compute a unit of an opinion for a whole document (bai, padman, & airoldi, ; bet- hard, yu, thornton, hatzivassiloglou, & danjurafsky, ; bune- scu & mooney, ; daille, ; didier, ; etzioni et al., ; freitag & mccallum, ; jacquemin & bourigault, ; luca & mazzini, ; riloff & jones, ; wilson, wiebe, & hwa, ; zhai & liu, ). turney ( ) makes a similar point, noting that for reviews, ‘‘the whole is not necessarily the sum of the parts”. even in a single sentence, a holder might express two different opinions. it seems necessary using more sophisticated techniques to determine the focus of each sentence, so that one can decide whether the author is talking about the topic (dave, lawrence, & pennock, ; fei, liu, & wu, ; kim & hovy, ; ku, wu, lee, & chen, ; mei, ling, wondra, su, & zhai, ). h. tang et al. / expert systems with applications ( ) – consequently, opinion extraction plays a very important role in sentiment detection. it not only focuses on extracting opinion information from reviews, but also on extracting relation between opinion and document topic. . . opinion information extraction the opinion information we mainly discussed in this paper in- cludes opinion-bearing word and opinion holder. an opinion-bear- ing word is a word or a phrase that carries a positive or negative sentiment directly such as ‘‘good”, ‘‘bad”, ‘‘foolish”, ‘‘virtuous”, etc. an opinion holder is an entity (person, organization, country, or special group of people) who expresses explicitly or implicitly the opinion contained in the sentence. for instance: ‘‘according to the review in internet, the keypad of ericsson w i is easy to use”. in above sentence, ‘‘the review in internet” is an opinion holder; ‘‘easy” is an opinion-bearing word. . . . opinion-bearing word extraction opinions can be recognized from various granularities such as a word, a sentence, a text, or even multiple texts, and each is impor- tant. here we focus on word level opinion detection, i.e., finding words or phrases that carry a positive or negative sentiment (opin- ion-bearing word) from subjectivity sentences or paragraphs. actually, opinion-bearing word is the smallest unit of opinion that can thereafter be used as a clue for sentence-level or text-level opinion detection. how to extract opinion-bearing words from document? a straightforward way can be conducted by following steps: . collect sentiment words from sentiment lexicon as opinion- bearing seed words. opinion-bearing seed words can be col- lected from several sources: general inquirer (gi), dictionary of affect of language (dal), and wordnet (kim & hovy, ; yi, nasukawa, bunescu, & niblack, ). . expand those selected opinion-bearing seed words of each sen- timent class by collecting synonyms from wordnet. however, it cannot simply assume that all the synonyms of positive words are positive since most words could have synonym relation- ships with both positive and negative classes. it should calcu- late the closeness of a given word to each category and determine the most probable class (kim & hovy, ). . refine some of the sentiment patterns from sentiment lexicon and training datasets (yi et al., ). for gi and dal, the sen- timent verb extraction is the same as the opinion-bearing seed words extraction. for wordnet, the sentiment verb can be extracted from the emotion cluster. the other sentiment pat- terns can be manually refined from the training datasets. senti- ment pattern database contains sentiment extraction patterns for sentence predicates. . extract sentences and text fragments from input documents containing subjectivity information (the approach we have dis- cussed in section ). apply sentiment analysis with expanded opinion-bearing seed words and sentiment patterns to those subjectivity sentences and text fragments. at last, the opin- ion-bearing words were extracted. . . . opinion holder extraction the goal of opinion holder extraction is to identify direct and indirect sources of opinions, emotions, sentiments, and other pri- vate states that are expressed in text. identifying opinion sources (or opinion holder) will be especially critical for opinion-oriented question–answering systems (e.g., systems that answer questions of the form ‘‘who feels about . . .?”) and opinion-oriented summa- rization systems, both of which need to distinguish the opinions of one source from another. there are mainly three research points in opinion holder extrac- tion. the first is: given a sentence, identifying opinion sources in it (choi, cardie, riloff, & patwardhan, ); the second is: given an opinion expression in a sentence, identifying its corresponding opinion sources (kim & hovy, ); the third is: given an opinion expression and a source, determines whether the source and the opinion expression is correspond (choi, breck, & cardie, ). . the first research point the first research point learns patterns of opinion sources using a graphical model and extraction pattern learning. it views the opin- ion holder extraction problem as an information extraction task and adopts a hybrid approach that combines crfs (conditional random fields) and a variation of autoslog (a supervised extrac- tion pattern learner that takes a training corpus of texts and their associated answer keys as input) (riloff, ). while crfs identi- fies source and autoslog learns extraction patterns.it starts with a sequence of words (x , x , . . . , xn) in a sentence and then: ( ) generate a sequence of labels (y , y , . . . , yn) indicating whether the word is a holder or not. ( ) presents a new variation of autoslog, autoslog-se, which generates patterns to extract sources. this algorithm is not perfect, however, so the resulting set of patterns needs to be manually reviewed by a person. in order to build a fully automatic system which has no use for manual review, choi et al. ( ) combined autoslog’s heuristics with statistics from the annotated training data to create a fully auto- matic supervised learner. . the second research pointthe second research point uses clas- sification and ranking to model the problem with maximum entropy (me) (kim & hovy, ). classification allocates each holder candidate to a set of pre-defined classes while ranking selects a single candidate as answer. it chooses the most prob- able candidate via a conditional probability.this method acts as following steps. ( ) generate all possible holder candidates, given a sentence and an opinion expression hei. after parsing the sentence, it extracts features such as the syntactic path information between each candidate hhi and the expression hei and a distance between hhi and hei. ( ) rank holder candidates according to the score obtained by the me ranking model, and pick the candidate with the highest score. given a set of holder candidates {h , h , . . . , hn} and opin- ion expression e. the conditional probability p(hj{h , h , . . . , hn}, e) can be calculated based on k feature functions fk(hj{h , h , . . . , hn}, e), as follows, h ¼ argmaxh½pðhjfh ; h ; . . . ; hng; eÞ� ¼ argmaxsh xk k¼ kk fkðhjfh ; h ; . . . ; hng; eÞ " # ð Þ where each kk is a model parameter indicating the weight of its feature function. . the third research point the third research point (choi et al., ) aimed to identify the opinion holder via extracting relations between opinion expres- sion entities and source entities. that is, given opinion expres- sion oi and source sj, it determines whether sj is the source of opinion expression oi. the global inference procedure is imple- mented via integer linear programming (ilp) to produce an optimal and coherent extraction of entities and relations. ilp formulation group consists of an objective function and a set of equality and inequality constraints among variables. the fol- lowing formulation is the objective function of binary ilp, h. tang et al. / expert systems with applications ( ) – f ¼ x i ðwoi oiÞþ x i ðw�oi o � i Þþ x j ðwsj sjÞþ x j ðw�sj s � j Þ þ x i;j ðwli;j li;jÞþ x i ðw�li;j l � i;jÞ ð Þ i; oi þ o�i ¼ ; j; sj þ s � i ¼ ; i; j; li;j þ l � i;j ¼ ð Þ where f is the objective function of ilp. oi and o � i are two variables for each opinion entity, oi = means to extract the opinion entity, and o�i ¼ means to discard the opinion entity. woi and w�oi are weights for oi and o � i , respectively, which are computed based on the labels of the adjacent variables of the crfs (choi et al., ). likewise, sj and s � j are two variables for each source entity to indi- cate whether it had been extracted, their weights wsj and w�sj are computed in the same way as opinion entities; li,j and l � i;j are two variables for each link relation between oi and sj. li,j = indicate oi and sj both be extracted, otherwise, l � i;j ¼ . their weights wlij and w�lij are based on probabilities from the binary link classifier.the following is a set of equality and inequality constraints among variables. i; oi ¼ x j li;j ð Þ j; sj þ aj ¼ x i li;j ð Þ j; aj � sj ð Þ i; j; i < j; xi þ xj ¼ ; x fs; og ð Þ formula ( ) enforces that only one link can emanate from an opin- ion entity. formulas ( ) and ( ) together allow a source to link to at most two opinions, where aj is an auxiliary variable between and . aj can be assigned to only if sj is already assigned to . for- mula ( ) is used to restrict all pairs of entities with overlapping spans. . . opinion-topic relation extraction opinion-topic relation refers to relationship between opinion expression (opinion-bearing words) and document topic (or fea- ture of topic). a feature of a topic is a term that satisfies one of the following relationships: � a part-of relationship with the given topic. � an attribute-of relationship with the given topic. � an attribute-of relationship with a known feature of the given topic. for instance, ‘‘i found the sony ericsson w i is a good mobile phone. the keypad is easy to use and texting is very simple as the buttons are small but well defined. the screen is bright and clear with good resolution. the software on this phone is excellent. i have had absolutely no problems with it what so ever, it never crashes, freezes or otherwise upsets me. the battery life on this phone is excellent. i have had mine for months, with it rarely being switched off in all this time. in my normal use (only about – min calls a day) it will last for about – days before needing a charge. charging is very quick too taking only around an hour to fully charge from about %. all in all an excellent little phone with very few faults. recommended!” in above text, ‘‘good”, ‘‘easy to use”, ‘‘bright” etc., are opinion expression, ‘‘sony ericsson w i” is topic, and ‘‘keypad”, ‘‘screen”, ‘‘software” etc., are features of topic. . . . feature term extraction yi et al. ( ) put forward the following three candidates as feature term to be extracted: . base noun phrases (bnp). bnp restricts the candidate feature terms to one of the following base noun phrase (bnp) patterns: nn, nn nn, jj nn, nn nn nn, jj nn nn, jj jj nn, where nn and jj are the part-of-speech (pos) tags for nouns and adjectives, respectively. . definite base noun phrases (dbnp). dbnp further restricts can- didate feature terms to definite base noun phrases, which are noun phrases of the form defined above that are preceded by the definite article ‘‘the”. given that a document is focused on a certain topic, the definite noun phrases referring to topic fea- tures do not need any additional constructs such as attached prepositional phrases or relative clauses, in order for the reader to establish their referent. thus, the phrase ‘‘the battery,” instead of ‘‘the battery of the digital camera,” is sufficient to infer its referent. . beginning definite base noun phrases (bbnp). bbnp refers to dbnp at the beginning of sentences followed by a verb phrase. this heuristic is based on the observation that, when the focus shifts from one feature to another, the new feature is often expressed using a definite noun phrase at the beginning of the next sentence. they developed and tested two feature term selection algo- rithms based on a mixture language model and likelihood ratio, while the likelihood test method gets better result. following is principle of the likelihood test method: let d+ be a collection of documents focused on a topic t, d� those not focused on t, and bnp a candidate feature term extracted from d+. then, the likeli- hood ratio � log k is defined as follows, � log k ¼� log maxp p lðp ; p Þ maxp ;p lðp ; p Þ p ¼ pðd dþjbnp dÞ p ¼ pðd dþjbnp dÞ ð Þ where l(p , p ) is the likelihood of seeing bnp in both d+ and d�. the higher the value of � log k, the more likely the bnp is relevant to the topic t. for each bnp, compute the likelihood score, � log k, as de- fined in formula ( ). then, sort bnp in decreasing order by their likelihood score. feature terms are all bnp’s whose likelihood ratio satisfying a pre-defined confidence level. alternatively simply only the top n bnp’s can be selected. . . . makes (topic j feature term, opinion) association in order to achieve high precision, we need focus on identifying semantic relationships between opinion expressions and topic (or feature of topic) terms, i.e. extracting opinion-bearing words asso- ciated with polarity of positive or negative for specific topic (or fea- ture of topic) from a document (agrawal & srikant, ; jon & tardos, ; liu, hsu, & ma, ; rosario & hearst, ). in order to identify opinion expressions and analyze their semantic relationships with the topic (or feature of topic) term, fol- lowing natural language processing techniques play an important role (nasukawa & yi, ): . pos tagging pos tagging can disambiguate some polysemous expressions such as ‘‘like,” which denote sentiment only when used as a verb instead of as an adjective or preposition. . syntactic parsing syntactic parsing is used to identify relationships between senti- ment expressions and the subject term. furthermore, in order to h. tang et al. / expert systems with applications ( ) – maintain robustness for noisy texts from various sources such as the www, it need to use a shallow parsing framework that iden- tifies phrase boundaries and their local dependencies in addition to pos tagging, instead of using a full parser that tries to identify the complete dependency structure among all of the terms. yi et al. ( ) extracts ternary expressions (t-expressions) and binary expressions (b-expressions), in order to make (topic j feature term, opinion) association. there are two types of t-expressions: . positive or negative sentiment verbs:h> /colorimagedict << /qfactor . /hsamples [ ] /vsamples [ ] >> /jpeg coloracsimagedict << /tilewidth /tileheight /quality >> /jpeg colorimagedict << /tilewidth /tileheight /quality >> /antialiasgrayimages false /cropgrayimages true /grayimageminresolution /grayimageminresolutionpolicy /warning /downsamplegrayimages true /grayimagedownsampletype /bicubic /grayimageresolution /grayimagedepth - /grayimagemindownsampledepth /grayimagedownsamplethreshold . /encodegrayimages true /grayimagefilter /dctencode /autofiltergrayimages true /grayimageautofilterstrategy /jpeg /grayacsimagedict << /qfactor . /hsamples [ ] /vsamples [ ] >> /grayimagedict << /qfactor . /hsamples [ ] /vsamples [ ] >> /jpeg grayacsimagedict << /tilewidth /tileheight /quality >> /jpeg grayimagedict << /tilewidth /tileheight /quality >> /antialiasmonoimages false /cropmonoimages true /monoimageminresolution /monoimageminresolutionpolicy /warning /downsamplemonoimages true /monoimagedownsampletype /bicubic /monoimageresolution /monoimagedepth - /monoimagedownsamplethreshold . /encodemonoimages true /monoimagefilter /ccittfaxencode /monoimagedict << /k - >> /allowpsxobjects false /checkcompliance [ /none ] /pdfx acheck false /pdfx check false /pdfxcompliantpdfonly false /pdfxnotrimboxerror true /pdfxtrimboxtomediaboxoffset [ . . . . ] /pdfxsetbleedboxtomediabox true /pdfxbleedboxtotrimboxoffset [ . . . . ] /pdfxoutputintentprofile (none) /pdfxoutputconditionidentifier () /pdfxoutputcondition () /pdfxregistryname () /pdfxtrapped /false /createjdffile false /description << /ara /bgr /chs /cht /cze /dan /esp /eti /fra /gre /heb /hrv (za stvaranje adobe pdf dokumenata najpogodnijih za visokokvalitetni ispis prije tiskanja koristite ove postavke. stvoreni pdf dokumenti mogu se otvoriti acrobat i adobe reader . i kasnijim verzijama.) /hun /ita /jpn /kor /lth /lvi /nld (gebruik deze instellingen om adobe pdf-documenten te maken die zijn geoptimaliseerd voor prepress-afdrukken van hoge kwaliteit. de gemaakte pdf-documenten kunnen worden geopend met acrobat en adobe reader . en hoger.) /nor /pol /ptb /rum /rus /sky /slv /suo /sve /tur /ukr /enu (use these settings to create adobe pdf documents best suited for high-quality prepress printing. created pdf documents can be opened with acrobat and adobe reader . and later.) /deu >> /namespace [ (adobe) (common) ( . ) ] /othernamespaces [ << /asreaderspreads false /cropimagestoframes true /errorcontrol /warnandcontinue /flattenerignorespreadoverrides false /includeguidesgrids false /includenonprinting false /includeslug false /namespace [ (adobe) (indesign) ( . ) ] /omitplacedbitmaps false /omitplacedeps false /omitplacedpdf false /simulateoverprint /legacy >> << /addbleedmarks false /addcolorbars false /addcropmarks false /addpageinfo false /addregmarks false /convertcolors /converttocmyk /destinationprofilename () /destinationprofileselector /documentcmyk /downsample bitimages true /flattenerpreset << /presetselector /mediumresolution >> /formelements false /generatestructure false /includebookmarks false /includehyperlinks false /includeinteractive false /includelayers false /includeprofiles false /multimediahandling /useobjectsettings /namespace [ (adobe) (creativesuite) ( . ) ] /pdfxoutputintentprofileselector /documentcmyk /preserveediting true /untaggedcmykhandling /leaveuntagged /untaggedrgbhandling /usedocumentprofile /usedocumentbleed false >> ] >> setdistillerparams << /hwresolution [ ] /pagesize [ . . ] >> setpagedevice a new dynamic modeling framework for credit risk assessment expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications journal homepage: www.elsevier.com/locate/eswa a new dynamic modeling framework for credit risk assessment maria rocha sousa a,∗, joão gama a,b, elísio brandão a a school of economics and management, university of porto, portugal b laboratory of artificial intelligence and decision support of the institute for systems and computer engineering, technology and science, portugal a r t i c l e i n f o keywords: credit risk modeling credit scoring dynamic modeling temporal degradation default concept drift memory a b s t r a c t we propose a new dynamic modeling framework for credit risk assessment that extends the prevailing credit scoring models built upon historical data static settings. the driving idea mimics the principle of films, by composing the model with a sequence of snapshots, rather than a single photograph. in doing so, the dynamic modeling consists of sequential learning from the new incoming data. a key contribution is provided by the insight that different amounts of memory can be explored concurrently. memory refers to the amount of historic data being used for estimation. this is important in the credit risk area, which often seems to undergo shocks. during a shock, limited memory is important. other times, a larger memory has merit. an application to a real-world financial dataset of credit cards from a financial institution in brazil illustrates our methodology, which is able to consistently outperform the static modeling schema. © elsevier ltd. all rights reserved. m m a p i c h i t s a f a b a i i t j m h c r i a u y s a d u t m i u n m c e u n h h . introduction in banking, credit risk assessment often relies on credit scoring odels, so called pd models (probability of default models). these odels output a score that translates the probability of a given entity, private individual or a company, becoming a defaulter in a future eriod. nowadays, pd models are at the core of the banking business, n credit decision-making, in price settlement, and to determine the ost of capital. moreover, central banks and international regulation ave dramatically evolved to a setting where the use of these models s favored, to achieve soundness standards for credit risk valuation in he banking system. since , with the worldwide implementation of regulations is- ued by the basel committee on banking supervision within basel ii ccord, banks were encouraged to strengthen their internal models rameworks for reaching the a-irb (advanced internal rating based) ccreditation (bcbs, ; bis, ). to achieve this certification, anks had to demonstrate that they were capable of accurately evalu- ting their risks, complying with basel ii requirements, by using their nternal risk models’ systems, and keep their soundness. banks own- ng a-irb accreditation gained an advantage over the others, because hey were allowed to use lower coefficients to weight the exposure of ∗ corresponding author. tel.: + . e-mail addresses: @fep.up.pt, jsc@inescporto.pt (m.r. sousa), gama@fep.up.pt (j. gama), ebrandao@fep.up.pt (e. brandão). other names can be used to refer to pd models, namely: credit scoring, credit risk odels, scorecards, credit scorecards, rating systems, or rating models, although some ave different meanings. c e s t r ttp://dx.doi.org/ . /j.eswa. . . - /© elsevier ltd. all rights reserved. redit at risk, the risk weighted assets, and benefit from lower capital equirements. a lot of improvements have been made in the exist- ng rating frameworks, extending the use of data mining tools and rtificial intelligence. yet, this may have been bounded by a certain nwillingness to accept less intuitive algorithms or models going be- ond standard solutions being implemented in the banking industry, ettled in-house or delivered through analytics providers. developing and implementing a credit scoring model can be time nd resource consuming, easily ranging from to months, from ata extraction until deployment. hence, it is not rare that banks use nchanged credit scoring models for several years. bearing in mind hat models are built using a sample file frequently comprising or ore years of historical data, in the best case scenario, data used n the models are shifted years away from the point they will be sed. should conditions remain unchanged, then this would not sig- ificantly affect the accuracy of the models, otherwise, their perfor- ance can greatly deteriorate over time. the recent financial crisis onfirmed that financial environment greatly fluctuates, in an un- xpected manner, posing renewed attention regarding models built pon time-frames that are by far outdated. by – , many fi- ancial institutions were using stale credit scoring models built with istorical data of the early-decade. the degradation of stationary redit scoring models is an issue with empirical evidence in the lit- rature (avery, calem, & canner, ; crook, thomas, & hamilton, ; lucas, ; sousa, gama, & gonçalves, b), however re- earch is still lacking more realistic solutions. dominant approaches rely on static learning models. however, as he economic conditions evolve in the economic cycle, either deterio- ating or improving, also varies the behavior of an individual, and his http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi= . /j.eswa. . . &domain=pdf mailto: @fep.up.pt mailto:jsc@inescporto.pt mailto:jgama@fep.up.pt mailto:ebrandao@fep.up.pt http://dx.doi.org/ . /j.eswa. . . m.r. sousa et al. / expert systems with applications ( ) – p h s i i c o c a a i c a u s ( p i s c d f c m t t t l e o t t l t s b a p c a l v t s o e a c c t d w t e c s c i m i m a b ability to repay his debt. furthermore, the default evolution echoes trends of the business cycle, and related with this, regulatory move- ments, and interest rates fluctuations. in good times, banks and bor- rowers tend to be overoptimistic about the future, whilst in times of recession banks are swamped with defaulted loans, high provisions, and tighten capital buffers turn highly conservative. the former leads to more liberal credit policies and lower credit standards, the later promotes sudden credit-cuts. hence, default needs to be regarded as time changing. traditional systems that are one-shot, fixed memory-based, trained from fixed training sets, and static settings are not prepared to process the evolving data. and so, they are not able to continuously maintain an output model consistent with the actual state of environ- ment, or to quickly react to changes (gama, ). these are some of the features of classic approaches that evidence the constraints of the existing credit scoring systems. as the processes underlying credit risk are not strictly stationary, consumers’ behavior and default can change over time in unpredictable ways. a few limitations to the ex- isting approaches, idealized in the classical supervised classification paradigm, can be traced in published literature: • the static models usually fail to adapt when the population changes. static and predefined sample settings often lead to an incomplete examination of the dynamics influencing the prob- lem (gama, ; hand, ). • certain assumptions that are implicit to the methods, often fail in real-world environments (yang, ). these assumptions relate to: – representativeness - the standard credit scoring models rely on supervised classification methods that run on -years-old static samples, in order to determine which individuals are likely to default in a future fixed period, year for pd mod- els (thomas, ; thomas, edelman, & crook, ). such samples are supposed to be representative of the potential borrowers consumers of the future, the through-the-door pop- ulation. they should also be sufficiently diverse to reflect dif- ferent types of repayment behavior. however, a wide range of research is conducted in samples that are not representative. – stability and non-bias - the distribution from which the design points and the new points is the same; classes are perfectly de- fined, and definitions will not change. not infrequently there are selective biases over time. simple examples of this occur- rence can be observed when a bank launches a new product or promotes a brand new segment of customers. it can also occur when macroeconomics shifts abruptly from an expansion to a recession phase, or vice versa. – misclassification costs - these methods assume that the costs of misclassification are accurately known, but in practice they are not. • the methods that are most widely used in the banking industry, logistic regression and discriminant analysis are associated with some instability with high-dimensional data and small sample size. other limitations regard to intensive variable selection effort and incapability of efficiently handling non-linear features (yang, ). • static models are usually focused in assessing the specific risk of applicants and obligors. however, a complete picture can only be achieved by looking at the return alongside risk, which requires the use of dynamic rather than static models (bellotti & crook, ). there is a new emphasis on running predictive models with the ability of sensing themselves and learn adaptively (gama, ). ad- vances on the concepts for knowledge discovery from data streams suggest alternative perspectives to identify, understand and effi- ciently manage dynamics of behavior in consumer credit in chang- ing ubiquitous environments. in a world where the events are not reordained and little is certain, what we do in the present affects ow events unfold in unexpected ways. so far, no comprehensive et of research to deal with time changing default had much impact nto practice. in credit risk assessment, a great deal of sophistication s needed to introduce economic factors and market conditions into urrent risk-assessment systems (thomas, ). the study presented in this paper is a large extension of a previ- us research that delivered the winning model within the brics ompetition in data mining and finance (sousa, gama, brandão et al., a; sousa et al., b). this competition opened to academics nd practitioners, was focused on the development of a credit risk ssessment model, tilting between the robustness of a static model- ng sample and the performance degradation over time, potentially aused by market gradual changes along few years of business oper- tion. participants were encouraged to use any modeling technique, nder a temporal degradation or concept drift perspective. in the re- earch attached to the winning model, sousa, gama, and gonçalves b) have proposed a two-stage model for dealing with the tem- oral degradation of credit scoring models, which produced motivat- ng results in a -year horizon. the winners first developed a credit coring method using a set of supervised learning methods, and then alibrated the output, based on a projection of the evolution in the efault. this adjustment considered both the evolution of the de- ault and the evolution of macroeconomic factors, echoing potential hanges in the population of the model, in the economy, or in the arket. in so doing, resulting adjusted scores translated a combina- ion of the customers’ specific risk with systemic risk. the winning eam (sousa, gama, & gonçalves) concluded that the performance of he models did not significantly differ among classification models, ike logistic regression (lr), adaboost, and generalized additive mod- ls (gam). however, after training in several windows lengths, they bserved that the model based on the longest window has produced he best performing model over the long-run, among all competitors. his finding allowed to realize that some specifics of the credit portfo- ios and macroeconomic environments may reveal quite stable along ime. for those cases, a model built with a static learning setting may eem appropriate, if tested during stable phases. the question yet to e answered was in which conditions credit risk models degrade? nd when so, if there is any alternative modeling technique to the revailing credit scoring models? the aim of this study is to find a learer understanding on which type of modeling framework allows rapid adaptation to changes, and in which circumstances a static earning setting still delivers well-performing models. with this in iew, we implemented a dynamical modeling framework and two ypes of windows for model training, which enable testing our re- earch questions: (a) in which conditions can a dynamic modeling utperform a static model?; (b) is the recent information more rel- vant to improve forecasting accuracy?; (c) does older information lways improve forecasting accuracy? this paper introduces a new dynamic modeling framework for redit risk assessment, imported from the emerging techniques of oncept drift adaptation, in streaming data mining and artificial in- elligence. the proposed model is able to produce more robust pre- ictions in stable conditions, but also in the presence of changes, hile the prevailing methods cannot. this is a promissory tool both o academics and practitioners, because unlike the traditional mod- ls, it has the ability of adjusting the predictions in the presence of hanges, like inversions in the economic cycles, major crisis, or intrin- ic behavioral circumstances (e.g. divorce, unemployment and finan- ial distress). besides the goal of enhancing the prediction of default n credit, the new modeling framework also enables developing a ore comprehensive understanding of the evolution of the credit rat- ng systems over time and anticipating unexpected events. further- ore, we study the implications to credit risk assessment of keeping long-term memory, and forgetting older examples, which have not een done so far. m.r. sousa et al. / expert systems with applications ( ) – i s d s f u a fi b c t g m t s f u i a m w c i m p d c fi l m a t e l y t a u u m u i a a k v a a p f v i e t t d c t o t d t l o s t m m i i b p a m c s m a s g w c o j o w o a a r p e d t a c b w t m c o m f i c b f t t few authors have explicitly tried a dynamic modeling framework n credit risk assessment, or connected concepts. based on a national ample of a credit reporting agency, avery et al. ( ) show that tra- itional modeling often fails to consider situational circumstances, uch as local economic conditions and individual trigger events, af- ecting the ability of scoring systems to accurately quantify individ- als’ credit risk. we can trace the few existing contributions in this rena over the most recent years. sun and li ( ) formally de- ne financial distress concept drift and build a dynamic modeling ased on instance selection. saberi et al. ( ) worked on the con- ept of granularity for selecting the optimum size of the testing and raining groups with a sample of credit cards of a bank operating in erman. pavlidis, tasoulis, adams, and hand ( ) proposed a ethodology for the classification of credit applications with the po- ential of adapting to population drifts. this paper follows in section with a brief description of the main ettings and concepts of the supervised learning problem and score ormulation. it also presents an overview of the methods typically sed in supervised learning, and specifically in credit score modeling. n section , we introduce the topic of concept drift in credit default nd some adaptation methods that can be promising for dynamic odeling credit risk. in section we present a case study, where e employ a set of these adaptation methods to a real-world finan- ial dataset. first, we characterize the database and provide some ntuition on the background of the problem. then, we explain the ethodology of this research. section provides the fundamental ex- erimental results. conclusions and future applications of the new ynamic modeling framework are traced in section . . settings and concepts in this work we import some of the emerging techniques in con- ept drift adaptation into credit risk assessment models. this is a eld of research that has been receiving much attention in machine earning over the last decade, as an answer for suitably shaping the odels and processes to a reality that is ever-changing over contexts nd time. the settings and definitions adopted in this paper replicate he general nomenclature surveyed by gama, Žliobaitė, bifet, pech- nizkiy, and bouchachia ( ). . . supervised learning problem credit risk assessment can be addressed as a classification prob- em, a subset of supervised learning. the aim is to predict the default ∈ {good, bad}, given a set of input characteristics x. the term at- ribute refers to each of the possible values that a characteristic can ssume; the term bin denotes a set of attributes or an interval of val- es in a continuous characteristics; the term example, or record, is sed to refer to one pair of (x, y). supervised learning classification ethods try to determine a function that best separates the individ- als in each of the classes, good and bad, in the space of the problem. the model building is carried on a set of training examples - train- ng set – collected from the past history of credit, for which both x nd y are known. the best separation function can be achieved with classification method. these methods include, among others, well- nown classification algorithms such as decision trees (dt), support ector machines (svm), artificial neural networks (ann), and gener- lized additive models (gam). hands-on software packages are avail- ble to the user for example in r, sas, matlab, and model builder for redictive analytics. in credit scoring models, the accuracy of such unctions is typically assessed in separate sets of known examples – alidation or out-of-sample data sets. the idea behind this procedure s to mimic the accuracy of that function in future predictions of new xamples where x is known, but y is not. according to the bayesian decision theory (duda, hart, & stork, ), a classification can be described by the prior probabilities of he classes p(y) and the class conditional probability density func- ion p(x|y) for the two classes, good (g) and bad (b). the classification ecision is made according to the posterior probabilities of the two lasses, which for class b can be represented as: p(b|x) = p(x|b)p(b)/p(x) ( ) where p(x) = p(b)p(x|b) + p(g)p(x|g). here, it is assumed that he costs for misclassifying a bad customer are the same as for the pposite situation, the equal costs assumption. it is worth recalling hat, in real-world financial environments, the costs of failing the pre- iction in a real bad are by far superior to failing in a real good. in he first case, there is essentially a loss of the exposure at default, oss given default (lgd), possibly mitigated with collateral. the sec- nd case affects the business, as it translates into a loss of margin. ousa and da costa ( ) show several possibilities to overcome his practical issue, by adapting the output of standard classification ethods under the equal costs assumption to imbalanced costs for isclassification, associated with the decision and prediction tasks. it s worth discussing the related issue of class imbalance in credit scor- ng datasets. quite often, these datasets contain a much smaller num- er of observations in the class of defaulters than in that of the good ayers (brown & mues, ; marqués, garcía, & sánchez, a). large class imbalance is therefore present which some techniques ay not be able to successfully handle. baseline methods to handle lass imbalance include oversampling the minority class or under ampling the majority class; tomek links is an example of the for- er, smote of the latter (chawla, bowyer, hall, & kegelmeyer, ). nother established approach to correct imbalance adopt a cost sen- itive classifier with the misclassification cost of the minority class reater than that of the majority class. within this approach, it is orth mentioning metacost, a general method for making classifiers ost-sensitive (domingos, ). all these methodologies, implicitly r explicitly, optimize the decision process for a specific business ob- ective. in other words, the optimization is made for a specific trade- ff between the error committed in identifying someone as defaulter hen one is in fact a non-defaulter individual and the opposite type f error of diagnosing someone as non-defaulter when one is in fact defaulter. this individualization is unconnected with our study and ny of these methods can be incorporated in the methodology under esearch. . . score formulation a credit scoring model is a simplification of the reality. the out- ut is a prediction of a given entity, actual or potential borrower, ntering in default in a given future period. having decided on the efault concept, conventionally a borrower being in arrears for more han days in the following months, those matching the criteria re considered bad and the others are good. other approaches may onsider a third status, the indeterminate, between the good and the ad classes, e.g. to days overdue, for which it may be unclear hether the borrower should be assigned to one class or to the other. his status is usually removed from the modeling sample, despite the odel can be used to score them. for simplicity, in this paper we will onsider the problem of two classes, although the proposed method- logy can easily be adapted to the other case. the output is a function of the input characteristics x, which is ost commonly referred as score, s(x). we also consider that this unction has a monotonic decreasing relationship with the probabil- ty of entering in default (i.e. reaching the bad status). a robust score- ard enables an appropriate differentiation between the good and the ad classes. it is achieved by capturing an adequate set of information or predicting the probability of the default concept (i.e. belonging to he bad class), based on previous known default occurrences. the no- ation of such probability, pr{bad|score based on x}, is: p(b|s(x)) = p(b|s(x), x) = p(b|x), ∀x ∈ x ( ) m.r. sousa et al. / expert systems with applications ( ) – & g c w a a o u v i a l c t t a p s t e m p i b a a d e m c s f w e c t m t t u i n i ( t v p u t t o h since p(g|x) + p(b|x) = , it naturally follows the probability of the complementary class: p(g|s(x)) = p(g|x) = − p(b|x), ∀x ∈ x ( ) among researchers and real-world applications, a usual written form of the score is the log odds score: s(x) = ln p(g|x) p(b|x) , and p(g|x) + p(b|x) = . ( ) in so saying, the score may vary from −∞, when p(g|x) = , to +∞, when p(g|x) = , i.e. s(x) ∈ r. the probability of the default event can be written in terms of the score: p(b|x) = /( + es(x)), ∀x ∈ x the most conventional way to produce log odds score is based in the logistic regression. however, other classification algorithms can also be used, adjusting the output to the scale of that function. in so saying, we assume that independently of the method used to de- termine the best separation between the two classes, good and bad, and then the resulting scorecard has the same property of the log odds score. although a grounded mathematical treatment may be tempting to tackle this problem, it goes beyond the scope of this work. notwithstanding, we provide some intuitions on the techni- cal material to survey. the basics of credit scoring and the most common approaches to build a scorecard, are further detailed in the operational research literature (anderson, ; crook, edelman, & thomas, ; mcnab & wynn, ; thomas, ; thomas et al., ). recent advances in the area also deliver methods to build risk based pricing models (thomas, ) and methodologies towards the optimization of the profitability to the lenders (einav, jenkins, & levin, ). . . supervised classification methods the first approach to differentiate between groups took place in fisher’s original work in ( ) for general classification prob- lems of varieties of plants. the objective was to find the best sep- aration between two groups, searching for the best combination of variables such that the groups were separated the most in the sub- space. durand ( ) brought this methodology to finance for distin- guishing between good and bad consumer loans. discriminant analysis was the first method used to develop credit scoring systems. altman ( ) introduced it in the prediction of cor- porate bankruptcy. first applications in retail banking were mainly focused on credit granting in two categories of loans: consumer loans, and commercial loans (for an early review and critique on the use of discriminant analysis in credit scoring see eisenbeis ( )). the boom of credit cards demanded the automation of the credit decision task and the use of better credit scoring systems, which were doable due to the growth of computing power. the value of credit scoring became noticed and it was recognized as a much better predictor than any other judgmental scheme. logistic regression (steenackers & goovaerts, ) and linear programming (see chen, zhong, liao, and li, for a review) were introduced in credit scoring, and they turned out to be the most used in financial industry (anderson, ; crook et al., ). the use of artificial intelligence tech- niques imported from statistical learning theory, such as classifi- cation trees (breiman, friedman, olshen, & stone, ; quinlan, ) and neural networks (desai, crook, & overstreet jr, ; jensen, ; malhotra & malhotra, ; west, ) have arisen in credit scoring systems. support vector machine (svm) is another method based in optimization and statistical learning, that received increased attention over the last decade in research in finance, ei- ther to build credit scoring systems for consumer finance or to predict bankruptcy (li, shiue, & huang, ; min & lee, ; wang, wang, & lai, ). genetic algorithms (chen & huang, ; ong, huang, tzeng, ), colony optimization (martens et al., ), and re- ression and multivariate adaptive regression splines (lee & chen, ) have also been tried. evolutionary computing (marqués, gar- ía, & sánchez, ), including genetic algorithms (chen & huang, ; ong et al., ) and colony optimization (martens et al., ), as also considered for credit scoring. regression (lee & chen, ) nd clustering (wei, yun-zhong, & ming-shu, ) techniques have lso been tailored to the problem. the choice of a learning algorithm is a difficult problem and it is ften based on which happen to be available, or best known to the ser (jain, duin, & mao, ). the number of learning algorithms is ast. many frameworks, adaptations to real-life problems, intertwin- ng of base algorithms were, and continue to be, proposed in the liter- ture, ranging from statistical approaches to state-of-the-art machine earning algorithms, from parametric models to non-parametric pro- edures (abdou & pointon, ; baesens et al., ). as an alterna- ive to using a single method, a trend that is still evolving relates to he use of hybrid systems (hsieh, ; lee, chiu, lu, & chen, ), nd ensemble of classifiers with which the outputs are achieved by a redefined sequence or rule, or a voting scheme (marqués, garcía, & ánchez, b; wang, hao, ma, & jiang, ). new concepts for adapting to changes (adams, tasoulis, anagnos- opoulos, & hand, ; pavlidis et al., ; sousa et al., b; yang, ) and modeling the dynamics (crook & bellotti, ; saberi t al., ) in populations start being exploited in credit risk assess- ent. . dynamic modeling for credit default . . concept drift in credit default credit default is mostly a consequence of financial distress. a erson, or a company, is in financial distress when is experiencing ndividual financial constraints or is being exposed to external distur- ances. in private individuals, financial constraints may result from brupt or intrinsic circumstances. in the first case, distress is usu- lly an outcome of sorrowful events like unemployment, pay cuts, ivorce, and disease. the second is most commonly related to over- xposure, low assets, erratic behavior, or bad management perfor- ance. in this paper we tackle the phenomenon of concept drift in redit default, which we now briefly explain. in the existing literature, concept drift is generally used to de- cribe changes in the target concept, which are activated by trans- ormations in the hidden context (schlimmer & granger jr, ; idmer & kubat, ) in dynamically changing and non-stationary nvironments. as a result of these transformations, the target concept an shift suddenly or just cause a change in the underlying data dis- ribution to the model. this means that with time, optimal features ay drift significantly from their original configuration or simply lose heir ability to explain the target concept. for example, a reduction of he minimum ltv (loan to value), tighten the space of possible val- es, which is noticed with a change in the distribution, and eventually n the credit default concept. when such drifts happen, the robust- ess of the model may significantly decrease, and in some situations t may no longer be acceptable. some authors distinguish real concept drift from virtual drift gama et al., ; sun & li, ; tsymbal, ). the former refers o changes in the conditional distribution of the output (i.e., target ariable) given the input features, while the distribution of the in- ut may remain unchanged. the later refers to gradual changes in the nderlying data distribution with new sample data flowing, whereas he target concept does not change (sun & li, ). real concept drift refers to changes in p(y|x), and it happens when he target concept of credit default evolves in time. such changes can ccur either with or without a change in p(x). this type of drift may appen directly as a result of new rules for defining the target classes, m.r. sousa et al. / expert systems with applications ( ) – g d g r p d d o w i i fi s a s m d t p m i ( i i l v a m i t w c t p t l d o c b t w t a d m t t s m t s i c f m g m d r t w d f p d fi i c t s b p s d ( b i i t fi t t b b d p m s c t s t p a c i i o a c i b u b t t t g ood or bad, as those settled by regulators, when new criteria for efault are demanded to the banks. examples of these include the uidelines for the minimum number of days past due or in the mate- iality threshold for the amount of credit in arrears, issued with the revious basel ii accord. another understanding of the real concept rift in credit default is associated with indirect changes in the hid- en context. in this case, credit default changes when evolving from ne stage of delinquency to another. for example, most of the people ith credit until five days past due tend to pay before the following nstallment, as most of them are just delayers. yet, the part of debtors n arrears that also fail the next installment are most likely to be in nancial distress, possibly as a result of an abrupt or intrinsic circum- tance, and therefore they require more care from the bank. when rrears exceed three installments, the debtor is most certainly with erious financial constraints, and is likely to fail his credit obligations. ore extreme delays commonly translate into hard stages of credit efault, which require intensive tracking labor or legal actions. virtual drifts happen when there are changes in the distribu- ion of the new sample data flowing without affecting the posterior robability of the target classes, p(y|x). with time, virtual drifts may ove to real concept drifts. other interpretations can also be found n literature, for describing an incomplete representation of the data widmer & kubat, ), and changes in the data distribution lead- ng to changes in the decision boundary (tsymbal, ). accord- ng to some authors, other events can also be seen as virtual drifts, ike sampling shift (salganicoff, ), temporary drifts (lazarescu, enkatesh, & bui, ), and feature change (salganicoff, ). as n example of virtual drift, we might consider the credit decision- aking along the recent financial crisis. the lenders had to anticipate f a borrower would enter in default in the future (i.e. being bad). al- hough the macroeconomic factors have worsened, employed people ith lower debt to income remained good for the lenders, and so they ontinued to have access to credit. although we are mostly interested to track and detect changes in he real target concept, p(y|x), the methodology introduced in this aper attempts to cover both real concept and virtual drifts applied o the default concept drift detection and model rebuilding. . . methods for adaptation traditional methods for building a scorecard consider a static earning setting. in so doing, this task is based in learning in a pre- efined sample of past examples and then used to predict an actual r a potential borrower, in the future. this is an offline learning pro- edure, because the whole training data set must be available when uilding the model. the model can only be used for predicting, af- er the training is completed, and then it is not re-trained alongside ith its utilization. in other words, when the best separation func- ion is achieved for a set of examples of the past, it is not updated for while, possibly for years, independently of the changes in the hid- en context or in the surrounding environment. new perspectives on odel building arise together with the possibility of learning online. he driving idea is to process new incoming data sequentially, so that he model may be continuously updated. one of the most intuitive ideas for handling concept drift by in- tance selection is to keep rebuilding the model from a window that oves over the latest batches and use the learn model for predic- ion on the immediate future. this idea assumes that the latest in- tances are the most relevant for prediction and that they contain the nformation of the current concept (klinkenberg, ). a framework onnected with this idea consists in collecting the new incoming data or sequential batches in predefined time intervals, e.g. year by year, onth by month, or every day. the accumulation of these batches enerates a panel data flow for dynamic modeling. in finance, it remains unclear whether it is best having a long emory or forgetting old events. if on the one hand, a long memory is esirable because it allows recalling a wide range of different occur- ences, in the other, many of those occurrences may no longer adjust o the present situation. a rapid adaptation to changes is achieved ith a short window, because it reflects the current distribution of efault more accurately. however, for the contrary reason, the per- ormance of models built upon shorter windows worsens in stable eriods. in credit risk assessment modeling, this matter has been in- irectly discussed by practitioners and researchers when trying to gure the pros and cons of using a through-the-cycle (ttc) or point- n-time (pit) schema to calibrate the output of the scorecards to the urrent phase of the economic cycle. for years, a pit schema was he only option, because banks did not have sufficient historical data eries. since the implementation of the basel ii accord worldwide, anks are required to store the data of default for a minimum -years eriod and consider a minimum of -years period for calibrating the corecards. an original idea of widmer and kubat ( ) uses a sliding win- ow of fixed length with a data processing structure first-in-first-out fifo). each window may consist of a single or multiple sequential atches, instead of single instances. at each new time step, the model s updated following two processes. in the first process, the model s rebuilt based on the training data set of the most recent window. hen, a forgetting process discards the data that move out of the xed-length window. incremental algorithms (widmer & kubat, ) are a less ex- reme hybrid approach that allows updating the prediction of models o the new contexts. they are able to process examples batch-by- atch, or one-by-one, and update the prediction model after each atch, or after each example. incremental models may rely on ran- om previous examples or in representative selected sets of exam- les, called incremental algorithms with partial memory (maloof & ichalski, ). the challenge is to select an appropriate window ize. . case study this research evolves from a one-dimensional analysis, where we ome across the financial outlook underlying the problem, to a mul- idimensional analysis along several points in time. the former, de- cribed in sections . , . , and . , is tailored to gain intuition on he default predictors and the main factors ruling the context of the roblem. the latter, in section . , is designed to gradually develop nd test a new dynamic framework to model credit risk. . . dataset and validation environment the research summarized here was conducted in a real-life finan- ial dataset, comprising , records, from a financial institution n brazil along two years of operation, from to . each entity n the modeling dataset is assigned to a delinquency outcome - good r bad. in this problem, a person is assigned to the bad class if she had payment in delay for or more days, along the first year after the redit has been granted. the delinquency rate in the modeling dataset s . %, which is in line with the high default rates in credit cards in razil, one of the countries with the highest default rates in the prod- ct. the full list of variables in the original data set is available in the rics official website. it contains variables, categorized in able , and one target variable with values identifying a record in he bad class and for the good class. . . data analysis and cleansing some important aspects of the datasets were considered, because hey can influence the performance of the models. these aspects re- ard to: m.r. sousa et al. / expert systems with applications ( ) – table predictive variables summary. type # information numerical age, monthly income, time at current address, time at current employer, number of dependents, and number of accounts in the bank. treated as nominal credit card bills due day, st to th zip digit codes, home (state, city, and neighborhood), marital status, income proof type, long distance dialing code, occupation code, and type of home. binary address type proof, information of the mother and fathers names, input from credit bureau, phone number, bills at the home address, previous credit experience, other credit cards, tax payer and national id, messaging phone number, immediate purchase, overdraft protection agreement, lives and work in the same state, lives and work in the same city, and gender. date application date. id customer, personal reference, and branch unique identifiers. % % % % % % , , , , , , c u m u la ti ve f re q u en cy income (r$) fig. . cumulative frequency of the monthly income for and . table information values for the tested combi- nations. combination iv age × income . age × occupation . income × marital status . income × occupation . income × proof of income . age × income × occupation . c t w n c i s t g u g c b r a t w u t t w i d a i n a w • significant percent of zero or missing values in exception to the vari- ables ‘lives and work in the same state’ and ‘previous credit expe- rience’, binary flags have % to % concentrated in one of the values, which turn them practically unworkable. the same occurs for the numerical variables number of dependents and number of accounts in the bank, both with more than % zeroes. the re- maining variables were reasonably or completely populated. • outliers and unreasonable values the variable age presents . % of applications assigned to customers with ages between and years. a small percent of values out of the standard ranges are observable in the variables credit card bills due day, monthly in- come and time at current employer. unreasonable values are de- tected in the first semester of , suggesting that the data were subjected to corrections from the second semester of on- wards. • unreliable and informal information little reliability on socio- demographic data is amplified by specific conditions in the back- ground of this problem. this type of scorecards is usually based in verbal information that the customer provides, and in most of the cases no certification is made available. in % of the applications, no certification for the income was provided, and % do not have proof for the address type. customers have little or no concern to provide accurate information. the financial industry is aware of this kind of limitations. however, in highly competitive environ- ments there is little chance to amend them, while keeping in the business. hence, other than regulatory imperatives, no player is able to efficiently overcome this kind of data limitations. as cur- rently there are no such imperatives in brazilian financial market, databases attached to this type of models are likely to keep lack- ing reliability in the near future. • bias on the distributions of modeling examples the most noticeable bias is in the variable monthly income, where values shift from one year to another, exhibited in fig. . this is most likely related to increases in the minimum wages and inflation. slight variations are also observable in the geographical variables, which are possibly related with the geographical expansion of the institution. in the remaining characteristics, the correlation between the frequency distributions of and range from to %, suggesting a very stable pattern during the analyzed period. . . data transformation and new characteristics . . . data cleansing and new characteristics we focused the data treatment on the characteristics that were reasonably or fully populated. fields state, city, and neighborhood ontain free text, and were subjected to a manual cleansing. at- ributes with or less records were assigned to a new class “other”. e could observe that there may be neighborhoods with the same ame in different cities; and hence we concatenated these new leansed fields, state and city, into the same characteristic. . . . data transformation variables were transformed using the weights of evidence (woe) n the complete modeling dataset, which is a typical measure in credit core modeling (fico, ). w oe = ln g/g b/b , where g and b are respec- ively the number of good and the number of bad in the attribute, and and b are respectively the total number of good and bad in the pop- lation sample. the larger the woe the higher is the proportion of ood customers in the bin. for the nominal and binary variables we alculated the woe for each class. numerical variables were firstly inned using sas enterprise miner, and then manually adjusted to eflect domain knowledge. in so doing we aim to achieve a set of char- cteristics less exposed to overfitting. cases where the calculation of he woe rendered impossible - one of the classes without examples – ere given an average value. the same principle was applied to val- es out of the expected ranges (e.g. credit card bills due day higher han ). . . . one-dimensional analysis the strength of each potential characteristic was measured using he information value (iv) in the period, iv = ∑ni= (g/g − b/b)w oe, here n is the number of bins in the characteristic. the higher is the v, the higher is the relative importance of the characteristic. in a one- imensional basis, for the entire period, the most important char- cteristics are age, occupation, time at current employer, monthly ncome and marital status, with information values of . , . , . , . , and . , respectively. remaining characteristics have . or less. . . . interaction terms using the odds in each attribute of the variables, we calculated ew nonlinear characteristics using interaction terms between vari- bles to model the joint effects. we tested six combinations, for which e present the information value in table . m.r. sousa et al. / expert systems with applications ( ) – d o r c e i i d c s m w t u s c b a t t s l b t p e i t o f b p t m a s o s t a w r t i s t f m m c d u a t t m s s i t a t r s t t w t t a s v l f q a y t n w s s m w f o t s t i d i t t s t w . . . time series descriptive analysis fig. a shows the real concept drift along – . the highest efault rates are noticed in the first quarter of , and at the end f . fig. b displays the evolution of the business in the same pe- iod. it exhibits two features of the business. first, we can see that the redit cards business follow an annual seasonality, increasing along ach year. second, the credit cards business is rising over time, which s related with the expansion of the branch network of the financial nstitution. the decrease of default rate during suggests that the ecision-making process might have been slightly enhanced, when omparing to the beginning of the period. . . dynamic modeling framework the dynamic modeling framework presented in this research con- iders that data is processed batch-by-batch. sequentially, at each onthly window, a new model is learned from a previous selected indow, including the most recent month. to mimic the time evolu- ion, we assumed that the current month gradually shifts from ntil the third quarter of . each learning unit for the model building was grounded on a static etting. the training of each unit consists of a supervised classifi- ation procedure, executed in three steps. first, characteristics are inned. second, the classification model is designed with generalized dditive models (gam) and a fold crossed-validation, upholding he classification algorithm used to develop the winning model in he brics in data mining and finance (brics-cci&cbic, ; ousa et al., b). concurrently, the best set of characteristics is se- ected until no other characteristic in the training dataset adds contri- ution to the information value (iv) of the model. in this application he threshold was set for a minimum increment of . . third, the erformance of the model is measured based on the gini coefficient, quivalent to consider the area under the roc curve (auc), which s a typical evaluation criteria among researchers and in the indus- ry (řezáč & řezáč, ). this coefficient refers to the global quality f the credit scoring model, and ranges between − and . the per- ect scoring model fully distinguishes the two target classes, good and ad, and has a gini coefficient equal to . a model with a random out- ut has a gini coefficient equal to zero. if the coefficient is negative, hen the scores have a reverse meaning. the extreme case − would ean that all examples of the good class are being predicted as bad, nd vice versa. in this case, the perfect model can be achieved just by witching the prediction. at each month, instances for modeling are selected from all previ- us available batches, according to a selection mechanism. we use in- tance selection methods to test the hypothesis under investigation. wo methods for tackling default concept drift were implemented - full memory time window, and a fixed short memory time window ith a forgetting mechanism. the full memory time window assumes that the learning algo- ithm generates the model based on all previous instances (fig. a). he process is incremental, so every time a new instance arises, it s added to the training set, and a new model is build. this schema hould be appropriate to detect mild concept drifts, but it is unable o rapidly adapt to major changes. models of this schema should per- orm suitably in stable environments. a shortcoming of this incre- ental schema is that the training dataset quickly expands which ay requires a huge storage capacity, and constrain the use of some lassification algorithms, to be able of processing the expanding ataset. in the fixed short memory time window, the model development ses the most recent window. with this schema, illustrated in fig. b, new model is build in each new batch, by forgetting past examples. he fundamental assumption is that past examples have low correla- ion with the current default concept. under this setting, the dynamic odeling should quickly adapt to changes. the most extreme case of hort memory time window is when only the current example is con- idered to train the new model, which represents to the online learn- ng without any memory of the past. a deficiency of this method is hat it often lacks of generalization ability in stable conditions that is mplified with extremely short windows. these modeling frameworks enable comparing these configura- ions between themselves, and also compare them with the model eached with a static learning setting. the research questions of this tudy should be answered following the reasoning: • if the full memory time window outperforms the other schema, then more recent data are not fundamental for the prediction; the environment of the decision-making should be in a stable phase. otherwise, the default concept is drifting, and so the most recent data are more relevant for the prediction. • if a model built with static learning in the first window of the period has the best performance, then older data can improve the prediction. this may happen, for example, when a new credit product is launched, and the credit decision-making criteria are adjusted afterwards. in such case, the oldest data are more rep- resentative, as they can illustrate a more diverse range of risk be- haviors. otherwise, over the long-run, dynamic modeling should outperform the model learnt with static setting. . experimental results we assessed the performance of the sequential models built with he dynamic modeling framework introduced in the previous sec- ion, through the period and . the experimental design as drawn for assessing the performance in the modeling period, in he short-term, and in the farthest-term. in each model rebuilding, he performance in the modeling period was assessed in the test set. dditionally, using two out-of-sample windows, we measured the hort-term performance of the model in the month following the de- elopment, and the farthest-term performance was measured in the ast quarter of . although we have considered monthly windows or developing the model, for the long-run assessment we chose a uarterly window instead of a single month. in so doing, potential typical properties of the decision-making process at the end of the ear were smoothed. in this section, we provide further evidence on temporal degrada- ion of static credit scoring. then, we challenge the robustness of the ew concept of dynamic modeling against a static model, developed ith a traditional framework. we finally present and discuss the re- ults for the two sliding-window configurations - full memory and hort memory. . . temporal degradation of static credit scoring the temporal degradation of the credit scoring is detected when easuring the performance of each model in the sequence generated ith the dynamic modeling. fig. a and b exhibit the gini coefficient or each model, measured in the modeling test set and two different ut-of-sample windows, one month after rebuild the model, and in he farthest quarter in the period ( q ). fig. a shows the performance along the entire period with the hort memory configuration. one month after rebuilding the model, he performance curve is always below the performance measured n the modeling period, showing that the performance consistently ecreases one month after rebuilding the model. when evaluat- ng the performance with the full memory configuration, in fig. b, he extent of degradation within a month is not consistent over he period. during the first semester of , performance mea- ured in the month after rebuilding the model is slightly superior to he one measured in the modeling period, and from that point on- ards, it is marginally inferior. this may suggest that the short-term m.r. sousa et al. / expert systems with applications ( ) – default average default % % % % % % -jan -apr -jul -oct -jan -apr -jul -oct (% ) r a te d e fa u lt application month (a) default rate. , , , , -jan -apr -jul -oct -jan -apr -jul -oct (# ) c a r d s c r e d it u n d e r w r it e d month (b) number of new credit cards contracts. fig. . default rate and new contracts in the period – . … learning learning learning learning n timemonth month month month n … (a) full memory time window. … learning learning learning learning n timemonth month month month n … (b) fixed short memory time window. fig. . configurations for tackling concept drift in credit default. modeling period month after rebuilding farthest quarter ( q ) . . . . . . . -jan -mar -jun -sep -dec -mar -jun - sep c o e ff ic ie n t g in i model rebuilding month (a) short memory. modeling period month after rebuilding farthest quarter ( q ) . . . . . . . -jan -mar -jun -sep -dec -mar -jun - sep c o e ff ic ie n t g in i model rebuilding month (b) full memory. fig. . gini coefficient of the sequence of models produced with the dynamical modeling. e t t c i performance is more similar to the performance in the modeling pe- riod when using the full memory configuration. the extent of degradation is higher, when the performance of the model is measured at the end of the period ( q ). the farthest is the point of the prediction from the point of the development; the highest is the extent of degradation of the performance. these ffects are consistently perceived on the two windowing configura- ions - short memory (fig. a) and full memory (fig. b). considering the real performance of the models one month after hey were built, the average degradation of the models sequentially onstructed, shown in table , is . in the short memory and . n the full memory configuration. in the farthest quarter in the period m.r. sousa et al. / expert systems with applications ( ) – table average degradation of the sequence of models produced with the dynamic modeling. memory type gini index degradation modeling month after farthest quarter month after farthest quarter period rebuilding ( q ) rebuilding ( q ) short , , , − , − , full , , , − , − , ( u q t p s n m s i c q p h o p - c t m m m t h i h t m t o i n u s w o t t h f m t t t c t t d h s b i d a u c r d a b e d d a d s w m t p t p n i t o t l l m d c o ( w c e i i n t e q ) the degradation reaches . in the short memory config- ration and . in the full memory schema. although degradation can be observed in all models of the se- uence, updating the model always yields the best discrimination be- ween the target classes - good and bads. . . dynamic versus static the proposed dynamic modeling framework enables a major im- rovement of the initial static model, which was trained with the ample in the first month of ( m ). fig. a shows the immediate performance achieved with the dy- amic modeling – full and short term memory – versus the static odel, measured in the month following the development. fig. b hows the performance of the models in each point in time, measured n the farthest quarter of the period. consistently, for both memory onfigurations and performance criteria, immediate or in the farthest uarter, either the static or the dynamic modeling performances im- rove until the third quarter of , which might reflect the en- ancement of the set of characteristics x that was partially corrected ver that period. in fig. a, we observe a certain overlap between the immediate erformance achieved with the two types of memory configurations short and full. for all the periods, the short-term performance in- reases until the third quarter of , and slightly decreases from hat point onwards. the extent of improvement with the dynamic odeling reaches . in . fig. b shows that the farthest-term performance of the first odel in the sequence of the dynamic modeling, same as the static odel, is significantly improved with the sequential rebuilding un- il the third quarter of , possibly as a consequence of the en- ancement of the set of characteristics. in this period, performance ncreases from . to . , meaning that the risk assessment is en- anced with the new dynamic modeling, rather than the static. from hat quarter onwards, the long-run predictions given by the dynamic odeling slightly improve, and always outperform the static frame. his suggests that, the new incoming data allow a better knowledge f the new context. although we know beforehand that the increase n performance is somewhat a consequence of the training being earer to the out-of-sample validation window, still we can see that sing the newest data improves the initial prediction given by the tatic model ( q ). . . memory - keep or lose it the new dynamic modeling framework enables investigating on hether it is preferable keeping a long-term memory or forgetting lder observations, or if they are equivalents in some contexts. from he second semester of onwards, the best results in the farthest- erm ( q ) are reached with the full memory configuration. owever, we realize that there is a certain overlap between the per- ormances of the sequence of models resulting from the two types of emory configurations, both for the short-term and for the farthest- erm. this suggests that, in the period, the information contained in he older examples remain appropriate for the default target, and that he context is not drifting as a result of particular changes in the set of haracteristics. hence, drifts in particular characteristics, like income, ranslate into virtual drifts because they did not have an impact in he distribution of target concept, p(y|x). to some extent, the imme- iate performance, exhibited in fig. a, decreases during from . to . %, which could be interpreted as the presence of a drift. owever, as the timeframe is small, it remains uncertain if it is a tran- itory outcome or a persistent drift in the context, potentially caused y changes in features that are not represented in the set of character- stics available for modeling in this application, like macroeconomic ata. . conclusions this research presents a new modeling framework for credit risk ssessment that extends the prevailing credit scoring models built pon historical data static settings. our framework mimics the prin- iple of films, by composing the model with a sequence of snapshots, ather than a single picture. within the new modeling schema, pre- ictions are made upon the sequence of temporal data, and are suit- ble for adapting to the occurrence of real concept drifts, translated y changes in the population, in the economy or in the market. it also nables improving the existing models based on the newest incoming ata. we present an empirical simulation using a real-world financial ataset of , credit cards, from a financial institution in brazil long two years of operation. a first conclusion is that monthly up- ates avoid the degradation of the model following the development. econdly, newest data consistently improve the forecasting accuracy, hen compared to the previous models in the sequence of dynamic odeling, both in a short-term as in a full-term memory configura- ion. particularly, the static model available at the beginning of the eriod is outperformed by every succeeding model, suggesting that he dynamic modeling framework has the ability of improving the rediction by integrating new incoming data. third, a slight domi- ance is achieved with the full-term memory, suggesting that older nformation remains meaningful for predicting default target within he analyzed period. in banking industry, prevailing credit scoring models are devel- ped from static windows and kept unchanged possibly for years. in his setting, the two basic mechanisms of memory, short-term and ong-term memory are fundamental to learning, but are still over- ooked in current modeling frameworks. as a consequence, these odels are insensitive to changes, like population drifts or financial istress. the usual outcomes are the default rates rising and abrupt redit cuts, as those that were observed in the u.s. in the aftermath f the last global crisis (as documented by sousa, gama, and brandão )). this problem could be overcome with the proposed frame- ork, since it would allow to gradually relearning along time and hanges. still, there are some real business problems with rebuilding mod- ls over time. first, lenders have little incentive to enhance the exist- ng rating systems frameworks because there is a recycling idea that t expensive and time-consuming to build new scorecards. then, they eed to be internally tested and validated, and then regulators need o approve them. second, regulators still promote models whose co- fficients do not change over time. this is one area where practice is m.r. sousa et al. / expert systems with applications ( ) – short memory full memory static model ( m ) . . . . . . . . . . -jan -mar -jun -sep -dec -mar -jun - sep c o e ff ic ie n t g in i model rebuilding month (a) short-term performance. short memory full memory static model ( m ) . . . . . . . . . . -jan -mar -jun -sep -dec -mar -jun - sep c o e ff ic ie n t g in i model rebuilding month (b) farthest-term performance, q . fig. . performance with the dynamic modeling – full and short memory – versus the static model ( m ). b b b c c c c c c d d d d e e f f g g h h j j k l l far distant from the technical advances, and new thoughts, like sim- plifying current decision layers, need to be encouraged. there are some important topics in default concept drift that we did not consider, which we defer for future research. while this pa- per provides convincing results, some additional simulations using real-world datasets from highly stressed economic environments and longer time frames would be valuable. second, modeling the delin- quency presents a specificity since a window of time is required in order to measure the outcome, i.e. the true class, before the new model is built. therefore, for forecasting, it turns out that there will be a time gap of the same length between the values of predictor variables used in the model and the first possible forecast period in the future. although this is not a problem of the proposed methodol- ogy, future research should bring new insights to overcome this issue, with a view on practicality. third, some good alternatives to using windows of data blocks are encouraged, which may be based on us- ing ensembles of the models learned in the past, possible combining the two components of memory, short-term and long-term memory, or a forgetting factor method. there is some material on this going back to adams et al. ( ). fourth, our empirical study considered a set of fixed predictors. therefore, future research should consider sets of predictor of variable length. this is important for detecting concept drift, because the set predictors that are being used may be quite limited to exhibit signs of change, even if they are occurring in the environment. finally, performance is reported in this paper, but the conditions leading to difference in performance are not explored. this is another future research direction. references abdou, h. a., & pointon, j. ( ). credit scoring, statistical techniques and evaluation criteria: a review of the literature. intelligent systems in accounting, finance and management, ( - ), – . adams, n. m., tasoulis, d. k., anagnostopoulos, c., & hand, d. j. ( ). temporally- adaptive linear classification for handling population drift in credit scoring. in y. lechevallier, & g. saporta (eds.), proceedings of compstat’ (pp. – ). physica-verlag hd. altman, e. i. ( ). financial ratios, discriminant analysis and the prediction of corpo- rate bankruptcy. journal of finance, ( ), – . anderson, r. ( ). the credit scoring toolkit: theory and practice for retail credit risk management and decision automation. oup oxford. avery, r. b., calem, p. s., & canner, g. b. ( ). consumer credit scoring: do situational circumstances matter? journal of banking & finance, ( ), – . baesens, b., van gestel, t., viaene, s., stepanova, m., suykens, j., & vanthienen, j. ( ). benchmarking state-of-the-art classification algorithms for credit scoring. journal of the operational research society, ( ), – . bcbs ( ). international convergence of capital measurement and capital standards: a revised framework - comprehensive version. bank for international settlements. bellotti, t., & crook, j. ( ). forecasting and stress testing credit card default using dynamic models. international journal of forecasting, ( ), – . bis ( ). implementation of basel ii: practical considerations. bank for international settlements. reiman, l., friedman, j., olshen, r., & stone, c. ( ). classification and regression trees. belmont, california: wadsworth international group. rics-cci&cbic ( ). ci algorithms competition (ciac): credit risk assessment sys- tem robustness against degradation and seasonal variation. http://brics- cci.org/ ci- algorithms- competition- ciac/ accessed . . . rown, i., & mues, c. ( ). an experimental comparison of classification algorithms for imbalanced credit scoring data sets. expert systems with applications, ( ), – . http://dx.doi.org/ . /j.eswa. . . . hawla, n. v., bowyer, k. w., hall, l. o., & kegelmeyer, w. p. ( ). smote: synthetic minority over-sampling technique. journal of artificial intelligence research, – . hen, d., zhong, y., liao, y., & li, l. ( ). review of multiple criteria and multi- ple constraint-level linear programming. procedia computer science, ( ), – . hen, m.-c., & huang, s.-h. ( ). credit scoring and rejected instances reassigning through evolutionary computation techniques. expert systems with applications, ( ), – . rook, j., & bellotti, t. ( ). time varying and dynamic models for default risk in consumer loans. journal of the royal statistical society: series a (statistics in society), ( ), – . rook, j. n., edelman, d. b., & thomas, l. c. ( ). recent developments in consumer credit risk assessment. european journal of operational research, ( ), – . rook, j. n., thomas, l. c., & hamilton, r. ( ). the degradation of the scorecard over the business cycle. ima journal of management mathematics, ( ), – . esai, v. s., crook, j. n., & overstreet jr, g. a. ( ). a comparison of neural networks and linear scoring models in the credit union environment. european journal of operational research, ( ), – . omingos, p. ( ). metacost: a general method for making classifiers cost-sensitive. in proceedings of the fifth acm sigkdd international conference on knowledge discov- ery and data mining. in kdd ’ (pp. – ). new york, ny, usa: acm. uda, r. o., hart, p. e., & stork, d. g. ( ). pattern classification. john wiley & sons. urand, d. ( ). risk elements in consumer installment financing. national bureau of economic research, inc. inav, l., jenkins, m., & levin, j. ( ). the impact of credit scoring on consumer lend- ing. the rand journal of economics, ( ), – . isenbeis, r. a. ( ). problems in applying discriminant analysis in credit scoring models. journal of banking & finance, ( ), – . ico ( ). introduction to scorecard for fico model builder, . isher, r. a. ( ). the use of multiple measurements in taxonomic problems. annals of eugenics, ( ), – . ama, j. ( ). knowledge discovery from data streams. london: chapman & hall/crc. ama, j., Žliobaitė, i., bifet, a., pechenizkiy, m., & bouchachia, a. ( ). a survey on concept drift adaptation. acm comput. surv., ( ), : – : . doi: . / . and, d. j. ( ). classifier technology and the illusion of progress. statistical science, ( ), – . sieh, n.-c. ( ). hybrid mining approach in the design of credit scoring models. expert systems with applications, ( ), – . ain, a. k., duin, r. p. w., & mao, j. ( ). statistical pattern recognition: a review. ieee transactions on pattern analysis and machine intelligence, ( ), – . ensen, h. l. ( ). using neural networks for credit scoring. managerial finance, ( ), – . linkenberg, r. ( ). learning drifting concepts: example selection vs. example weighting. intelligent data analysis, ( ), – . azarescu, m. m., venkatesh, s., & bui, h. h. ( ). using multiple windows to track concept drift. intelligent data analysis, ( ), – . ee, t.-s., & chen, i.-f. ( ). a two-stage hybrid credit scoring model using artificial neural networks and multivariate adaptive regression splines. expert systems with applications, ( ), – . http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://brics-cci.org/ci-algorithms-competition-ciac/ http://dx.doi.org/ . /j.eswa. . . http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://dx.doi.org/ . / http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref m.r. sousa et al. / expert systems with applications ( ) – l l l m m m m m m m m o p q r s s s s s s s s s t t t t w w w w w w y ee, t.-s., chiu, c.-c., lu, c.-j., & chen, i.-f. ( ). credit scoring using the hybrid neural discriminant technique. expert systems with applications, ( ), – . i, s.-t., shiue, w., & huang, m.-h. ( ). the evaluation of consumer loans using sup- port vector machines. expert systems with applications, ( ), – . ucas, a. ( ). updating scorecards: removing the mystique. in readings in credit scoring: foundations, developments, and aims (pp. – ). new york: oxford uni- versity press. alhotra, r., & malhotra, d. k. ( ). differentiating between good credits and bad credits using neuro-fuzzy systems. european journal of operational research, ( ), – . aloof, m. a., & michalski, r. s. ( ). incremental learning with partial instance memory. artificial intelligence, ( ), – . arqués, a. i., garcía, v., & sánchez, j. s. ( a). on the suitability of resampling tech- niques for the class imbalance problem in credit scoring. journal of the operational research society, ( ), – . arqués, a. i., garcía, v., & sánchez, j. s. ( b). two-level classifier ensembles for credit risk assessment. expert systems with applications, ( ), – . arqués, a. i., garcía, v., & sánchez, j. s. ( ). a literature review on the application of evolutionary computing to credit scoring. journal of the operational research so- ciety, ( ), – . artens, d., de backer, m., haesen, r., vanthienen, j., snoeck, m., & baesens, b. ( ). classification with ant colony optimization. ieee transactions on evolutionary com- putation, ( ), – . cnab, h., & wynn, a. ( ). principles and practice of consumer credit risk manage- ment. cib publishing. in, j. h., & lee, y.-c. ( ). bankruptcy prediction using support vector machine with optimal choice of kernel function parameters. expert systems with applica- tions, ( ), – . ng, c.-s., huang, j.-j., & tzeng, g.-h. ( ). building credit scoring models using ge- netic programming. expert systems with applications, ( ), – . avlidis, n., tasoulis, d., adams, n., & hand, d. ( ). adaptive consumer credit classi- fication. journal of the operational research society, ( ), – . uinlan, j. r. ( ). induction of decision trees. machine learning, ( ), – . ˇ ezáč, m., & řezáč, f. ( ). how to measure the quality of credit scoring models. finance a uver: czech journal of economics & finance, ( ), – . aberi, m., mirtalaie, m. s., hussain, f. k., azadeh, a., hussain, o. k., & ashjari, b. ( ). a granular computing-based approach to credit scoring modeling. neurocomput- ing, ( ), – . alganicoff, m. ( ). tolerating concept and sampling shift in lazy learning using pre- diction error context switching. artificial intelligence review, ( - ), – . chlimmer, j. c., & granger jr, r. h. ( ). incremental learning from noisy data. ma- chine learning, ( ), – . ousa, m. r., & da costa, j. p. ( ). a tripartite scorecard for the pay/no pay decision- making in the retail banking industry. frontiers in artificial intelligence and applica- tions, . ousa, m. r., gama, j., & brandão, e. ( ). links between scores, real default and pric- ing: evidence from the freddie mac’s loan-level dataset. journal of economics, busi- ness and management, ( ), – . ousa, m. r., gama, j., brandão, e., et al. ( a). introducing time-changing economics into credit scoring. technical report. universidade do porto, faculdade de econo- mia do porto. ousa, m. r., gama, j., & gonçalves, m. j. s. ( b). a two-stage model for dealing with temporal degradation of credit scoring. in proceedings of brics-cci & cbic. teenackers, a., & goovaerts, m. ( ). a credit scoring model for personal loans. in- surance: mathematics and economics, ( ), – . un, j., & li, h. ( ). dynamic financial distress prediction using instance selection for the disposal of concept drift. expert systems with applications, ( ), – . homas, l. c. ( ). consumer credit models: pricing, profit and portfolios: pricing, profit and portfolios. oxford university press. homas, l. c. ( ). consumer finance: challenges for operational research. journal of the operational research society, , – . homas, l. c., edelman, d. b., & crook, j. n. ( ). credit scoring and its applications. philadelphia: society for industrial and applied mathematics. symbal, a. ( ). the problem of concept drift: definitions and related work. dublin: computer science department, trinity college. ang, g., hao, j., ma, j., & jiang, h. ( ). a comparative assessment of ensemble learn- ing for credit scoring. expert systems with applications, ( ), – . ang, y., wang, s., & lai, k. ( ). a new fuzzy support vector machine to evaluate credit risk. ieee transactions on fuzzy systems, ( ), – . ei, g., yun-zhong, c., & ming-shu, c. ( ). a new dynamic credit scoring model based on the objective cluster analysis. in practical applications of intelligent sys- tems. in advances in intelligent systems and computing: (pp. – ). springer berlin heidelberg. est, d. ( ). neural network credit scoring models. computers & operations re- search, ( ), – . idmer, g., & kubat, m. ( ). effective learning in dynamic environments by explicit context tracking. in machine learning: ecml- (pp. – ). springer. idmer, g., & kubat, m. ( ). learning in the presence of concept drift and hidden contexts. machine learning, ( ), – . ang, y. ( ). adaptive credit scoring with kernel learning methods. european journal of operational research, ( ), – . http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref http://refhub.elsevier.com/s - ( ) - /sbref a new dynamic modeling framework for credit risk assessment introduction settings and concepts . supervised learning problem . score formulation . supervised classification methods dynamic modeling for credit default . concept drift in credit default . methods for adaptation case study . dataset and validation environment . data analysis and cleansing . data transformation and new characteristics . . data cleansing and new characteristics . . data transformation . . one-dimensional analysis . . interaction terms . . time series descriptive analysis . dynamic modeling framework experimental results . temporal degradation of static credit scoring . dynamic versus static . memory - keep or lose it conclusions references microsoft word - esa .doc page an adaptive knowledge acquisition system using generic genetic programming man leung wong department of computing and decision sciences lingnan university tuen mun hong kong mlwong@ln.edu.hk page an adaptive knowledge acquisition system using generic genetic programming abstract the knowledge acquisition bottleneck greatly obstructs the development of knowledge-based systems. one popular approach to knowledge acquisition uses inductive concept learning to derive knowledge from examples stored in databases. however, existing learning systems cannot improve themselves automatically. this paper describes an adaptive knowledge acquisition system that can learn first-order logical relations and improve itself automatically. the system is composed of an external interface, a biases base, a knowledge base of background knowledge, an example database, an empirical ilp learner, a meta- level learner, and a learning controller. in this system, the empirical ilp learner performs top-down search in the hypothesis space defined by the concept description language, the language bias, and the background knowledge. the search is directed by search biases which can be induced and refined by the meta-level learner based on generic genetic programming. it has been demonstrated that the adaptive knowledge acquisition system performs better than foil on inducing logical relations from perfect or noisy training examples. the result implies that the search bias evolved by evolutionary learning is better than that of foil which is designed by a top researcher in the field. consequently, generic genetic programming is a promising technique for implementing a meta-level learning system. the result is very encourage as it suggests that the process of natural selection and evolution can successfully evolve a high performance learning system. area: evolutionary computation, genetic programming, knowledge acquisition . introduction the knowledge acquisition bottleneck greatly obstructs the development of knowledge-based systems. one popular approach to knowledge acquisition uses inductive concept learning to derive knowledge from examples stored in databases. the knowledge acquired can be expressed in different knowledge representations such as first-order logical relations, decision trees, decision lists, and production rules. existing learning systems such as cart (breiman et al. ), c . (quinlan page ), assistant (cestnik et al. ), aq (michalski et al ), and cn (clark and niblett ) use attribute-value language for representing the training examples and the induced knowledge and allow a finite number of objects in the universe of discourse. this representation limits them to learn only propositional descriptions in which concepts are described in terms of values of a fixed number of attributes. dzeroski and lavrac show that inductive logic programming (ilp) can be used to induce knowledge represented as first-order logical relations (dzeroski and lavrac , dzeroski ). ilp is more powerful than traditional inductive learning methods because it uses an expressive first-order logic framework and facilitates the application of background knowledge. in this formalism, domain knowledge represented in the form of relations can be used in the induced relational descriptions of concepts. moreover, ilp has a strong theoretical foundation from logic programming and computational learning theory. the task of inducing first-order logical relations can be formulated as a search problem (mitchell ) in a hypotheses space of logical relations. various approaches (quinlan ; , muggleton and feng ) differ mainly in the search strategy and the heuristics used to guide the search. the search space is extremely large, so strong heuristics are required to manage the problem. most systems are based on a greedy search strategy. they generate a sequence of logical relations from general to specific (or from specific to general) until a consistent relation is found. each relation in the sequence is obtained by specializing (or generalizing) the previous one. for example, foil (quinlan , ) applies a hill climbing search strategy guided by an information-gain heuristic to search relations from general to specific. but these strategies and heuristics are not always applicable because these systems may become trapped in local maxima. in order to overcome this problem, non greedy strategies should be adopted. moreover, existing ilp systems cannot improve themselves automatically. in this paper, we describe an adaptive knowledge acquisition system that induces first-order logical relations and improves itself during learning. we formulate page the definitions of inductive concept learning and adaptive knowledge acquisition in the next section. the system is based on a generic genetic programming approach that is presented in section . a generic top-down first-order learning algorithm is described in the next section. the fifth section contains a description of a meta-level learner that induces search bias. the experimentation and some evaluations of the system are reported in section six. finally, the conclusion is presented in the last section. . inductive concept learning and adaptive knowledge acquisition the goal of machine learning is to develop techniques and tools for building intelligent learning machines. machine learning paradigms include inductive, deductive, genetic-based, and connectionist learning. multi-strategy learning integrates several learning paradigms. this section focuses on supervised inductive concept learning. if u is a universal set of observations, a concept c is formalized as a subset of observations in u. inductive concept learning finds descriptions for various target concepts from positive and negative training instances of these concepts. in machine learning, formal languages for describing observations and concepts are called object and concept description languages respectively. typically, object description languages are attribute-value pair descriptions and first-order languages of horn clauses. concepts can be described extensionally or intensionally. a concept is described extensionally by listing the descriptions of all of its instances (observations). thus extensional concepts are represented in the object description language. on the other hand, intensional concepts are expressed in a separate concept description language that permits compact and concise concept descriptions. typical concept description languages are decision trees, decision lists, production rules, and first-order logic. inductive concept learning can be viewed as searching the space of hypothesis descriptions. a bias is a mechanism employed by a learning system to constrain the search for target hypotheses. a search bias determines how to conduct the search in page the hypothesis space while a language bias determines the size and structure of the hypothesis space. a strong search bias, such as the hill-climbing search strategy, employs existing knowledge about the size and structure of the hypothesis space to exploit promising solutions of the space, thus it can find the target concept quickly. but it may trap the system in a local maximum. a weak search bias, such as depth-first and breath-first search, explores the space completely; the learner is guaranteed to find the target concept that can be represented by the concept description language. nevertheless, a weak bias is very inefficient. in other words, the search bias introduces the efficiency/completeness tradeoff into a learning system. a strong language bias defines a less expressive description language such as the propositional logic. the hypothesis space created by the bias is comparatively smaller and the learning can be performed more efficiently. but the learner may fail to find the target concept which is not contained in the small hypothesis space. a weak bias defines a larger space and thus the target concept is more likely to be expressible in the space. the disadvantage is that the learner is less efficient. the language bias introduces the efficiency/expressiveness tradeoff into a learning system. background knowledge b is declarative prior knowledge that can be used by either the search bias to direct the search more efficiently, or the language bias to express the hypothesis space in a more natural and concise way. background knowledge plays an important role in relational concept learning. relational concept learning induces a new relation for the target concept (i.e., the target predicate) from training examples and known relations from the background knowledge. the training examples, the hypothesis space, and the background knowledge are represented in first order horn clause languages (muggleton ). tradeoffs between expressiveness and efficiency are introduced by some additional restrictions on the three languages representing the examples, the hypothesis space, and the background knowledge. the background knowledge b provides definitions of known predicates qi which can be used in the definition of the target predicate p. it also provides page additional information to ease the learning. the information includes argument types, symmetry of predicates in pairs of arguments, input/output modes, rule models, predicate sets, parametrized languages, integrity constraints, determinations and any knowledge that can modify the operation of the search and language biases (lavrac and dzeroski ). an adaptive knowledge acquisition system is a relational concept learning system that can improve itself on the learning capability. it maintains various sets of background knowledge and biases. it improves itself by modifying its biases and background knowledge. since a hypothesis space for learning is defined through the concept description language, the language bias, and the background knowledge, the size and structure of the hypothesis space can be modified by changing the language bias and the background knowledge. the search strategy and heuristics are changed if the system's search biases are modified. here, we formulate the task of an adaptive knowledge acquisition system in table . given: -a set e of positive e+ and negative e- training examples of the target predicate p. training examples are represented as ground atoms -a concept description language l -a set of learning biases biases -a set of various background knowledge bks find: -a modified set of learning biases biases' -a modified set of background knowledge bks' -a concept definition h for the target predicate p expressible in l such that h is complete and consistent with respect to (w.r.t.) the training examples e and a background knowledge b' in bks' h is complete if every positive example e+ in e+ is covered by h w.r.t. the background knowledge b. i.e. b ∪ h |= e+ h is consistent if no negative example e- in e- is covered by h w.r.t. the background knowledge b. i.e. b ∪ h |≠ e- table : the definition of adaptive knowledge acquisition page external interface learning controller bkbase biases base example database meta-level learner empirical ilp learner data flow control flow figure : the logical organization of an adaptive knowledge acquisition system the logical organization of our system is depicted in figure . its components are introduced as follows: ( ) external interface: it provides a user-friendly interface between the system and users. it accepts training examples, a set bks of background knowledge, and a set biases of biases and transfers them through the learning controller to the example database, bkbase, and biases base respectively. the interface also provides commands for users to query about the results of an adaptive learning task and to directly control the operations of the learning controller. ( ) biases base: it is a knowledge base that stores all learning biases. biases can be retrieved, added, deleted, and modified through the interface of this knowledge base. ( ) bkbase: it stores various background learning knowledge that can be used in inductive learning. background knowledge can be retrieved, added, deleted, and modified through the interface of bkbase. since each entity of it is in fact a complex structure representing background knowledge, bkbase is implemented using object-oriented techniques. page ( ) examples database: it stores the training examples. ( ) empirical ilp learner: it induces first-order logical relations from the training examples, given a concept description language, a specific background knowledge, a search bias, and a language bias. a search of the hypothesis space can be performed bottom-up or top-down. bottom-up techniques start from the training examples and search the space by employing various generalization operators. top-down techniques start from the most general concept descriptions, and search the space by using various specialization operators. top-down techniques are better suited for learning from imperfect examples because a large number of data are available in every specialization step and the system can employ various statistical techniques to decide how to perform the specialization. moreover, top-down search can easily be guided by the search bias. in section , a generic top-down first-order learning algorithm is described. ( ) meta-level learner: it learns search biases, language biases, and background knowledge. search and language biases can be represented declaratively or procedurally. in this paper, we apply ggp to implement the meta-level learner that induces procedural biases. the description of ggp is presented in section . ( ) learning controller: it is a knowledge-based system that controls the empirical ilp learner and the meta-level learner. the knowledge used by the learning controller can be updated by the meta-level learner. . generic genetic programming (ggp) generic genetic programming (ggp) is a novel approach that combines genetic programming (koza ; , kinnear ) and inductive logic programming (quinlan ; , muggleton ). using ggp, programs in various programming languages can be evolved. the approach is also powerful enough to handle context-sensitive information and domain-dependent knowledge which can be used to accelerate the learning speed and/or improve the quality of the programs. page ggp can induce programs in various programming languages. this is achieved by accepting or choosing grammars of different languages to produce programs in these languages. most modern programming languages are specified in the notation of bnf (backus-naur form) which is a kind of context-free grammars (cfgs). however, ggp is based on logic grammars because cfgs (hopcroft and ullman , lewis and rapadimitrion ) are not expressive enough to represent context-sensitive information for some languages and domain-dependent knowledge of the target program being induced. this section first introduces the formalism of logic grammars followed by the descriptions of ggp. . . introduction to logic grammars logic grammars are the generalizations of cfgs. their expressivenesses are much more powerful than those of cfgs, but equally amenable to efficient execution. in this paper, logic grammars are described in a notation similar to that of definite clause grammars (pereira and warren , pereira and shieber , sterling and shapiro ). the logic grammar for some simple s-expressions in table will be used throughout this section. : start -> [(*], exp(w), exp(w), exp(w), [)]. : start -> {member(?x,[w, z])}, [(*], exp- (?x), exp- (?x), exp- (?x), [)]. : start -> {member(?x,[w, z])}, [(+], exp- (?x), exp- (?x), exp- (?x), [)]. : exp(?x) -> [(/ ?x . )]. : exp- (?x) -> {random( , ,?y)}, [(/ ?x ?y)]. : exp- (?x) -> {random( , ,?y)}, [(- ?x ?y)]. : exp- (w) -> [(+ (- w ) )]. table : a logic grammar a logic grammar differs from a cfg in that the logic grammar symbols, whether terminal or non-terminal, may include arguments. the arguments can be any term in the grammar. a term is either a logic variable, a function or a constant. a variable is represented by a question mark ? followed by a string of letters and/or digits. a function is a grammar symbol followed by a bracketed n-tuple of terms and page a constant is simply a -arity function. arguments can be used in a logic grammar to enforce context-dependency. thus, the permissible forms for a constituent may depend on the context in which that constituent occurs in the program. another application of arguments is to construct tree structures in the course of parsing, such tree structures can provide a representation of the semantics of the program. the terminal symbols enclosed in square brackets correspond to the set of words of the language specified. for example, the terminal [(- ?x ?y)] creates the constituent (- . . ) of a program if ?x and ?y are instantiated respectively to . and . . non-terminal symbols are similar to literals in prolog, exp- (?x) in table is an example of non-terminal symbols. commas denote concatenation and each grammar rule ends with a full stop. the right-hand side of a grammar rule may contain logic goals and grammar symbols. the goals are pure logical predicates for which logical definitions have been given. they specify the conditions that must be satisfied before the rule can be applied. for example, the goal member(?x, [w, z]) in table instantiates the variable ?x to either w or z if ?x has not been instantiated, otherwise it checks whether the value of ?x is either w or z. if the variable ?y has not been bound, the goal random( , , ?y) instantiates ?y to a random floating point number between and . otherwise, the goal checks whether the value of ?y is between and . domain-dependent knowledge can be represented in logic goals. for example, consider the following grammar rule: a-useful-program -> first-component(?x), {is-useful(?x, ?y)}, second-component(?y). this rule states that a useful program is composed of two components. the first component is generated from the non-terminal first-component(?x). the logic variable ?x is used to store semantic information about the first component produced. the logic goal then determines whether the first component is useful according to the semantic information stored in ?x. domain-dependent knowledge page about which program fragments are useful is represented in the logical definition of this predicate. if the first component is useful, the logic goal is- useful(?x, ?y) is satisfied and some semantic information is stored into the logic variable ?y. this information will be used in the non-terminal second- component(?y) to guide the search for a good program fragment as the second component of a useful program. the special non-terminal start corresponds to a program of the language. in table , some grammar symbols are shown in bold-face to identify the constituents that cannot be manipulated by genetic operators. for example, the last terminal symbol [)] of the second rule is revealed in bold-face because every s-expression must be ended with a ')'. the number before each rule is a label for later discussions. it is not part of the grammar. . . representations of programs one of the fundamental contributions of ggp is in the representations of programs in different programming languages appropriately so that initial population can be generated easily and the genetic operators such as reproduction, mutation, and crossover can be performed effectively. a program can be represented as a derivation tree that shows how the program has been derived from the logic grammar. ggp applies deduction to randomly generate programs and their derivation trees in the language declared by the given grammar. these programs form the initial population. for example, the program (* (/ w . ) (/ w . ) (/ w . )) can be generated by ggp given the logic grammar in table . it is derived from the following sequence of derivations: start => [(*] exp(w) exp(w) exp(w) [)] => [(*] [(/ w . )] exp(w) exp(w) [)] => [(*] [(/ w . )] [(/ w . )] exp(w) [)] => [(*] [(/ w . )] [(/ w . )] [(/ w . )] [)] => [(* (/ w . ) (/ w . ) (/ w . ))] this sequence of derivations can be represented as the derivation tree depicted in figure . page in literature, the terms derivation trees and parse trees are usually used interchangeably. however, we will use the term derivation trees to refer to the tree structures in our framework and the term parse trees to refer to those in gp. the bindings of logic variables are enclosed in a pair of braces. the sub-trees enclosed in a dashed rectangular are frozen. in other words, they are generated by bold-faced grammar symbols and they cannot be modified by genetic operators. an advantage of logic grammars is that they specify what is a legal program without any explicit reference to the process of program generation and parsing. furthermore, a logic grammar can be translated into an efficient logic program that can generate and parse the programs in the language declared by the logic grammar (pereira and warren , pereira and shieber , abramson and dahl ). in other words, the process of program generation and parsing can be achieved by performing deduction using the translated logic program. consequently, the program generation and analysis mechanisms of ggp can be implemented using a deduction mechanism based on the logic programs translated from the grammars. [(*] exp(w) exp(w) exp(w) [)] start [(/ ?x . )] {?x/w} [(/ ?x . )] {?x/w} [(/ ?x . )] {?x/w} figure : a derivation tree of the s-expression in lisp (* (/ w . ) (/ w . ) (/ w . )) this method of translating a logic grammar into a logic program is common in the field of natural language processing (pereira and warren , pereira and shieber , abramson and dahl ). the original idea of this approach is to page rephrase the special purpose formalism of cfgs into a general purpose first-order predicate logic (kowalski , colmerauer , pereira and warren ). this approach is further refined and generalized to definite clause grammars (dcgs) which can handle the properties of context-dependency of natural languages effectively. since dcgs, a kind of logic grammars, can be translated into efficient logic programs automatically, parsers and generators for the corresponding natural languages can be obtained easily. in other words, researchers in the field of natural language processing only declare the grammar for a particular natural language, and the translation process will produce the corresponding parser and generator for them. moreover, for some cases, the same logic program can be used as both a parser and generator at the same time. alternatively, initial programs can be induced by other learning systems such as foil (quinlan ; ) or given by the user. ggp analyzes each program and creates the corresponding derivation tree. . . the evolution process of ggp in ggp, populations of programs are genetically bred using the darwinian principle of survival and reproduction of the fittest along with genetic operations appropriate for creating programs. ggp starts with an initial population of programs generated randomly, induced by other learning systems, or provided by the user. logic grammars provide declarative descriptions of the valid programs that can appear in the initial population. a fitness function must be defined by the user to evaluate the fitness values of the programs. typically, each program is run over a set of fitness cases and the fitness function estimates its fitness by performing some statistical operations (e.g. average) to the values returned by this program. the initial programs in generation are normally incorrect and have poor performances. however, some programs in the population will be fitter than others. fitness of each program in the generation is estimated and the following process is iterated over many generations until the termination criterion is satisfied. the reproduction, sexual crossover, and asexual mutation are used to create new page generation of programs from the current one. the reproduction involves selecting a program from the current generation and allowing it to survive by copying it into the next generation. either fitness proportionate or tournament selection can be used. the crossover is used to create offspring programs from two parental programs selected. mutation creates a modified offspring program from a parental program selected. unlike crossover, the offspring program is usually similar to the parent program. logic grammars are used to constraint the offspring programs that can be produced by these genetic operations. this algorithm will produce populations of programs which tend to exhibit increasing average of fitness. finally, ggp returns the best program found in any generation of a run as the result. a high level algorithm of ggp is presented in table . . generate an initial population of programs. . execute each program in the current population and assign it a fitness value according to the fitness function . if the termination criterion is satisfied, terminate the algorithm. the best program found in the run of the algorithm is designated as the result. . create a new population of programs from the current population by applying the reproduction, crossover, and mutation operations. these operations are applied to programs selected by fitness proportionate or tournament selections. . rename the new population to the current population. . proceed to the next generation by branching back to the step . table : the high level algorithm of ggp . a generic top-down first-order learning algorithm this section presents a generic top-down first-order learning algorithm based on foil (quinlan , ). the algorithm is depicted in table . the algorithm consists of three steps. in the pre-processing step, missing argument values in training examples are handled by assigning default or random values to them. a training example will be removed if it has too many missing values. if there are no or page inadequate negative examples in the training set, they can be generated. different ways of creating negative examples have been proposed (lavrac and dzeroski ). input: e: training examples l: the concept description language biassearch: the search bias biaslang: the language bias b: background knowledge t: the target concept output: a relation p which contains a set of clauses. each clause c ∈ l. function learning(e, l, biassearch, biaslang, b, t) ( ) pre-processing of the training examples e and producing a modified set of examples e': e' := preprocessing(e). ( ) let ecurrent := e'; let p := {}; repeat -let c := t ←; -find a specialization c' of c. this step constructs a clause c' from c by calling clause-construct(c, ecurrent, b, l, biassearch, biaslang); -if a specialization can be found -add c' to p to produce a new relation p'. i.e. p' := p ∪ {c'}; -remove all positive examples covered by p' from ecurrent to get an updated training set e' e'current := ecurrent - { positive examples in ecurrent covered by p' w.r.t. the background knowledge b}; -let ecurrent := e'current; -let p := p' else -set the flag no-more-improvement to true; until the covering termination criterion is satisfied. i.e. covering-termination(p, no-more-improvement, ecurrent, b) returns true; ( ) post-processing the relation p and producing p'. i.e. p' := post-processing(p); return(p'); table : a generic top-down first-order learning algorithm page the second step performs the construction of a logical relation. this step employs four local variables: ecurrent (current training examples set), e'current (updated training examples set), p (current relation) and p' (modified relation). the main component of this step is the covering loop which implements michalski's covering algorithm (michalski et al. a). the covering loop construct a relation by iteratively executing the following sub-steps: (a) construct a clause that covers some positive examples in ecurrent. (b) append the clause to the current relation p and generate a modified relation p'. (c) remove all positive examples from ecurrent which are covered by p' with respect to the background knowledge b. the covering loop terminates if the terminating conditions are satisfied. a typical condition is that either all positive examples are covered or no more improvement can be achieved by searching for a new clause. the final step attempts to improve the accuracy of the relation induced when classifying unseen examples and to simplify the relation. the covering loop calls the 'clause-construct' function which is the core of the generic algorithm. a hill-climbing 'clause-construct' algorithm is presented in table . the function constructs a clause cn = t ← l , l , ..., ln starting from the most general clause c = t ← with an empty body. a sequence of clauses c , c , c , c , ...., cn are generated by a number of specialization steps. at each step, the current clause ci = t← l ,l , ..., li is refined by appending a specific literal lj to its body. a literal lj is constructed from the background knowledge b restricted by the concept description language l and language bias biaslang. the language may limit lj to be function-free while biaslang may prevent new variable to be introduced in lj. the aim of the procedure is to find a clause which covers most positive examples while excludes all or most negative examples. in a hill-climbing search, the procedure keeps the current page best clause and refines it using the estimated best specialization at each step, until the stopping condition is satisfied the 'clause-construct' function calls the 'find-extension' function to find the extension ei of the current training examples given the partially developed clause ci = t(x , x , ..., xn) ← l , l , ..., li and the background knowledge b. each training example is a n-tuple where xi, ≤i≤n, are some constants. to find the extension, the function initializes a clause c = t(x , x , ..., xn), then the literal l is added to the body of c to produce a new clause c . the literal l is either of the form xj = xk, xj ≠ xk, pm(y , y , ...,ysm ) or not pm(y , y , ...,ysm ). if the literal contains k new variables, the arity of each tuple in the generated training set e increases to (n + k). e can be found by performing a natural join of ecurrent with the relation corresponding to literal l . the process is repeated for literals l , l , ..., li until the extension ei is found. the most important component of the hill-climbing 'clause-construct' algorithm is the 'scoring' function that estimates the performance of each literal. an accurate estimation directs the search towards the global maxima while a misleading one traps the system into local-maxima. by providing different 'scoring' functions to the generic learning algorithm, various learning algorithms can be generated. the performances of a good and a bad learners can be significant different as shown in sub-section . . page input: c: an initial clause c = t← ecurrent: the current training examples b: background knowledge l: the concept description language biassearch: the search bias biaslang: the language bias output: a clause that covers some positive examples in ecurrent while excludes all or most negatives examples in ecurrent function clause-construct(c, ecurrent, b, l, biassearch, biaslang) there is a scoring function stored in biassearch, save this function to scoring; repeat -set best to a bad literal such as x = x where x is a variable appearing in the head of the clause; -set best-score to ; -find the extension ei of ecurrent using the clause c w.r.t. b. i.e. ei := find-extension(c, ecurrent, b); -let ni + be the number of positive tuples in ei; -let ni − be the number of negative tuples in ei; -current-information := − log ( ni + / (ni + + ni − )); -for all literal l from b that satisfy the constraints imposed by the language l and bias biaslang -set c ' = c ∪ {l}; -find the extension ei+ of ecurrent using the clause c ' i.e. ei+ := find-extension(c ', ecurrent, b); -let ni+ + be the number of positive tuples in ei+ ; -let ni+ − be the number of negative tuples in ei+ ; -let the number of positive tuples in ei that have been represented by one or more tuples in ei+ be ni ++ ; -find the score of the literal l by using the scoring function i.e. literal-score := scoring (ni ++ , ni+ + , ni+ − , current-information); -if literal-score > best-score then -best := l; -best-score := literal-score; -if best == x=x then -no-more-improvement := true; else -append best to the body of c; until clause-termination(c, no-more-improvement, ecurrent, b) is true; post-processing the clause c to find an improvement i.e. c' := find-improvement(c); if acceptable(c') -return(c'); else -return(no-specialization-can-be-found); table : a hill-climbing 'clause-construct' algorithm page . inducing procedural search biases in this section, ggp is used in the meta-level learner to induce procedural search biases (i.e. the 'scoring' function). in order to employ ggp, a logic grammar must be defined (table ). in the grammar, the terminal symbols n-pos-i-plus- , n-neg-i-plus- , and n-pos-i represent respectively ni + + , ni + − and ni + + . with reference to the algorithms in tables and , assume that ei is the extension of current training examples ecurrent by current clause ci, ni + and ni − are respectively the number of positive and negative tuples in ei. ei can be extended by using the literal l to ei+ . ni+ + and ni+ − are respectively the number of positive and negative tuples in ei+ . ni ++ is the number of positive tuples in ei that have been represented by one or more tuples in ei+ . the terminal symbol current-information is defined as − log ( ni + / (ni + + ni − )) . start -> function. s-exp -> term. s-exp -> function. function -> [(], op , s-exp, [)]. function -> [(], op , s-exp, s-exp, [)]. op -> [ protected-log ]. op -> [ + ]. op -> [ - ]. op -> [ * ]. op -> [ % ]. op -> [ info ]. term -> [ n-pos-i-plus- ]. term -> [ n-neg-i-plus- ]. term -> [ n-pos-i ]. term -> [ current-information ]. term -> { random(- , , ?a) }, [ ?a ]. table : a logic grammar for learning procedural search bias the terminal symbols +, -, and * represent functions that perform ordinary addition, subtraction, and multiplication respectively. the symbol % represents page function that normally returns the quotient. however, if division by zero is attempted, the function returns . . the symbol protected-log is a function that calculates the logarithm of the input argument x if x is larger than zero, otherwise it returns . . the symbol info represents the basic function that calculates − log ( x / ( x + y )) given x and y as inputs. the logic goal random(- , , ?a) generates a random floating point number between - and and instantiates ?a to the random number generated . . the evolution process the evolution process of the adaptive knowledge acquisition system is depicted in figure . firstly, the biases base is initialized with a population of different 'scoring' functions generated randomly using the logic grammar depicted in table . to estimate the fitness of a specific 'scoring' function, it is combined with the generic top-down learner to produce a specific learner. the performance of this specific learner is then evaluated by using a fitness function. this measure is assigned as the fitness of the specific 'scoring' function. ggp employs selection, crossover, and mutation to generate potentially better functions. the modified functions are stored in the biases base and the whole evolution process iterates until the best function is found or no computational resource is available page biases base ggp empirical ilp learner bkbase example database modified biases initial biases bias performance of bias figure : the evolution process of the adaptive knowledge acquisition system . . the experimentation setup in this paper, learning curves are used to estimate the performances of various learning systems. the example space is divided randomly into disjoint training and testing sets. the learner is trained on progressively larger portions of the training set and the performance of the induced logical relation is estimated on the disjoint testing set. this process of dividing, training, and testing is repeated for trials and the results are averaged to generate a learning curve. as a running example, we use a traditional problem discussed in the literature (muggleton and feng ). in the problem of learning the list predicate member, the data consist of all lists of lengths to defined over three constants. the background knowledge b contains definitions of list construction predicates: null which holds for an empty list and component which decomposes a list into its head and tail. the example space contains positive and negative examples. the training sets contain to examples, one-half of each training set is positive examples. the testing set consists of positive and negative examples. page . . fitness calculation adjusted and normalized fitness values are used as in koza ( ). they are calculated from the raw fitness which is estimated by the fitness function. various fitness functions have been tried and two of them are described here. the impact of fitness function on the generality of the evolved function is also demonstrated. the problem domain of learning the member predicate is used here. for the first fitness function, a random set of positive and negative examples is used. a specific 'scoring' function is combined with the generic top-down learner to produce a specific learner called adapted-ilp hereafter. adapted-ilp induces first-order logical relations using the random example set. the quality of the induced logical relations is evaluated by counting the total number of misclassified examples from the same training set. this measure is used as the raw fitness of the specific 'scoring' function. using this fitness function, only poor 'scoring' functions have been evolved. the learning curve of a poor learner is depicted in figure . for the second fitness function, the raw fitness is developed in several steps. at the beginning of each generation, four instances of the learning task are created randomly from the member domain. each learning task has a training and a disjoint testing data. the training set contains positive and negative examples. for each learning task, a specific adapted-ilp induces logical relations from the training set and the relations are evaluated by counting the number of misclassified examples from the testing set. the performance of the adapted-ilp is the sum of numbers of misclassified examples for all learning tasks. this measure is then used as the raw fitness of the corresponding 'scoring' function. this fitness function can force the evolution of good 'scoring' functions. the learning curve of a good learner is shown in figure . page training size . . . . . . a cc ur ac y poor ilp learner good ilp learner figure : the learning curves of good and poor learner . experimentation and evaluations this section compares the performance of our system with that of foil (quinlan ; ). standard learning tasks in the literature are used in these experiments (quinlan , muggleton and feng ). . . the member predicate the learning curves for this problem are depicted in figure . it is interested to find that our system has higher accuracy than foil. the difference is significant at % level of significance when the training size is less than . training size . . . . . a cc ur ac y foil the adaptive ilp system figure : learning curves for the member problem page . . the member predicate in a noisy environment difference amount of noise is introduced into the training examples in order to study the performances of both systems in learning relations in noisy environment. to introduce n% of noise into the examples, n% positive examples are labeled as negative ones while n% negative examples are labeled as positive ones. in this experiment, the percentages of introduced noise are % ( . ) and % ( . ). their learning curves are summarized in figure . our system performs better than foil at all noise level. training size . . . . . a cc ur ac y foil ( . ) the adaptive ilp system ( . ) foil ( . ) the adaptive ilp system ( . ) figure : learning curves for the member problem in a noisy environment . . the multiply predicate in the problem of learning the arithmetic predicate multiply (muggletion and feng ), the data contain integers in the range from zero to ten. the background knowledge is composed of definitions for arithmetic predicates plus, decrement, zero, and one. the example space has positive and negative examples respectively. the training sets consist of to examples, one-tenth of each training set is positive and the remainder is negative. the learning curves for multiply are presented in figure . our system performs better than foil when the size of training set is less than . the difference is significant at % level of significance. page training size . . . . a cc ur ac y foil the adaptive ilp system figure : learning curves for the multiply problem . . the uncle predicate another traditional testbed for relational learners is the domain of family relationships (quinlan ). in this experiment, the uncle predicate is induced and the background predicates are parent, sibling, married, male, and female. the learning curves are presented in figure . training size . . . . . . a cc ur ac y foil the adaptive ilp system figure : learning curves for the uncle problem page . conclusion in this paper, we formulate an adaptive knowledge acquisition system which is composed of an external interface, a biases base, a knowledge base of background knowledge, an example database, an empirical ilp learner, a meta-level learner, and a learning controller. an implementation of the adaptive knowledge acquisition system has been developed. in the implementation, the empirical ilp learner performs top- down search in the hypothesis space defined by the concept description language, the language bias, and the background knowledge. the search is directed by search biases which can be induced and refined by a meta-level learner implemented by using generic genetic programming. generic genetic programming (ggp) is a novel, general, and powerful approach of evolving search biases. it is also powerful enough to use logic grammars to represent context-sensitive information and domain-dependent knowledge. the idea of using formal grammars to direct search for knowledge in the hypothesis space or to reduce the size of the space has also been independently studied by other researcher recently (cohen , gruau , whigham ; ). it has been demonstrated that the induced bias is better than that of foil on many standard learning tasks. from these experiments, it can be concluded that the adaptive knowledge acquisition system has superior learning ability compared to foil. since they are different in their search biases only, the result implies that the search bias induced by ggp is better than that of foil for the learning problems. this result is surprising because the search biases of the adaptive knowledge acquisition system are initialized by a random process. these biases are normally poor, but the process of natural selection and evolution can successfully evolve a good bias. it is important to mention that the search bias is rather general because it has reasonable performance on many traditional learning problems using the same bias acquired automatically. this paper illustrates that ggp is a plausible approach for implementing a meta-level learning system. for future work, in order to find a page general, efficient, and effective bias, a large number of learning tasks of different kinds, such as the member, append, quick sort, ackermann, uncle, and grandfather problems, with various characteristics should be used. this adaptive learning approach, though computationally intensive, is rather exciting, as it opens up many opportunities for creating or improving learning algorithms. page reference abramson, h. and dahl, v. ( ). logic grammars. berlin: springer-verlag. bremen, l., friedman, j. h., olshen, r. a. and stone, c. j. ( ). classification and regression trees. belmont: wadsworth. cestnik, b., kononenko, j. and bratko, i. ( ). assistant : a knowledge elicitation tool for sophisticated users. in i. bratko and n. lavrac (ed.), progress in machine learning, - . wilmslow: sigma press. clark, p. and niblett, t. ( ). the cn induction algorithm. machine learning; , - . cohen, w. ( ). compiling prior knowledge into an explicit bias. in proceedings of the ninth international workshop on machine learning, - . ca: morgan kaufmann. colmerauer, a. ( ). metamorphosis grammars. in l. bolc (ed.), natural language communication with computers. berlin: springer-verlag. dzeroski, s. ( ). inductive logic programming and knowledge discovery in databases. in u. m. fayyad, g. piatetsky-shapiro, p. smyth, and r. uthurusamy (eds.), advances in knowledge discovery in data mining, - . menlo park, ca: aaai press. dzeroski, s. and lavrac, n. ( ). inductive learning in deductive databases. ieee transactions on knowledge and data engineering, , - . gruau, f. ( ). on using syntactic constraints with genetic programming. in p. j. angeline and k. e. kinnear, jr. (eds.) advances in genetic programming , - . ma: mit press. hopcroft, j. e. and ullman, j. d. ( ). introduction to automata theory, languages, and computation. ma: addison-wesley. koza, j. r. ( ). genetic programming ii: automatic discovery of reusable programs. cambridge, ma: mit press. koza, j. r. ( ). genetic programming: on the programming of computers by means of natural selection. cambridge, ma: mit press. kinnear, k. e. jr., editor ( ). advances in genetic programming. cambridge, ma: mit press. kowalski, r. a. ( ). logic for problem solving. amsterdam: north-holland. page lavrac, n. and dzeroski, s. ( ). inductive logic programming: techniques and applications. london: ellis horword. lewis, h. r. and rapadimitrion, c. h. ( ). elements of the theory of computation. nj: prentice hall. michalski, r. s., mozetic, i., hong, j. and lavrac, n. ( ). the multi-purpose incremental learning system aq and its testing application on tree medical domains. in proceedings of the national conference on artificial intelligence, - . san mateo, ca: morgan kaufmann. mitchell, t. m. ( ). generalization as search, artificial intelligence, , - . muggletion, s., editor ( ). inductive logic programming. london: academic press. muggleton, s. and buntine, w. ( ). machine invention of first-order predicates by inverting resolution. in proceedings of the fifth international conference on machine learning, - . ca: morgan kaufmann muggleton, s. and feng, c. ( ), efficient induction of logic programs, in proceedings of the first conference on algorithmic learning theory, - . pereira, f. c. n. and warren, d. h. d. ( ). definite clause grammars for language analysis - a survey of the formalism and a comparison with augmented transition networks artificial intelligence, , - . pereira, f. c. n. and shieber, s. m. ( ). prolog and natural-language analysis. ca: csli. quinlan, j. r. ( ). learning logical definitions from relations. machine learning, , - . sterling, l. and shapiro, e. ( ). the art of prolog. cambridge, ma: mit press. whigham, p. a. ( ). search bias, language bias and genetic programming. in proceedings of the first genetic programming conference, - . cambridge, ma: mit press. whigham, p. a. ( ). inductive bias and genetic programming. in proceedings of the first international conference on genetic algorithms in engineering systems: innovations and applications, - . uk: iee. annals of emerging technologies in computing (aetic) vol. , no. , obada alhabashneh, "fuzzy-based adaptive framework for module advising expert system”, annals of emerging technologies in computing (aetic), print issn: - , online issn: - x, pp. - , vol. , no. , st january , published by international association of educators and researchers (iaer), doi: . /aetic. . . , available: http://aetic.theiaer.org/archive/v /v n /p .html. research article fuzzy-based adaptive framework for module advising expert system obada alhabashneh mutah university, karak, jordan o.alhabashneh@mutah.edu.jo received: rd november ; accepted: nd december ; published: st january abstract: in the enrolment process, selecting the right module and lecturer is very important for students. the wrong choice may put them in a situation where they may fail the module. this could lead to a more complicated situation, such as receiving an academic warning, being de-graded, as well as withdrawn from the program or the university. however, module advising is time-consuming and requires knowledge of the university legislation, program requirements, modules available, lecturers, modules, and the student's case. therefore, the creation of effective and efficient systems and tools to support the process is highly needed. this paper discusses the development of a fuzzy-based framework for the expert recommender system for module advising. the proposed framework builds three main spaces which are: student-space (ss), module- space (ms), and lecturer-space (ls). these spaces are used to estimate the risk level associated with each student, module, and lecturer. the framework then associates each abnormal student case in the students’ grade history with the estimated risk level in the ss, ms, and ls involved in that particular case. the fuzzy- based association-rule learning is then used to extract the dominant rules that classify the consequent situation for each eligible module if it is to be taken by the student for a specific semester. the proposed framework was developed and tested using real-life university data which included student enrollment records and student grade records. a five-fold cross-validation process was used for testing and validating the classifying accuracy of the fuzzy rule base. the fuzzy rule base achieved a % accuracy level in classifying the risk level for enrolling on a specific module for a specific student case. however, the average classifying accuracy achieved was . % which is acceptable for this problem domain as it involves human behavior modeling and decision making. keywords: intelligent academic advisor; module adviser; expert system; fuzzy logic; fuzzy rule-based. . introduction every academic semester, students enroll in different modules offered by their universities. enrolment might sound like a simple process, but it involves many complications and problems for students particularly in these universities where the credit hour system is adopted. according to this system, although there are some constraints such as the academic plan, the students still have the freedom to select the sections to enroll in based on their preferences including the module, timeslot, and lecturer of the offered sections in the timetable. however, the lack of knowledge about the modules and the lecturers, coupled with the contradictory advice they receive from their colleagues and the complexity of the academic plan of their programs may affect the students' ability to select the right choice. this creates a need for professional advice as taking a wrong choice may lead to unwanted situations such as failing the module, or more seriously, academic warning, program withdrawal or even leaving the university [ - ]. to address this problem, universities usually have academic advisors to help the students in selecting their modules and sections for the upcoming semester based on different factors including http://aetic.theiaer.org/ http://aetic.theiaer.org/ http://www.theiaer.org/index.htm http://aetic.theiaer.org/archive/v /v n /p .html mailto:o.alhabashneh@mutah.edu.jo aetic , vol. , no. www.aetic.theiaer.org the student’s cumulative grade point average (gpa), the available modules, timeslots, module’s lecturer, and the academic plan of the student. however, module advising is a challenging task as it is complicated, time-consuming, and needs qualified and knowledgeable human resources. this urges the need for developing intelligent systems for module-advising to aid the students in selecting the right modules at the right time to ensure a smooth passage for students throughout their academic plans to finish their degrees with fewer problems. intelligent recommender systems (rss) are widely used in different application areas including online shopping, movies, social networking, and others. however, they are less common in education and this area is still under development [ - ]. rss can provide the students with personalized recommendations on suitable modules based on the student’s academic history. such systems can also incorporate previous student's experiences to provide an improved recommendation. furthermore, they adapt to the changes in the data and give the recommendation based on that. the need for rss and their requirements in the education sector has been discussed in [ - ]. more recently, the use of rss in education has received more attention and their potential has been proven [ , ]. this paper discusses the development of a fuzzy-based framework to be used in recommender systems for module advising. the proposed framework builds the recommendation based on creating three main spaces which are: student-space (ss), module-space (ms), and lecturer-space (ls). these spaces are used to estimate the risk level associated with each student, module, and lecturer. the framework then associates each abnormal student case in the students’ grade history with the estimated risk level in the ss, ms, and ls involved in that particular case. fuzzy-based association- rule learning [ , ] is then used to extract and summarize a fuzzy rule base. the fuzzy rule base is used to predict the risk level associated with the combination of a specific student, module, and lecture and build the recommendation to students based on that. the proposed approach provides a novel mechanism to estimate the consequent risk associated with the student selection of a specific module that is taught by a specific lecturer. the risk estimation is based on creating and analyzing the three space elements. . related work . . course/module advising in higher education academic advising is vital for a good academic experience of students as it positively impacts their success and retention [ , ]. however, it requires specific knowledge of the student's situation, history, program’s academic plan, modules, and lecturers [ ]. gordon et al. [ ], defined academic advising as “situations in which an institutional representative gives insight or direction to a college student about an academic, social, or personal matter”. module advising is one of the main tasks of the academic advisor, which refers to the process by which the academic advisor supports the student in selecting the right modules in which to enroll. having the importance and the complexity of module advising, different researchers have argued that there is a need for developing intelligent systems to support this task [ , , , ]. developing such systems aims to minimize the demand for human advisors and gives them more time to focus on other important advising tasks including career development and financial issues. [ - ]. a novel approach for module long term plaining called interactive decision support for course planning (idisc+) was proposed by mohamed [ ]. the approach used optimization techniques to support both the student and the advisor in building a recommended long-term academic plan (towards graduation) to be followed by the student. laghari [ ] developed an automated course advising system (acas) for module planning. the system distributes the modules of the academic plan on different semesters based on the history of other students. there has been significant research effort in applying information technology for module advising. roushan et al. [ ] introduced an internet-based approach to support the module-advising process which integrated the process of advising with the enrolment taking into account the constraints of the program plan and the university policy. however, the system did not replace the human academic-advisor but rather facilitated the advisor role by providing an automated tool for aetic , vol. , no. www.aetic.theiaer.org communication and information exchange. a decision-support system was proposed to support academic advisors in preparing a pre-enrolment plan for the students and assist in the offering of appropriate modules for the upcoming semester [ ]. al-ghamdi et al.[ ] developed an advisor expert system (pas) for postgraduate students. the proposed system was designed to assist them in the selection of the most relevant modules without referring to a human module-advisor. a web- based advising system [ ] was proposed which supports three types of users, including students, advisors, and heads of department to make sure that a complete picture is available for the students. mattei et al.[ ] developed a decision-theory advising tool to enhance the advisor-student relationship. the tool allowed students to browse the module offerings, possible future scenarios, and their outcomes. shatnawi et al. [ ] used the enrolment and marking history from similar cases and applied an association rule-based system to give general recommendations when selecting the modules on which to enroll. . . content-based different researchers have used a content-based academic recommendation system for module selection. for example, lin et al.[ ] utilized a multi-agent approach and ontology to provide a dynamic and personalized recommended module list. the multi-agent approach, which included various agents, used a preference-driven planning algorithm supported by the ontology to build the recommendations. darimola et al.[ ] integrated case-based reasoning (cbr) with rule-based reasoning (rbr) techniques to provide an intelligent approach for module advising. the approach also used historical data to build a list of recommended modules for the following semester. . . collaborative filtering collaborative filtering was also used for academic advising. huang et al. [ ] proposed a score- based prediction approach for course recommendation. the approach used a user domain collaborative filtering to create the recommended course list. courses were clustered based on the student feedback in [ ] and the resulted clusters were combined with fuzzy based rule association technique to create the course recommendation. nafea et al. [ ] developed a learning-style-based collaborative-filtering approach for module recommendations which utilized different metrics to identify similar profiles including k-means, cosine, and person correlation. chang et al. [ ] proposed a user-based collaborative-filtering approach to predict student grades. mortenson et al. [ ] introduced a collaborative filtering approach for module selection which utilizes an artificial immune mechanism for the prediction. bydžovská [ ] investigated the effect of student and module features on the enrolment patterns and designed a collaborative-filtering-based system to predict the module grade. yao et al. [ ] attempted to increase the fairness of the module recommendation by addressing the biased-recommendation problem against minority groups of students. they developed four different fairness metrics that can be optimized using the learning objectives . . knowledge-based different knowledge-based recommendation approaches have been proposed for module selection. xu et al. [ ] developed a knowledge-based approach to offer a personalized module sequence to the new students. this approach utilized a dynamic learning algorithm that learns from the performance of other students on a specific module. koutrika [ ] argued that recommendation methods should not be 'hard-wired', but it should be flexible. in that sense, a new paradigm for the recommendation was introduced in which a recommendation approach can be defined declaratively as a high-level parameterized workflow comprising traditional relational operators and new operators that generate or combine recommendations. keston et al. [ ] utilized semantic web expert system technologies to build a knowledge base that is used by an intelligent web-based application to provide the required recommendations. engin et al. [ ] developed an expert rule-based system for module recommendation. the rules were captured from the real advisors and then injected into a rule-base using oracle policy automation (opa). aetic , vol. , no. www.aetic.theiaer.org hashemi and blondin [ ] included several factors to be taken into consideration when recommending modules for students such as the frequency of the module offering, balancing the module load, and shortening the graduation path. all these factors and others were included in a rule base which was used for recommending modules. ayman [ ] proposed an expert system for module selection which included both prescriptive and developmental advising models and utilized object- oriented databases for data and rule representation. abdullah al-ghamdi et al. [ ] proposed a rule- based advising system. the system was designed for postgraduate students and built recommendations to support the students in the selection of modules related to the topic of their research thesis. nambia and dutta [ ] introduced a dynamic and flexible rule-based advising system in which the rules were separated from the execution to enable the student to try different scenarios by updating the xml file where the rules are stored. nguyen et al. [ ] proposed a knowledge-based framework that utilized a learning data-warehouse for discovering patterns in the student behavior including module selection and achievements. these patterns were then used to make the recommendations. . . hybrid other researchers proposed hybrid approaches in which more than different perspectives were integrated. daramola. et al. [ ] designed an intelligent expert system for module advising which integrated rule-based reasoning (rbr) with case-based reasoning (cbr) using the academic history of the students. sobecki [ ] applied ant colony optimization (aco) to provide an efficient module advisor system. the system predicted the final grade of the students in a module, based on a domain- specific representation integrated with the aco. abdulwahhab [ ] integrated genetic algorithms (gas) and a decision tree for short-term module-scheduling. . . fuzzy based a few researchers have used fuzzy logic to develop module advising systems. goodarzi and rafe [ ] developed a fuzzy-based expert system for student advising. the proposed system was a web-based module that can be integrated with the university portal. the module fuzzifies the business rules and the gpa of the students to advise them on which modules should be taken in the following term. adak [ ] used fuzzy techniques to recommend elective modules to students. it analyzed transcripts of students to extract fuzzy rules to relate the module to the students who intend to take them. baloul and williams [ ] used the order of preference by similarity to ideal solution (topsis) technique to develop a fuzzy model for probation students to minimize the risk of taking the wrong modules in the early stage of their study. . the proposed approach this paper discusses the development of a fuzzy-based framework of an expert system for module advising. the proposed framework assumes that recommendations are based on three main elements which are: student-space (ss), module-space (ms), and lecturer-space (ls). ss contains the current student status which includes their accumulative average (if there is any academic warning), the closest abnormal academic status and to which extent it's close to that status. it also contains the knowledge domains of the academic program and the progress of the students in each domain. ms contains the average module mark for the last five years and the knowledge domains that the module belongs to. ls contains the average of the marks for the lecturer for all modules and the average of the lecturer's marks for each module. the three spaces and the result of the final calculation are then combined in a matrix called the case-space (cs). fuzzy based association-rule learning is then used to extract the dominant rules to classify the consequent case for each eligible module if it was taken by the student for a specific semester. the fuzzy logic is used to handle the uncertainty involved in modeling the human decision and to provide a transparent and interpretable mechanism module taking risk estimation. aetic , vol. , no. www.aetic.theiaer.org the main purpose of the proposed framework is to estimate the consequent risk level (low, medium, and high) taking a specific module. the risk level is assigned based on a list of unwanted cases associated with the student failing or not progressing in the module. these cases include a gpa decrease, moving down from one gpa category to a lower one (degrading), receiving an academic warning, withdrawn from the academic program, withdrawn from the university, and graduation pending. in addition to the risk-level estimation, the framework provides the student with a justification (interpretation) for the estimated risk-level based on the student's situation, the targeted module, and the lecturer. the framework was developed based on real-life university data which included a historical record of enrolment associated, student's marks, cumulative gpa, module offerings, academic plans for the programs, and finally the knowledge domains for both the programs and the modules. the proposed approach is depicted in figure figure : the proposed framework fig. shows the use case diagram of the proposed system. figure : system use case diagram aetic , vol. , no. www.aetic.theiaer.org the proposed approach consists of six main steps: . . step- : creating the three spaces and abnormal case matrix in this step, the student-space (ss), module-space (ms), lecturer-space (ls), and abnormal-case matrix are created as follows: . . . step- .a: creating the three spaces the three spaces ss, ms, and ls are created as shown in fig. error! reference source not found.. the spaces are extracted from the university database. figure : the three spaces and abnormal case matrix student space: this space contains information about the student as shown in table . table : student space element description . student id a unique identification number for the student . cumulative gpa the accumulative average of the student out of . student average mark for the program knowledge domains this is a matrix that includes the knowledge domains of the student academic program and the average mark of the student for the modules which belong to each domain. . potential risk situation this refers to the approximate abnormal student situation for the current situation of the student. (i.e., the current abnormal situation of the student boundary between two gpa categories, such as “very good” and “satisfactory”). each abnormal-situation type is given a code and a percent of to indicate the level of risk which is given to the code. . current student situation this refers to the student's current abnormal situation. i.e., if the current abnormal situation of the student boundary between two gpa categories, such as “very good” and “satisfactory”, or the student's gpa has dropped. each abnormal-situation type is given a code and a percent of to indicate the level of risk which is given to the code. module space: this space contains information about the module as shown in table . table : module space element description . module id a unique identification number of the module . general average of grades the average grades for the students in the module . average grade for each program a matrix includes the academic programs which contain the module and the average module grade for each program . average grade for each lecturer who taught the module a matrix which includes the lecturers who taught the module and the average grade for the module given by each lecturer . fail rate the percentage of students who failed the module aetic , vol. , no. www.aetic.theiaer.org lecturer space: this space contains information about the module as shown in table . table : lecturer space element description . lecturer id a unique identification number of the lecturer . general average of grades the average grade of the lecturer . fail rate the percentage of students who failed the modules taught by the lecturer. . . . step- .b: creating abnormal case matrix in this sub-step, the abnormal-case matrix is created by selecting the student id, module id, lecturer id, and the student abnormal situation for abnormal cases in the student enrolment records from the university database. the abnormal case is usually indicated by a flag or a symbol in the database as shown in figure . figure : abnormal-case matrix with case mapping . . step- : case risk analysis and calculation for each problematic case in the dataset, a risk weight is calculated for the three spaces: ss, ms, and ls and the associated abnormal situation. the product of this step is shown in fig. . the risk was calculated for the three spaces using equations developed based on common sense and the opinion of field experts such as the academic registry members, lectures, and head of departments. the results than have been presented to the experts and evaluated by them. figure : case-risk analysis and calculation risk estimation of student-space: 𝑺𝑹(𝑺𝑺) = (𝟏𝟎𝟎 − ( 𝑮𝑷𝑨(𝑺𝑰𝑫) + 𝑲𝑫𝑨(𝑴𝑰𝑫,𝑺𝑰𝑫) 𝟐 ) + 𝑷𝑨𝑪 + 𝑪𝑨𝑪)/𝟑 equation : risk estimation of student-space where sr is the risk estimation of the student space in relation to the current case. it takes the student space as input and returns a percentage value out of . ss is the student space. gpa is the student cumulative gpa, sid is the student id, kda is the student average mark of the knowledge domain to which the module belongs. pac is the scaled weight (out ) of the potential abnormal situation. cac is the scaled weight (out ) of the current abnormal situation of the student. risk estimation of module-space: aetic , vol. , no. www.aetic.theiaer.org 𝑴𝑹(𝑴𝑺) = (𝟏𝟎𝟎 − 𝑴𝑨𝑮(𝑴𝑰𝑫) + 𝑴𝑷𝑨(𝑴𝑰𝑫,𝑷𝑰𝑫) + 𝑴𝑳𝑨(𝑴𝑰𝑫,𝑳𝑰𝑫) 𝟑 + 𝑴𝑭𝑹(𝑴𝑰𝑫))/𝟐 equation : risk estimation of module-space where mr is the risk estimation of the module in relation to the current student situation. it takes the module space (ms) as an input and returns a percentage out of , ms is the module space. mag is the general average mark of the module. mid is the module identification number; mpa is the average module mark for the student's program. pid is the identification number of the student's academic program. lid is the lecturer identification number. cai is the average mark of the module for the lecturer. mfr is the module fail rate which is a percentage of the students who failed the module. risk estimation of lecturer-pace: 𝑳𝑹(𝑳𝑺) = ((𝟏𝟎𝟎 − 𝑳𝑨𝑮(𝑳𝑰𝑫)) + 𝑳𝑭𝑹(𝑳𝑰𝑫))/𝟐 equation : risk estimation of lecturer space where lr is the risk estimation that comes from the lecturer side out , ls is the lecturer-space. lag is the general average mark of the lecturer. lid is the lecturer-identification number. lfr is the lecturer fail rate which is a percentage of the students who failed the modules taught by the lecturer. situation risk (sr) estimation: the resulting situation types were provided to a set of experts (i.e., academic-registry staff, advisors, and heads of department) and they were asked to assign a specific weight which reflects the risk level. the risk level takes a value between and . the net risk for each situation type is calculated by taking the average of the weights given by the experts. . . step- : linguistic labelling in this step, the final risk weights of the three spaces and the resulting situation are given linguistic labels of h-high, m-medium, and l-low. these linguistic labels are generated using type- and based on predefined fuzzy sets based on the mendel wang method [ ]. in the proposed approach, the member function of the fuzzy sets uses a triangle shape as shown in figure and is based on three parameters {a, b, c} as shown in equation . 𝑇𝑟𝑖𝑎𝑛𝑔𝑙𝑒(𝑥;𝑎;𝑏;𝑐) = { , 𝑥 ≤ 𝑎. 𝑥 − 𝑎 𝑏 − 𝑎 , 𝑎 ≤ 𝑥 ≤ 𝑏. 𝑐 − 𝑥 𝑐 − 𝑏 , 𝑏 ≤ 𝑥 ≤ 𝑐. , 𝑐 ≤ 𝑥. equation : triangle fuzzy set where the parameters {a, b, c} (with a then) where the linguistic labels of the three spaces are the antecedents and the linguistic label of the resulting situation is the consequent as shown in figure . ss ms ls result ↓ ↓ ↓ ↓ h h h → h antecedents consequent figure : rule base creating . . step- : fuzzy rule-base compression and validation the rule base extracted in the previous step could contain many rules based on the size of the dataset. it may have a large amount of repetition (i.e., repeated rules) and contradiction. the contradiction here means rules that have the same antecedent with different consequents such as shown in the following example in table : table : contradictory rule patterns r(ss) (ms is result ↓ ↓ ↓ ↓ h h h → h h h h → l h h h → m to address these issues a five-fold cross-validation [ ] process was used to train the rule base to summarize it to the most dominant unique patterns. in this validation method the dataset is divided into equal-size sets (d , .., d ). for each fold the following steps are applied: . one of the subsets dn is selected as a testing set tn and the other subsets are groups in one training set d^. . the rule compression(summarization) technique [ ] is applied to the training set to produce a summarized rule-base . the summarized rule-base is applied to the test set to predict the risk level. . the predicted risk levels are compared with the actual risk levels from the dataset, to identify the accuracy of the rule-base in predicting the risk level the compression technique: this technique uses two measures for rule quality, which are “generality” and “reliability” and are used to identify rule patterns with maximum firing strength. generality measures the number of instances in the extracted rule base which supports each rule pattern. reliability measures the confidence level of each rule pattern [ ]. aetic , vol. , no. www.aetic.theiaer.org in the proposed approach, generality is calculated using scaled fuzzy support, and the reliability is calculated by multiplying the scaled fuzzy support by the firing strength of the rule pattern. the support of the rule pattern refers to the number of the rules in the rule-base that the pattern represents. the “confidence” refers to the strength of a specific rule pattern against the contradictory patterns (i.e. because other rule patterns have the same antecedent and a different consequent) [ ]. fuzzy support is used to identify the unique rule-patterns with their occurrence in the extracted rule base. the fuzzy support can be scaled for each unique rule-pattern using the total number of instances in the rule-base which have the same consequent. equation shows how the scaled fuzzy support for a unique rule pattern is calculated: 𝑠𝑐𝐹𝑢𝑧𝑧𝑆𝑢𝑝(𝑅𝑃𝑙) = 𝑵𝑹𝑷𝒍 𝑵𝑹𝑷𝒍 + 𝑵𝑹�̂�𝒍 equation : fuzzy support where l = to m, l is the index of the rule pattern, rpi is a unique rule pattern i.e. (h, h, h →m), 𝑵𝑹𝑷𝒍 is the number of instances in the extracted rule base supporting the rule pattern rpi. 𝑵𝑹�̂�𝒍is the number of instances in the extracted rule base which support other patterns with the same consequent i.e. ({m, h, h →m}, {m, m, h →m},…). the “confidence” is a measure of the uniqueness of the pattern as it indicates its strength against the contradictory patterns, which are the other patterns having the same antecedents but a different consequent. equation shows how confidence is calculated 𝑠𝑐𝐶𝑜𝑛𝑓(𝑅𝑃𝑙) = 𝑠𝑐𝐹𝑢𝑧𝑧𝑆𝑢𝑝(𝑹𝑷𝒍) 𝐶𝑜𝑹𝑷𝒍 equation : confidence where 𝐶𝑜𝑹𝑷𝒍 is the number of instances in the extracted rule base which supports the contradictory rule- patterns with 𝑅𝑃𝑙. the final scaled-weight of the rule pattern is calculated as the product of the fuzzy support and the fuzzy confidence as shown in equation . 𝑠𝑐𝑊𝑖 = 𝑠𝑐𝐹𝑢𝑧𝑧𝑆𝑢𝑝 × 𝑠𝑐𝐶𝑜𝑛𝑓 equation : final scaled weight each of the generated unique rule patterns is assigned the scaled fuzzy weight measure scwi as follows: table : the scaled fuzzy weight of the unique rule-patterns ss ms ls result scwi ↓ ↓ ↓ ↓ h h h → h . h h h → l . h h h → m . the scaled fuzzy weight of the unique rule patterns is then used to select the rules with the highest weights among the contradictory-patterns set. the result of the compression process is a summarized rule-base that contains dominants and consistent rule patterns. the resulting rule base will be used later to estimate the risk level considering a specific module, lecturer, and student. . . step- , compressed rule base selection the accuracy levels of the five compressed rule-bases resulting from the previous step are compared and the compressed rule-base with the highest accuracy level was selected to be included in the system. . experiment and results . . dataset a real-life university data was used as a dataset for this paper. the dataset included student- enrolment records, student-mark records, module records, academic program records, and records for lecturers. the dataset consisted of the records of students who faced problems during their aetic , vol. , no. www.aetic.theiaer.org studies. these problems included academic warnings, program withdrawals, program changes, and cumulative average grade down-gradings. . . creating the three spaces (ss, ms, and ls) the university uses an oracle database for the student information system (sis), which was used to create a view for each of the three space types. a view was also created to include the risk estimation for each space associated with the risk estimation provided by the expert. . . creating the fuzzy rule-base to create the fuzzy rule base, matlab fuzzy logic tool kit was used to create the fuzzy sets, which were then used to create the linguistic labels for the data extracted from the three spaces. . . fuzzy rule-base training and compression a five-fold cross-validation technique was used to train the rule-base, then, for each fold, the fuzzy rule-base was divided into two different subsets, which were training and testing. the %- % rule was used to identify the size of each set. the compression technique discussed in section ( ) was applied on the training set to produce the compressed rule-base. the resulting compressed rule- base was applied to the test set to predict the risk level. the predicted risk-levels were compared with the actual risk-levels provided by the experts to determine the accuracy of the rule-base in predicting the risk level as shown in table . table : rule base classifying accuracy sample passed? predicted actual situation type l_risk l_id m_risk m_id s_risk s_id l l a m h m **** m h d l h h **** h h e h m h **** h h d h l l **** m m b m h m **** m m f m l m **** h h d h m h **** m h d l h h **** the accuracy level of the test results for each fold is shown in table . as shown in the table the proposed approach achieved % classifying accuracy and . as an average classifying accuracy. the comparison between the proposed approach and the other approaches is not a straightforward task as the main focus of the proposed approach differs from those of the others. the proposed approach aims to address the risk estimation while some others focused on long-term planning or supported the process of the academic advisory as a whole. however, the results can still be compared to indicate the performance of the proposed approach. for example, the multi-agent system proposed in [ ] achieved only % user satisfaction with its effectiveness. besides, the user satisfaction with the intelligent advising system proposed in [ ] was . % which indicates at the performance proposed system in this paper is acceptable. table : -fold accuracy test results fold number of rules accuracy % % % % % average . % aetic , vol. , no. www.aetic.theiaer.org . . best compressed rule base selection the accuracy levels of the five compressed rules bases resulting from the previous step were compared and the compressed rule-base with the highest accuracy level was selected to be included in the system as shown in table . table : best rule set i ss ms is result scwi ↓ ↓ ↓ ↓ h h h → h . h h m → h . h h l → m . h m h → h . h m m → h . h m l → m . h l h → h . h l m → m . m l l → l . m h h → h . m h m → m . m h l → m . m m h → m . m m m → m . m m l → m . m l h → m . m l m → m . l l l → l . l h h → h . l h m → m . l h l → m . l m h → m . l m m → m . l m l → l . l l h → l . l l m → l . l l l → l . . conclusion this paper has presented a fuzzy-logic based approach of an expert system for module advising. the approach creates three spaces which are student-space, module-space, and lecturer space, and associate them with the abnormal situation type for each abnormal case. each space and the abnormal situation type is then given a linguistic label using fuzzy sets. the linguistic labels are then associated with each other for each case to generate a fuzzy rule. the fuzzy rules for all cases are combined in one rule-base and then compressed to extract those rules with the highest firing strength. the fuzzy logic was used to handle the uncertainty implied in the human judgment of the student case as well as to provide a transparent and interpretable mechanism for predicting the risk level of enrolling a student on a module. the approach was developed using real-life university data and achieved an acceptable level of accuracy of % which is expected to improve as more data is captured and used to train the rule base. although the achieved accuracy might sound like it needs a bit of enhancement but having that the machine learning approach used for training the rule base the accuracy is expected to enhance as more data instances are included. also, the accuracy level acceptance depends on the problem and context, especially with cases in which are molding human decisions or behavior. although there are different approaches have been proposed in this area, this paper introduces a novel mechanism that creates three spaces to estimate the risk of the student situation associated with a specific module. the three spaces which are namely: students, lecturer, and module provide a multi-angle view on the student case and makes the estimation more realistic. also, applying the fuzzy logic provides a means to handle the uncertainty included in the human decision-making regarding module selection. aetic , vol. , no. www.aetic.theiaer.org the fuzzy rule base also provides a transparent mechanism to make the recommendation which means it doesn’t provide the risk level only but also the justification of that recommendation based on the risk level of the three spaces. references [ ]. lam, s. s. and choi, s. p. m. ( ). implementing an efficient preference-based academic advising system. international journal of applied management science, ( ), pp – , doi: . /ijams. . . [ ]. almutawah, k. a. ( ). a decision support system for academic advisors. international journal of business information systems, ( ), pp . doi . /ijbis. . . [ ]. daramola, o., emebo, o., afolabi, i. and ayo, c. ( ). implementation of an intelligent course advisory expert system cased-based course advisory expert system. in (ijarai) international journal of advanced research in artificial intelligence, ( ), available: www.ijarai.thesai.org [ ]. bendakir, n. and aïmeur, e. ( ). using association rules for course recommendation. aaai workshop - technical report, vol. ws- - , pp – , available: https://www.aaai.org/papers/workshops/ /ws- - /ws - - .pdf [ ]. o’mahony, m. p. and smyth, b. ( ). a recommender system for on-line course enrolment: an initial study. recsys’ : proceedings of the acm conference on recommender systems, pp – . doi: . / . . [ ]. sandvig, j. and burke, r. ( ). aacorn: a cbr recommender for academic advising. technical report tr - , available: http://facweb.cs.depaul.edu/research/techreports/tr - .doc. [ ]. ajanovski, v. v. ( ). guided exploration of the domain space of study programs. in th joint workshop on interfaces and human decision making for recommender systems (intrs), available: http://ceur- ws.org/vol- /paper .pdf. [ ]. young, p. ( ). a recommender system for personalized exploration of majors, minors, and concentrations *, recsys poster proceedings ( ), available: http://ceur-ws.org/vol- /recsys _poster .pdf. [ ]. sharma, a. and tiwari, n. ( ). design and analysis of fuzzy based association rule mining. international journal of computer applications and information technology, (i), pp , available: https://www.ijcait.com/ijcait/ / .pdf. [ ]. wang, l.-x. x. and mendel, j. m. ( ). generating fuzzy rules by learning from examples. ieee transactions on systems, man and cybernetics, ( ), pp – . doi: . / . . [ ]. young-jones, a. d., burt, t. d., dixon, s. and hawthorne, m. j. ( ). academic advising: does it really impact student success? in quality assurance in education. vol. , available: www.emeraldinsight.com. [ ]. mohamed, a. ( ). a decision support model for long-term course planning. decision support systems, vol. , pp – . doi: . /j.dss. . . . [ ]. harding, b. ( ). students with specific advising needs. in v. n. gordon, w. r. habley and t. j. grites (eds.), academic advising: a comprehensive handbook ( nd ed.), pp. – , jossey-bass. [ ]. laghari, m. s. ( ). automated course advising system. international journal of machine learning and computing, pp – . doi: . /ijmlc. .v . . [ ]. roushan, t., chaki, d., hasdak, o., chowdhury, m. s., rasel, a. a., rahman, m. a. and arif, h. ( ). university course advising: overcoming the challenges using decision support system. th int’l conf. computer and information technology, iccit , pp – . doi: . /iccitechn. . . [ ]. talal al-nory, m. ( ). simple decision support tool for university academic advising. proceedings of international symposium on information technologies in medicine and education, itme , vol. , pp – . doi: . /itime. . . [ ]. abdullah, a.-g., sumaia, a.-g., fadel, a., al-ruhaili, f. and thamary, a.-a. ( ). an expert system for advising postgraduate students. international journal of computer science and information technology, ( ), pp – . [ ]. albalooshi, f. and shatnawi, s. ( ). he-advisor: a multidisciplinary web-based higher education advisory system. in global journal of computer science and technology, ( ), pp. - . [ ]. mattei, n., dodson, t., guerin, j. t., goldsmith, j. and mazur, j. m. ( ). lessons learned from development of a software tool to support academic advising. in proceedings of the zone conference of the american society for engineering education - “engineering education: industry involvement and interdisciplinary trends”, asee zone . ieee computer society. doi: . /aseezone . . . [ ]. shatnawi, r., althebyan, q., ghalib, b. and al-maolegi, m. ( ). building a smart academic advising system using association rule mining. arxiv, available: http://arxiv.org/abs/ . aetic , vol. , no. www.aetic.theiaer.org [ ]. lin, f., dunwei, s. l., frank, w., kinshuk, z. and mcgreal, r. ( ). e-advisor: a multi-agent system for academic advising, available: http://io.acad.athabascau.ca/~oscarl/pub/abshl .pdf [ ]. huang, l., wang, c. d., chao, h. y., lai, j. h. and yu, p. s. ( ). a score prediction approach for optional course recommendation via cross-user-domain collaborative filtering. ieee access, vol. , pp – , doi: . /access. . . [ ]. asadi, s. and shokrollahi, z. ( ). developing a course recommender by combining clustering and fuzzy association rules. journal of ai and data mining, ( ), pp. – . doi: . /jadm. . . . [ ]. nafea, s. m., siewe, f. and he, y. ( ). a novel algorithm for course learning object recommendation based on student learning styles. proceedings of international conference on innovative trends in computer engineering, itce , pp – , doi: . /itce. . . [ ]. chang, p. c., lin, c. h. and chen, m. h. ( ). a hybrid course recommendation system by integrating collaborative filtering and artificial immune systems. algorithms, ( ), doi: . /a . [ ]. newell, c. and miller, l. ( ). design and evaluation of a client-side recommender system. recsys - proceedings of the th acm conference on recommender systems, pp – . doi: . / . . [ ]. bydžovská, h. ( ). are collaborative filtering methods suitable for student performance prediction? lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics), vol. , pp. – . doi: . / - - - - _ . [ ]. yao, s. and huang, b. ( ). beyond parity: fairness objectives for collaborative filtering. advances in neural information processing systems, vol. -decem, pp. – , http://arxiv.org/abs/ . . [ ]. xu, j., xing, t. and van der schaar, m. ( ). personalized course sequence recommendations. ieee transactions on signal processing, ( ), pp. - , doi: . /tsp. . . [ ]. koutrika, g., bercovitz, b. and garcia-molina, h. ( ). flexrecs: expressing and combining flexible recommendations. sigmod-pods’ - proceedings of the international conference on management of data and th symposium on principles of database systems, pp – . doi: . / . . [ ]. keston, l. and goodridge, w. ( ). adviseme: an intelligent web-based application for academic advising. international journal of advanced computer science and applications, ( ). doi: . /ijacsa. . . [ ]. engin, g., aksoyer, b., avdagic, m., bozanli, d., hanay, u., maden, d. and ertek, g. ( ). rule-based expert systems for supporting university students. procedia computer science, vol. , pp – . doi: . /j.procs. . . . [ ]. hashemi, r. r. and blondin, j. ( ). sassy: a petri net based student-driven advising support system. itng - th international conference on information technology: new generations, pp – . doi: . /itng. . . [ ]. ayman, m. ( ). a prototype student advising expert system supported with an object-oriented database. international journal of advanced computer science and applications, ( ). doi: . /specialissue. . . [ ]. nambiar, a. n. and dutta, a. k. ( ). expert system for student advising using jess. iceit - international conference on educational and information technology, proceedings, vol . doi: . /iceit. . . [ ]. thanh binh, n., thi anh duong, h., hieu, t., duc nhuan, n. and hong son, n. ( ). an integrated approach for an academic advising system in adaptive credit-based learning environment. vnu journal of science, natural sciences and technology, vol. , pp – , avaiable: https://repository.vnu.edu.vn/bitstream/ / / /tc_ .pdf. [ ]. sobecki, j. and tomczak, j. m. ( ). student courses recommendation using ant colony optimization. lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics), vol. lnai(part ), pp – , doi: . / - - - - _ . [ ]. abdulwahhab, r. s., al makhmari, h. s. and al battashi, s. n. ( , march ). an educational web application for academic advising. ieee th gcc conference and exhibition, gccce , doi: . /ieeegcc. . . [ ]. goodarzi, m. h. and rafe, v. ( ). educational advisor system implemented by web-based fuzzy expert systems. journal of software engineering and applications, ( ), pp – , doi: . /jsea. . . [ ]. adak, m. f., yumusak, n. and taskin, h. ( , april ). an elective course suggestion system developed in computer engineering department using fuzzy logic. international conference on industrial informatics and computer systems, ciics . doi: . /iccsii. . . aetic , vol. , no. www.aetic.theiaer.org [ ]. baloul, f. m. and williams, p. ( ). fuzzy academic advising system for on probation students in colleges of applied sciences - omn. proceedings - international conference on computer, electrical and electronics engineering: “research makes a difference”, icceee , pp. – , doi: . /icceee. . . [ ]. moss, h. b., leslie, d. s. and rayson, p. ( ). using j-k fold cross validation to reduce variance when tuning nlp models. arxiv, vol. , pp – , available: http://arxiv.org/abs/ . . [ ]. wu, d., mendel, j. m. and joo, j. ( ). linguistic summarization using if-then rules. ieee world congress on computational intelligence, wcci , doi: . /fuzzy. . . © by the author(s). published by annals of emerging technologies in computing (aetic), under the terms and conditions of the creative commons attribution (cc by) license which can be accessed at http://creativecommons.org/licenses/by/ . . context-aware systems: a literature review and classification expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a context-aware systems: a literature review and classification jong-yi hong *, eui-ho suh, sung-jin kim posmis lab, industrial management engineering building, pohang university of science technology, ( - ) san , hyoja-dong, nam-gu, pohang, kyungbuk, south korea a r t i c l e i n f o keywords: context-aware systems classification framework literature reviews - /$ - see front matter crown copyright � doi: . /j.eswa. . . * corresponding author. tel.: + ; fax e-mail addresses: xman@postech.ac.kr (j.-y. hong suh), sciv@postech.ac.kr (s.-j. kim). a b s t r a c t nowadays, numerous journals and conferences have published articles related to context-aware systems, indicating many researchers’ interest. therefore, the goal of this paper is to review the works that were published in journals, suggest a new classification framework of context-aware systems, and explore each feature of classification framework. this paper is based on a literature review of context-aware sys- tems from to using a keyword index and article title search. the classification framework is developed based on the architecture of context-aware systems, which consists of the following five lay- ers: concept and research layer, network layer, middleware layer, application layer and user infrastruc- ture layer. the articles are categorized based on the classification framework. this paper allows researchers to extract several lessons learned that are important for the implementation of context- aware systems. crown copyright � published by elsevier ltd. all rights reserved. . introduction emerging ubiquitous or pervasive computing technologies offer ‘anytime, anywhere, anyone’ computing by decoupling users from devices (dey, ; hill et al., ; kwon, choi, & park, ; kwon, yoo, & suh, ; schilit, adams, & want, ). to provide adequate service for the users, applications and services should be aware of their contexts and automatically adapt to their changing contexts-known as context-awareness (bolchini, schreiber, & tan- ca, ; dey, ; zhu, mutka, & ni, ). context is very important, since it provides information about the present status of people, places, things and devices in the environment (korpipää, mäntyjärvi, kela, keränen, & malm, ; kwon, ). context is any information that can be used to characterize the situation of an entity. an entity is a person, place, or object that is considered rel- evant to the interaction between a user and an application, includ- ing location, time, activities, and the preferences of each entity (dey, ). context-awareness means that one is able to use con- text information. a system is context-aware if it can extract, inter- pret and use context information and adapt its functionality to the current context of use (byun & cheverst, ). the term context- aware computing is commonly understood by those working in context-aware, where it is felt that context is a key in their efforts to disperse and transparently weave computer technology into our lives. one goal of context-aware systems is to acquire and utilize information on the context of a device in order to provide services that are appropriate to the particular people, place, time, event, etc. published by elsevier ltd. all r : + . ), ehsuh@postech.ac.kr (e.-h. these systems aim to provide context-aware access to information, communication and computation. in late s, there was a period of beginning activity on con- text-aware computing. a few of context-aware computing has met the interest. however, the activity seems to be increasing dra- matically. nowadays, to overcome new challenges and require- ments found in context-aware systems, many researchers have made efforts to design and implement network, user infrastructure and middleware which effectively provide users with context- aware services. numerous articles of journals and conferences have published research related to context-aware systems. in other words, many people are interested in context-aware systems. therefore, we feel that this is a good time for a review analysis, since it has been over a year since many papers were published. currently, it is difficult to compare articles, because the available research is published in quite different journals. accordingly, the main objectives of this review are: � to classify and summarize research relevant for context-aware systems. � to provide a conceptual framework for the integration and clas- sification of articles. � to derive suggestions for context-aware systems researchers based on the literature review. chen and kotz ( ) surveyed the literature related with con- text-aware computing in mobile computing. they defined the terms context and context awareness, listed the context-aware applications that have been built, discussed approaches to sense and model the context, and looked into supporting infrastructures, security and privacy issues. however, in , articles published in ights reserved. mailto:xman@postech.ac.kr mailto:ehsuh@postech.ac.kr mailto:sciv@postech.ac.kr http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa potentially related articles: articles retrieved for further review: articles included in this paper: exclusion of articles based on the selection criteria: exclusion of articles based on titles/abstracts/full texts: potentially related articles: articles retrieved for further review: articles included in this paper: exclusion of articles based on the selection criteria: exclusion of articles based on the selection criteria: exclusion of articles based on titles/abstracts/full texts: exclusion of articles based on titles/abstracts/full texts: fig. . the procedures to select the articles. j.-y. hong et al. / expert systems with applications ( ) – journal were not enough to analyze context-aware systems. fur- thermore, a classification standard does not involve all fields of context-aware systems. therefore, a new survey research on con- text-aware systems is needed. baldauf, dustdar, and rosenberg ( ) suggested the abstract layer architecture for context-aware systems and introduced various existent middleware and server- based approaches to ease the development of context-aware appli- cations, where each layer is explained based on various research. however, suggestions for further study, literature review related with each layer, introducing framework of many research and pre- senting current research issues are not broad. the remainder of this paper is structured as follows. in section , a methodology to extract the literature is illustrated. section presents general characteristics of the literature research. section illustrates the proposed classification framework. section de- scribes each feature of classification framework (concept and re- search, network, user infrastructure, middleware, application and service). section contains discussions and suggestions. finally, section concludes the paper with brief concluding remarks. . procedures a total of articles from to were obtained and re- viewed. articles were found via computerized search of the topic areas. the search was narrowed using the terms context-aware. a detail illustration of methodology for extracting articles is followed. . . the selection criteria the selection criteria were used to select and accept the con- text-aware articles. if the papers did not meet the selection criteria, then they were excluded. the three criteria are described as follows: � the literature was based on a search in the descriptor for ‘con- text-aware’. various online journal databases were selected to search context-aware literatures. context-aware articles are found in comprehensive subjects such as computer science, engineering, etc. subjects of online database which are summa- rized according to fields of context-aware systems are shown in table . � this paper surveys the articles published from to . the reason for selecting this time period is that many journals and conferences have published researches related to context-aware computing since . table the online databases and subjects. online database subjects ieee xplore biomedical communication and information communication networks/systems computational intelligence computer engineering computer graphics control systems distributed computing/real-time systems socio-political aspects of technology software design development wiley interscience computer science engineering � this paper covers only journal articles. other publication forms (the conference proceedings, unpublished working papers, mas- ter’s and doctoral dissertations, newspapers and books, etc.) were not included. furthermore, lncs (lecture notes of com- puter science) is excluded, because of numerous articles having contents of computer science. the reason for choosing the jour- nal is that both practitioners and academicians use journals fre- quently to obtain knowledge and spread their study findings. thus, journal articles indicate the highest level of research. . . data sources and procedures to extract articles the papers were selected according to the procedures shown in fig. . first of all, the articles were searched using six online dat- abases. the total number of articles is . the number of articles by each online database is as follows: science direct ( ), ieee xplore ( ), springer link online libraries ( ), acm digital li- brary ( ), wiley interscience ( ), ingenta journals ( ), and ebsco (electronic journal service) ( ). next, articles were excluded because they did not have the word of context-aware systems in the titles and abstracts. next, the articles were carefully reviewed to select those that considered context-aware as the core part. two hundred and thirty seven articles remained because articles did not meet the second selection criteria. based on these procedures, a total of articles met all the selection criteria. online database subjects science direct business, management and accounting computer science decision science engineering psychology social science ebsco (electronic journal service) computer science library science mathematics acm digital library computer science engineering social sciences ingenta journals computer science education engineering springer link online libraries computer science engineering fig. . classification of articles by publication year. table classification of articles based on the journal. journal articles number of articles ieee pervasive computing personal and ubiquitous computing ieee internet computing wireless personal communications ieee intelligent systems mobile networks and applications ieee transactions on software engineering the others expert systems with applications computer communications journal of systems and software pervasive and mobile computing world wide web ieee wireless communications interacting with computers total j.-y. hong et al. / expert systems with applications ( ) – . the general characteristics of the articles . . classification of articles by publication year the number of articles by publication year is depicted in fig. . numerous context-aware articles have grown considerably since . the number of articles in have becomes times more than the number of articles in . the number of articles from to had been increased continuously, and the number of articles from to has been almost the same. it is abso- lute that the concern about context-aware systems was increased and will be continued. . . classification of articles by online database the article by online database is categorized in table . there are a total of articles from online databases. we use online databases to search the articles. in table , ieee xplore is the high- est percentage of articles ( articles, %), because it offers arti- cles of many journals (ieee pervasive computing, ieee internet computing, ieee wireless communications and ieee transactions on software engineering) which have subject relevant to context- aware systems. science direct stores journal articles of various study fields, therefore this database is the second highest of articles ( articles, . %). other online databases are springer link on- line libraries ( articles, . %), ingenta journals ( articles, %), acm digital library ( articles, %), wiley interscience ( articles, %) and ebsco (electronic journal service) ( articles, %). when the same article is found repetitively among different online database, the article number of database having more num- ber of articles is increased. . . classification of articles by journal the articles by journal articles are categorized in table . there were journals that published context-aware articles. most of table classification of articles by publication year. online database number of articles ieee xplore science direct springer link online libraries ingenta journals acm digital library wiley interscience ebsco (electronic journal service) total them were related to the computer science, electronic engineering and information management. table specifies journals that pub- lished four or more context-aware articles and the others that pub- lished only one article or no more than related with context- aware systems are omitted. pervasive computing and ubiquitous computing is closely connected with context-aware computing. therefore, journals which focus on pervasive computing or ubiqui- tous computing have published articles relevant to context-aware systems. as table shows, ieee pervasive computing published the most articles on context-aware systems ( articles, %). ex- pert systems with applications and personal and ubiquitous com- puting published articles ( %). distribution of articles according to the journal shows that various journals have published articles relevant to context-aware systems. . classification framework . . abstract architecture of context-aware systems the abstract architecture of context-aware systems is drawn based on the literature that explores the context-aware prototype, systems, and application to offer classification criteria for dividing the literature appropriately. to make the abstract architecture, so- cam, acai, nama, pecan, x-caf, cyberdesk, webpads, capias, hycon service framework, culliver’s genie, intelligent agent framework, context-aware agent architecture, reference frame- work for multi-target user interfaces, etc. are reviewed in detail. as the results of literature review related with context-aware architecture, we present general abstract layer architecture of con- text-aware systems. as fig. shows, this architecture consists of four layers: ( ) network layer involves a network supporting con- text-aware systems and sensor collecting low-level of context information; ( ) middleware layer manages processes and stores context information; ( ) based below layers, application layer pro- vides users with appropriate service; and ( ) to offer suitable inter- fig. . abstract layer architecture of context-aware systems. j.-y. hong et al. / expert systems with applications ( ) – face to users, interface of context-aware systems is managed in user infrastructure layer. . . classification framework of context-aware systems the classification framework is developed for classifying the lit- eratures related with context-aware systems, based on the abstract architecture of context-aware systems. the classification frame- work developed here consists of the following five layers: concept and research layer, network layer, middleware layer, application layer and user intrastate layer. without the concept and research layer, layers of classification framework and layers of context- aware systems architecture are the same. as fig. shows, each layer has detailed categories for dividing the literatures in layers. fig. depicts a graphical classification framework for context- aware systems. context-aware applications and services require technology from the foundation of wireless network, user infra- structure, and middleware. moreover, application and technology related with network, middleware and user infrastructure are based on concept and research activities. concept and research layer involves overview, algorithm, development guideline, frame- work, context data management, evaluation and privacy and secu- rity categories. network layer consists of protocol, sensing, network requirement and network implementation. middleware layer is classified as agent-based middleware, metadata based mid- dleware, tuple space based middleware, osgi based middleware, reflective middleware and sensor selection middleware. user infra- structure layer is divided into interface and usability categories. application and service layer consists of smart space, tour guide, information systems, communication systems, m-commerce and web service. concept & research framework evaluation cont dat mg overview algorithm network infrastructur protocol network imp leme middleware lay agent based metadata based ada obje ba tup le sp ace based user infrastructure user interface application laye smart sp ace tour guide is commu sys fig. . classification framework . result of article classification the articles by subjects are categorized in tables – . table shows the number of articles, percentage of subject, and percent- age of all subjects, while tables – represent all references of the context-aware systems articles. the context-aware concept and research has the highest per- centage of context-aware articles ( articles, %). in context- aware concept and research, the category can be divided into subjects. most of the articles are related to ‘‘development guide- line”, ‘‘context data management” ( articles) and ‘‘algorithm” ( articles). the ‘‘algorithm” is divided into categories, namely ‘‘agent” ( articles), ‘‘context reasoning” ( articles) and ‘‘service recommendation” ( articles). there are many articles related to guidelines, context data management and algorithm to develop the context-aware application and systems. other topics are ‘‘over- view” ( articles), ‘‘framework” ( articles), ‘‘privacy and security” ( articles), and ‘‘evaluation” ( articles). the second highest per- centage of context-aware articles is related to context-aware appli- cation ( articles, %), which has six categories. the majority of this section is related to the ‘‘m-commerce” ( articles)”. the ‘‘smart spaces” ( articles) are the second portion of context- aware application and are composed of ‘‘home” ( articles), ‘‘hospi- tal” ( articles) and ‘‘class room” ( article). other topics are ‘‘tour guide” ( articles) and ‘‘communication systems” ( articles), ‘‘information systems” ( articles) and ‘‘web service” ( articles). network infrastructure has the third highest percentage of con- text-aware articles ( articles, %). the articles are mostly re- lated to the ‘‘internet protocol” ( articles), ‘‘handoff management” ( articles), ‘‘sensing” ( articles), ‘‘network requirement” ( articles) and ‘‘network implementation” ( arti- develop - ment guideline ext a t privacy & security e layer sensing ntation er network requirement reflective based osgi based ptive ctive sed layer usability r nication tems m-commerce web service of context-aware systems. table classification of articles by subject. classification criteria number of articles percentage of subject (%) percentage of all subjects (%) . concept and research . . . algorithm . . . . . agent . . . . . context reasoning . . . . . service recommendation . . . . development guideline . . . . framework . . . . context data management . . . . evaluation . . . . privacy and security . . . . overview . . . network infrastructure . . . protocol . . . handoff management . . . . sensing . . . . network requirement . . . . network implementation . . . middleware . . . agent-based middleware . . . . reflective middleware . . . . metadata based middleware . . . . tuple space based middleware . . . . adaptive and objective based middleware . . . . osgi based middleware . . . user infrastructure . . . interface . . . . usability . . . application . . . smart space . . . . . home . . . . . hospital . . . . . class room . . . . tour guide . . . . information systems . . . . communication systems . . . . m-commerce . . . . web service . . total j.-y. hong et al. / expert systems with applications ( ) – cles). the articles related to the middleware have articles ( %). the category of this middleware part can be also divided into six subjects. the articles are related to ‘‘agent-based middleware” ( articles), ‘‘reflective middleware” ( articles), and ‘‘adaptive and objective based middleware” ( articles). the fewest of con- text-aware articles is related to user infrastructure ( articles, %). the user infrastructure can be divided into two parts, ‘‘inter- face” ( articles) and ‘‘usability” ( articles). in result, the concept and research has the highest percentage of context-aware articles, in which the most explored subject is the algorithm. this represents that the context-aware studies are in developing stages and much research have been conducted in this area. the fewest of context-aware articles is related to user infrastructure. fig. shows the number of distribution of articles by the year and classification framework. articles of concept and research layer and application and service layer have been increased con- tinuously by increasing the total number of articles from to . almost articles which have issues of privacy and security are published in and . it means that the concern about privacy and security has increased recently. context-aware sys- tems can provide users with better service based on analyzing physical context and personal context such as user schedule, user preferences and so on. however, grasping user context causes risk of security and privacy relevant to personal informa- tion. articles which focus on context reasoning are distributed almost from to , because many algorithms for finding user context such as bayesian network algorithm, decision tree, case based reasoning and rule based reasoning has been already developed. approximately % of articles involved in application layer were published from to . in accordance with developing fields of the other layer, context-aware computing application and service have been implemented recently. how- ever, articles in user infrastructure, network infrastructure and middleware layer are decreased after the peak publishing point of each layer. for example, articles in middleware layer were in- creased until , and then decreased until. this phenomenon was caused by recent research that covers several category of framework from concept to application. in other words, the ten- dency of recent research focuses not only on basic concept but also implantation or case-study. therefore, articles of application category in our classification framework have been increased continuously. . . concept and research layer this represents the references according to the detailed catego- ries in concept and research layer in table . ‘‘context-aware con- cept and research” involves theories and foundation to construct context-aware systems. algorithm is an essential part in develop- ing the context-aware systems and is divided into three sections such as agent, context reasoning, and service recommendation. the algorithm category has research related to an ai methodology that provides the foundation for a technology of intelligent sys- tems. here, it is applied as three indexes within the concept and re- search layer: ( ) designing and modeling algorithm, message bus encoded by xml, role and communication methods of agents, ( ) increasing accuracy and efficiency of algorithm for extracting high-level context from low-level context, and ( ) developing algo- rithm or method to recommend appropriate service according to the context of user. context data management represents the framework to manage the context and context information. devel- opment guideline describes the guideline related to the develop- ment of many kinds of context-aware applications and services. the framework of context-aware systems illustrates acquiring, dis- covering, interpreting and accessing various contexts to build con- text-aware services. the framework of context-aware systems is similar to an architectural sketch of a house. it gives users a general idea of what the systems will look like and how it will implement systems. the framework shows the general capabilities of the sys- tems, a user interfaces system functions, system (data) flow, sys- tem management, etc. the evaluation is focused on the evaluation method of the system that is developed in each paper. the overview deals with the general definition, characteristics and history of context-aware. ontology is generally defined as a ‘representation of a shared conceptualization of a particular domain’, and further defined by ‘axioms that offer additional constraints and meaning as well as rules and heuristics that can derive additional useful information’. ontology language has been developed in the semantic web and has converged to a new w c standard called web ontology lan- guage (owl; smith, welty, & mcguinness, ). semantic web metadata standards such as resource description framework (rdf) and owl provide standard representation for context data table references of concept and research layer. classification criteria references concept and research algorithm agent arcos and plaza ( ), hattori, cho, ohsuga, isshiki, and honiden ( ), kao and yuan ( ), krause, smailagic, and siewiorek ( ) context reasoning anagnostopoulos, ntarladimas, and hadjiefthymiades ( ), brunato and battiti ( ), castro and munz ( ), chetan, ranganathan, and campbell ( ), drakatos, pissinou, douligeris, and makki ( ), ladd, bekris, rudys, kavraki, and wallach ( ), niemegeers and heemstra de groot ( ), ranganathan, al-muhtadi, and campbell ( ), sajal, nirmalya, and abhishek ( ), samaan and karmouch ( ), satoh ( ), smailagic and kogan ( ), weber, sorkine, lipman, and gotsman ( ), xiong and hassan ( ) service recommendation byun and cheverst ( ), hatala, wakkary, and kalantari ( ), kocaballı and koçyiğitgranular ( ), kouadri and younas ( ), kwon and kim ( ), meeuwissen, reinold, and lieminferring ( ), pertselakis, ferles, tsiolis, and stafylopatis ( ), soldatos, stamatis, azodolmolky, pandis, and polymenakos ( ), syukur and loke ( ), yuan and peng ( ) development guideline anhalt et al. ( ), augustin et al. ( ), avrahami, gergle, hudson, and kiesler ( ), bolchini et al. ( ), casas, cuartielles, marco, gracia, and falco ( ), henricksena and indulska ( ), hong and landay ( ), julien, payton, and roman ( ), kwon ( ), kwon, choi, and kim ( ), qing, károly, christian, paulo, and bernhard ( ), qiu, chang, lin, and shi ( ), van sinderen, van halteren, wegdam, meeuwissen, and eertink ( ) framework dix et al. ( ), doulkeridis, loutas, and vazirgiannis ( ), johnson ( ), jones, grandhi, terveen, and whittaker ( ), kao and yuan ( ), panagiotakis and alonistioti ( ), rogers and muller ( ), roussaki, strimpakou, and anagnostou ( ), van kranenburg, bargh, iacob, and peddemors ( ), vuković, kotsovinos, and robinson ( ) context data management abecker, bernardi, hinkelmann, kühn, and sintek ( ), chaari, ejigu, laforest f., and scuturici ( ), demiris, vlahakis, makri, papaioannou, and ioannidis ( ), korpipää et al. ( ), liao, he, and tang ( ), reignier, brdiczka, vaufreydaz, crowley, and maisonnasse ( ), roussaki, strimpakou, and pils ( ), satoh ( ), serrano and serrat ( ), simons and wirtz ( ), smith, ma, and ryan ( ), qi and venkatasubramanian ( ), zimmermann, specht, and lorenz ( ) evaluation griswold et al. ( ), morla and davies ( ), o’grady et al. ( ) privacy and security bhatti, bertino, and ghafoor ( ), bhatti, samuel, eltabakh, amjad, and ghafoor ( ), falkovych and nack ( ), hill et al. ( ), jiang and landay ( ), jorns, jung, and quirchmayr ( ), jutla, bodorik, and zhang ( ), minami and kotz ( ), tentori, favela, and rodriguez ( ), vassilis, loukas, dimitris, and stavros ( ), yee, korba, lin, and shih ( ), zhang, qi, zhao, and hou ( ) overview brown and randel ( ), chalmers ( ), dey ( ), intille ( ), lieberman and selker ( ), liechti ( ), loke ( ), prekop and burnett ( ) table references of network layer. classification criteria references network infrastructure internet protocol benedetto and nardis ( ), boukerche, pazzi, and araujo ( ), chen and mohapatra ( ), friday et al. ( ), giaffreda, de carvalho, melia, giaffreda, and de carvalho ( ), kellerer, wagner, balke, and schulzrinne ( ), lin, tsai, and lai ( ), singh and acharya ( ), tyan and mahmoud ( ), yau and karim ( ), zhu et al. ( ) handoff management balasubramaniam and indulska ( ), bellavista, corradi, and foschini ( ), marias, prigouris, papazafeiropoulos, hadjiefthymiades, and merakos ( ), mcnair and zhu ( ), qing et al. ( ), rocha and endler ( ), yang, wu, yang, and liu ( ) sensing brilingaitė and jensen ( ), gellersen, schmidt, and beigl ( ), gellersen, kortuem, schmidt, and beigl ( ), goulev, stead, mamdani, and evans ( ), gross, egla, and marquardt ( ), jonker, persa, caarls, de jong, and lagendijk ( ), knight et al. ( ), matsushita ( ), michahelles and samulowitz ( ), pfeifer ( ), schmidt and laerhoven ( ), xu, fu, lee, and winter ( ), zhong and gilbert ( ) network requirement dixit ( ), harter, hopper, steggles, ward, and webster ( ), kanter ( ), yang et al. ( ) network implementation aguiar, sarma, bijwaard, marchetti, and pacyna ( ), beigl, gellersen, and schmidt ( ), cano, ferrandez-bell, and manzoni ( ), gonzalez-castano, garcia-reinoso, gil-castineira, costa-montenegro, and pousada-carballo ( ), riva, nadeem, borcea, and iftode ( ), wang et al. ( ) table references of middleware layer. classification criteria references middleware agent-based middleware alahuhta, löthman, helaakoski, koskela, and röning ( ), bellavista, corradi, montanari, and stefanelli ( ), bellavista, corradi, and stefanelli ( ), chen et al. ( ), julien and roman ( ), khedr and karmouch ( ), khedr and karmouch ( ), kumar et al. ( ), riekki et al. ( ), román et al. ( ), soldatos, dimakis, stamatis, and polymenakos ( ) reflective middleware capra, emmerich, and mascolo c. ( ), chan and chuang ( ), chan, chuang, cao, and leong ( ), qin, shi, and suo ( ), sacramento et al. ( ) metadata based middleware bellavista, corradi, montanari, and stefanelli ( a), bellavista, corradi, montanari, and stefanelli ( b) tuple space based middleware cabri, leonardi, mamei, and zambonelli ( ), curino et al. ( ), murphy and picco ( ) adaptive and objective based middleware alex, kumar, and shirazi ( ), trumler, bagci, petzold, and ungerer ( ), yau and karim ( a), yau and karim ( b), yau et al. ( ), wu et al. ( ) osgi based middleware choi, shin, and shin ( ), gu, pung, and zhang ( ), gu, pung, and zhang ( ), yu, zhou, yu, zhang, and chin ( ) j.-y. hong et al. / expert systems with applications ( ) – in context-aware systems. in other words, ontology will play a cru- cial role in enabling the processing and sharing of information and knowledge on the middleware. the reasons are as follows: � providing a shared and common understanding of a domain that can be communicated across people and application systems (chen & finin, ). table references of application and service layer. classification criteria references applications and services smart space home baek, lee, lim, and huh ( ), intille ( ), schulzrinne, wu, sidiroglou, and berger ( ) hospital agarwal, joshi, finin, yesha, and ganous ( ), bottazzi, corradi, and montanari ( ), favela, rodríguez, preciado, and gonzález ( ), favela et al. ( ), kjeldskov and skov ( ), muñoz, rodriguez, favela, martinez-garcia, and gonzález ( ), rodríguez, favela, martínez, and muñoz ( ) class room sung et al. ( ) tour guide bellotti, berta, de gloria, and margarone ( ), cano, manzoni, and toh ( ), cheverst, mitchell, davies, and smith ( ), cheverst, smith, mitchell, friday, and davies ( ), o’hare and o’grady ( ), pashtan, heusser, and scheuermann ( ) information systems celentano and gaggi ( ), kwon ( a), kwon, yoo, et al. ( ), kwon, yoo, and suh ( ), norrie et al. ( ), signer, grossniklaus, and norrie ( ), wilson, doyle, weakliam, bertolotto, and lynch ( ) communication systems fogarty, lai, and christensen ( ), goularte, pimentel, and moreira ( ), raento, oulasvirta, petit, and toivonen ( ), ranganathan, campbell, ravi, and mahajan ( ), schmidt, takaluoma, and mantyjarvi ( ), schilit, hilbert, and trevor ( ), sumi and mase ( ), sumi and nishida ( ), udugama, kuladinithi, gorg, pittmann, and tionardi ( ), yuan and chen ( ) m-commerce anagnostopoulos, tsounis, and hadjiefthymiades ( ), becerra-fernandez, cousins, and weber ( ), bouvin, christensen, frank, and hansen ( ), broens, halteren, sinderen, and wac ( ), chakraborty et al. ( ), genco, sorce, reina, and santoro ( ), kwon ( ), kwon and sadeh ( ), kwon, shin, and kim ( ), maiden, omo, seyff, grunbacher, and mitteregger ( ), mandato, kovacs, hohl, and amir-alikhani ( ), riva and toivonen ( ), roussos, marsh, and maglavera ( ), skov and høegh ( ), stylianos, nikia, kia, and christos ( ), syukur and loke ( ),wac, halteren, bults, and broens ( ), wohltorf, cissée, and rieger ( ), yu et al. ( ) web service blake, kahan, and nowlan ( ), debaty, goddi, and vorbau ( ), ceri, daniel, facca, and matera ( ), gandon and sadeh ( ), kanter ( ), kwon ( b), kwon, choi, et al. ( ), kwon, yoo, et al. ( ), lee ( ), pashtan, kollipara, and pearce ( ) table references of user infrastructure layer. classification criteria references user infrastructure interface alexander and matthias ( ), bell, feiner, and höllerer ( ), calvary et al. ( ), hatala and wakkary ( ), hong, dickson, chiu, shen, and kafeza ( ), korhonen et al. ( ), kurvinen, lähteenmäki, salovaara, and lopez ( ), lieberman and chu ( ), lum and lau ( ), mäntyjärvi and seppänen ( ), rehman, stajano, and coulouris ( ), selker ( ), smailagic and siewiorek ( ) usability barnard, yi, jacko, and sears ( ), burrell and gay ( ), kaasinen ( ) fig. . distribution of articles number by year and classification framework. j.-y. hong et al. / expert systems with applications ( ) – � facilitating knowledge sharing and reuse in an open and dynamic distributed systems (kwon, choi, et al., ; kwon, yoo, et al., ; gruber, ). � deriving fresh knowledge and facts based on reasoning contex- tual data and information by using inference engines. � allowing devices and agents not expressly designed to work together to interoperate, achieving ‘‘serendipitous interoperabil- ity” (wang, dong, chin, hettiarachchi, & zhang, ). although some research only focuses on ontology, the number is very few. ontology is generally used for reasoning and express- ing context data. therefore, they do not need to have a separate category for ontology. . . network layer as shown in table , we categorize the network infrastructure layer into internet protocol, handoff management, sensing, net- work requirements and network implementation. context-aware computing needs to dynamically adapt to changes in this context and connect entities based on network. therefore, many researches are conducted to offer appropriate network for providing context- aware computing. internet protocol category involves presenting mechanisms and design of session initiation protocol (sip), mobile ipv , an integral part of constructing self-configuring mobile ad hoc networks and object discovery. handoff management category has article about the process of transferring an ongoing call or data session from one channel connected to the core network to an- other. article topic of handoff management category is treating heterogeneous network for seamless communication environ- ments. the sensing category is responsible for capturing, abstract- ing and acquiring context information of the situations in the environment. when implementing context-aware systems, the ba- sic idea is to use the sensor data and predict the current situation. therefore, the sensing category consists of sensing algorithm, sens- ing technology and wearable sensing mechanism. network requirements category means collecting requirements for the adaptation method which can support seamless computing infra- structures, framework for integration for user services out to the j.-y. hong et al. / expert systems with applications ( ) – mobile devices and heterogeneous network infrastructures. net- work implementation category involves realizing the network which can provide context-aware computing environments. although many researches related with context-aware computing illustrate network layer, the research which are specially focused on network are only involved in this category. a part of research related with constructing context-aware systems involves an illus- tration of sensing. however, these researches do not belong to the network category, because network sensing is not their core part. . . middleware layer table indicates the references according to the detailed cate- gories in middleware layer. the context-aware systems can be made up of several basic parts. almost all the middleware parts gather context information, process it and derive meaningful ac- tions from it. context-aware systems must have middleware sup- port for providing customized services. the middleware allows agents to acquire contextual information easily, reason about it using different logics and then adapt themselves to changing con- texts. therefore, the middleware facilitates the development of context-aware agents. there are many types of middleware. this paper divides the currently available middleware for context- aware systems into seven categories. the essential part of agent- based middleware is mobile agents. mobile agents are regarded as a highly potential technology for many application areas. they emerge as a middleware technology suitable to develop context- aware services, and have been developed to implement the needed active infrastructure and the ma-based middleware design and implementation. reflective middleware possesses the unique abil- ity to model itself through self-representation, such that the manipulation of its behavior may be changed. a middleware sys- tems with self-representation is causally connected if changes made to the self-representation directly affect the implementation of the middleware. the opposite is true if changes to the middle- ware implementation change the self-representation. metadata provides information on users, devices resources, and the preferred binding strategies. metadata based middleware facilities perform runtime binding management actions based on metadata and con- text and location visibility. tuple spaces have been considered for use in sensor networks and as a blackboard with a mechanism that permits to retrieve partially known information. this is a shared repository of information, which can be accessed via well-defined primitives. adaptive middleware is a middleware for context- aware applications in smart-home setups. in this scheme, the mid- dleware matches the quality of context (qoc) requirements with the qoc available with the sensors. the application’s qoc require- ments are mapped to a utility function using the qoc attributes of the sensors available. osgi-based middleware reliably manages context-aware services to support context acquisition, discovery, and reasoning. the middleware using agents for managing complexity in the delivery of context-aware systems have recently emerged. auton- omous, proactive, rational and social agents are engaged in oppor- tunistic collaboration in the solution of appropriate service, because context-aware systems are comprised of many dynami- cally interacting components. numerous incarnations of agents are found from the reactive agent that responds in a stimulus re- sponse manner, the agent that minimizes the need for a complex representational context model, to the agent that support reason- ing in a collaborative context. . . application and service layer table indicates the references according to detailed categories in applications and services layer. there are many types of context- aware application, which provide the user with smart environment such as home, hospital, class room, etc. when a user wants to go to unfamiliar places such as museum, they play an important role as a tour guide. in addition, context-aware applications include infor- mation systems, especially decision support systems, communica- tion systems as social community, m-commerce, and web service. based on the analysis of the literature review related with con- text-aware application and service, characteristics of context- aware computing application and service are extracted as follows: � compute device to operate independently of human control. to provide the user with services autonomously, context-aware applications are customizable by using user context and prefer- ence. automatic execution of service is conducted when in a cer- tain context. � context-aware applications act in anticipation of future goals or problems. context-aware applications not only handle current task, situation and action but also, anticipate future behavior, moving point and problem of user. � in context-aware application, various devices are interconnected among each other and they recognize each other on a certain distance, because proximal devices distance between user and device or between devices are important factors to offer appro- priate service to user. � context-aware applications provide a rich set of capabilities and services to the nomad as she moves from places to place in a transparent and convenient form. . . user infrastructure layer table shows the references according to the detailed catego- ries in user infrastructure layer. it is characteristic for handheld de- vices and their users that they are continuously moving in several simultaneous fuzzy contexts. the usefulness of a given context- aware systems is most likely to suffer from the small screen and a dynamic context of users. mobility is at its core the very essence of context-awareness. the dynamic environment sets special requirements for usability and acceptance of context-aware sys- tems. therefore, research of user interface (ui) and usability of handheld device are carried out. in ui category, a content adapta- tion system for overcoming small interface such as pda and cellu- lar phone, user modeling and human–computer interaction for considering the emergence of ubiquitous and mobile computing environments are presented. in ui research, it is suggested that the contents were confusing when a screen changed due to a loca- tion change. usability category involves investigating the user needs based on user interviews, field evaluations with users, and expert evaluations of context-aware services. appearing context- aware devices and services introduces a new level of complexity when it comes to usability evaluation. usability evaluation meth- ods have been measured against desktop environments. some researchers investigated general guidelines for evaluating con- text-aware systems, while others have evaluated novel interfaces and ways of interacting. they suggested a hybrid evaluation meth- od comprised of cognitive walkthrough and heuristic evaluation for ui supporting mobility of context-aware systems. . discussions and suggestions context-aware systems are still developing in order to improve. many researchers have been concerned about context-aware sys- tems and context-awareness, and as shown in fig. , the concern has increased from to . moreover, reports or white paper of many project have been published increasingly. however, con- j.-y. hong et al. / expert systems with applications ( ) – text-aware systems are not fully implemented in real life. as shown in table , much research has focused on the concept and research layer. the scope of applications or services in most articles is limited to small regions; laboratory, school, hospital, smart room and so on. in addition, strategic alternatives or business models for gaining the revenue by using context-aware systems are very few. furthermore, technologies related to context-aware systems are merely standardized. the architecture, method of context model- ing, inference algorithm, network implementation and devices of users in each project are different. furthermore, middleware, applications and services make use of different level of contexts and adapt the way they behave according to the current context. therefore, according to the level and type of contexts along with the goal of context-aware systems, the context modeling method- ology, inference algorithm, agent structure, and interaction meth- od of agent are changed. although the interaction between agents in the same middleware and cooperation between compo- nents of the same architecture are investigated, a standard for interaction, cooperation and operation in the different context- aware application has not been studied. finally, we found that con- text sensing, context managing and context-aware services and applications are included in ubiquitous computing environment. context awareness is a key factor for new applications in the area of ubiquitous computing. the immediate future of context-aware application and service may increasingly lie in its everyday life; as researchers are begin- ning to realize the impossibility of developing systems and appli- cation only in laboratory environments. therefore, possible future research agenda may include answers to the following questions. . . how to extract and use the cognitive context in context-aware application? most of the context-aware systems focus on the external con- text, called physical context. external context means context data collected by physical sensors. it involves context data of the phys- ical environment, location data, distance, function on to other ob- jects, temperature, sound, air pressure, time, lighting levels surrounding users, and so on. the external context is important, and very useful for context-aware systems, because context-aware systems provide recommended services for a person based on ana- lyzing the external data. however, to provide personalized services according to user preferences, task and emotional state of user, the cognitive domains, such as information retrieval, decision making, situation monitoring, and so on, are needed. however, a few authors have addressed utilizing the cognitive elements of a user’s context. several researchers have proposed models to capture the internal elements of context. this model differs from many of the previous approaches, because it focuses on extracting a user’s cog- nitive activities, rather then extracting the user’s movement based on a physical environment. cognitive context information is key information to satisfy user by providing personalized context- aware computing services. in the early stage of research, only the physical context information was collected to be aware of user’s context. recently, some of the literatures which focus on cognitive context have been introduced. however, it is insufficient to estab- lish context-aware systems that reflect cognitive context. to pro- vide suitable and personalized service for the user, more studies need to be conducted in this area. . . what are design patterns of context-aware systems? design patterns are defined as recurring solutions to design problems you see over (berczuk ). they focus on the reuse of recurring architectural design themes, and have a recurring de- sign problem that is generated in a specific domain and suggest a solution (buschmann, meunier, rohnert, sommerlad, & stal, ). design patterns from another format sharing knowledge such as development guidelines are distinguished by the following reasons. � to help designers address high-level problems as well as low- level ones by showing the hierarchical structure between design patterns. � to capture the essence of recurring problems and their solutions in a compact form. � to have many examples of actual designs, alternatives to apply the solution, and some of the tradeoffs in applying the solution. design patterns have been proposed in many domains as a for- mat for capturing and sharing design knowledge between practi- tioners. they can also assist designers in developing context- aware systems. design patterns of context-aware systems will pro- vide an effective way for sharing solutions to design problems and reuse prior design knowledge. also, the designer of the context- aware systems will avoid dangerous design problem and easily solve the difficult recurring problems. some effort has aimed at unearthing common design patterns for context-aware systems. however, suggested design patterns are not specific, nevertheless, general and very few research has been conducted. therefore, re- search on design pattern of context-aware systems should be con- ducted, because there are many potential benefits in developing a design pattern. . . which is the best inferring algorithm to extract user context and provide service to user? various algorithms used in context-aware systems are classified into two parts. first, algorithm is utilized to infer high-level con- text of user. according to levels of abstraction, context is divided into low-level context and high-level context. low-level context is raw data collected directly from physical sensors, while high-le- vel context is inferred from low-level context. this part involves the algorithm of context reasoner to extract high-level context and inferring algorithm to extract correct position of user, near ob- ject, and environments. second, algorithm is recommending suit- able service. according to the activity, location and preference of user, recommended services of context-aware systems are differ- ent. although location is critical to determine service, user profile (such as sex, age, style, preference, and so on) also has influence on recommending service. therefore, for increasing user satisfac- tion by recommending service that customer wants to receive, an inferring algorithm based on context and user profile is used. in the two parts, bayesian networks for handling causal relationships between various events, probabilistic logic, fuzzy logic, decision tree, neural network and support vector machine (svm) are ap- plied. moreover, many researches focus on improving accuracy and efficiency of these algorithms. however, the research for com- paring these algorithms according to the context data type, amount of context data, research scenario and components of envi- ronments has not been carried out. since finding the suitable ser- vice quickly and correctly from on low-level context is very important in limited computing environments, research related with the best inferring algorithm for extracting user context and providing the service with users should be conducted. . . how to deal with concurrently enormous data, information and knowledge having different format to serve suitable service to users? the raw data of low-level context are usually gathered from dif- ferent physical sensors. data type, formats and abstraction level j.-y. hong et al. / expert systems with applications ( ) – from different physical sensors are different. devices and physical sensors of context-aware systems use various scale and unit, and low-level context has different elements. context-aware systems store data, information and knowledge that have different relation- ship, format and abstraction level in the context base. furthermore, context-aware systems collect context history storing sensor data over time to offer proactive service. context history stores huge amount of data on location, temperature, lighting level, task, uti- lized devices, state of devices, selected services and so on. to quickly provide suitable service to users, context-aware systems should manage variety, diversity and numerous amount of context. however, previous research suggested only a concept to control this problem. therefore, a methodology and real implication for treating variety, diversity and numerous mount of context are needed. . . how to extract the best solution when the context of users is conflicted? conflicts occur on context-aware computing environments when many users use the several physical sensors and service sources. thus, the conflicts of context-aware systems are classified into sensing conflict, service resource conflict and user preference conflict. � sensing conflict: not matching results from several physical data sources. if the coordinates of gps and spotting of a camera are different, then sensing conflict is generated. � service resource conflict: selecting only several users among many users that want to be provided with the same service, due to limited service resource. if context-aware systems offer services without possessing all the necessary service resources, then service resource conflict occurs. � user preference conflict: providing only several users with per- sonalized service, because the preference of users is not the same, although the same context of users is identified. if family members like different tv programs at the same time, then pref- erence difference between family members leads to user prefer- ence conflict. the problem of conflicts has been approached by using informa- tion fusion, time stamps and fuzzy algorithm. however, it is still not solved perfectly, because it is difficult to provide personalized services under the limited resource. therefore, a study on method- ology and actual guideline for harmonizing all conflicts are required. . . how to reflect the preference of users for satisfying user needs? the preference of each user is different according to user con- text and profile. predicting the preferences of users and providing the personalized services based on users’ preferences have carried out by various research. especially, predicting the preferences of users in e-commerce users for providing the personalized services is one of the important issues. recommendation systems providing personalized services have been carried out by many researchers. however, the research considering users’ preferences on context- aware computing is a relatively insufficient in the context-aware systems research field (byun & cheverst, ). there were some limitations in previous research for providing the personalized ser- vices on context-aware systems. first, the users have to input their preferences directly. second, they did not provide automated ser- vices. finally, it is difficult to provide new user with the personal- ized services. thus, it is difficult to provide the users with the automatic personalized services due to these limitations. hence, proposing the context-aware systems to provide the personalized services based on context history in context-aware computing is needed. . . how to save users information in context-aware systems? as table shows, the research has been conducted in the area of security and privacy. the context-aware systems that do not consider user privacy and security in systems that deal with per- sonal data will not be publicly accepted. therefore, general and several approaches to implement security and privacy in context- aware systems have been studied. although privacy and security are an important aspect of context-aware systems, applicable con- text-aware systems utilizing these approaches have not been investigated yet, due to their complexity. in other words, most re- searches have only focused the concept for protecting privacy and security. the more research related with privacy and security is needed for context-aware systems to become a reality. a user utilizing context-aware systems uses many available computers embedded in everyday artifacts, like pda, navigation, and so on. each user uses many personal computing devices, and at the same time, the same publicly available device is used by many users. furthermore, context-aware systems store and handle sensitive and personal data. common to context-aware systems is the lack of appropriate security and privacy mechanisms that can effectively ensure a secure user authentication. therefore, the need for authentication arises as the information and data stored in con- text base is highly personal and should be protected. hence, more research on context-aware systems that can be protected from unauthorized access is needed. . . how to evaluate performance of context-aware systems? it is essential to improve managing and planning context-aware systems based on performance evaluation, because ‘‘what you measure is what you get”. various models such as balanced score card (bsc), is success model, control objectives for information and related technology (cobit), information economics, and so on, have been applied to suggest measures that can evaluate the performance of information systems by considering diverse per- spectives. due to the unique characteristics of context-aware sys- tems such as portability, mobility, proactiveness, nomadicity, and so on, the current technology for evaluating information systems (is) does not provide techniques to formally define, verify, imple- ment, and analyze them. therefore, research on the model and per- formance measure reflecting the characteristics of context-aware systems should be conducted. . conclusions the review was organized according to the framework devel- oped with the purpose of providing a comprehensive overview of research on context-aware systems. the review introduced the context-aware systems concept, network infrastructure, middle- ware, application, user infrastructure and presented an exhaustive list of each layer of context-aware systems. in this paper, we re- viewed the literatures for the concept and applications, and exam- ined them using dimensions related to ongoing and emerging issues in context-aware systems. this paper is based on a literature review of context-aware systems from to using a key- word index and article title search. overall, we found that the activity related with context-aware computing seems to increase dramatically and can be classified into designing and implement- ing network, user infrastructure on handheld device, middleware which are enough to provides users with context-aware services, concept and research being the base of context-ware application, j.-y. hong et al. / expert systems with applications ( ) – and application and service. it provides sufficient literature for the researchers on the use of context-aware systems, and we hope that it will motivate the researchers and practitioners to examine con- text-aware systems issues and its applications. the framework of context-aware systems offers a general development guideline, fundamental component and relationships within component of context-aware systems. furthermore, on the basis of this frame- work and the literature review, selected opportunities for future research were discussed. it is concluded that different social science methodologies, such as psychology, cognitive science, and human behavior could imple- ment context-aware computing. integration of qualitative, quanti- tative and scientific methods and integration of context-aware theory and concept studies may broaden our horizon on this sub- ject. finally, the ability to continually change and obtain new understanding is the power of context-aware computing. although considerable attention was given to the framework de- sign and classification of the literature review, some limitations still exist. first, some relevant articles might have been overlooked. much the literature has been found by the title, keyword or abstract only. although the title in most cases describes the content quite well this is not always so. in order to conduct a comprehensive lit- erature review, the topical focus was kept relatively narrow on con- text-aware systems. moreover, this paper makes a brief literature review of context-aware systems from to in order to ex- plore how context-aware systems have developed in this period. white paper, practical reporter and many articles related with con- text-aware systems were not reviewed, because seven online dat- abases were searched in this paper. period and searching limitation may not satisfy the need of readers looking for a review on context-aware systems. finally, the correlation between the framework and real projects of context-aware systems are not illus- trated. the review did not involve the role and interaction methods of logical components in architectures of socam, cass, cobra, guide, hydrogen, gaia, context toolkit, stu project and so on, since only articles of journals were extracted and analyzed. references abecker, a., bernardi, a., hinkelmann, k., kühn, o., & sintek, m. ( ). context- aware, proactive delivery of task-specific information: the knowmore project. information systems frontiers, ( ), – . agarwal, s., joshi, a., finin, t., yesha, y., & ganous, t. ( ). a pervasive computing system for the operating room of the future. mobile networks and applications, ( – ), – . aguiar, r. l., sarma, a., bijwaard, d., marchetti, l., & pacyna, p. ( ). pervasiveness in a competitive multi-operator environment: the daidalos project. communications magazine, ieee, ( ), – . alahuhta, p., löthman, h., helaakoski, h., koskela, a., & röning, j. ( ). experiences in developing mobile applications using the apricot agent platform. personal and ubiquitous computing, ( ), – . alex, h., kumar, m., & shirazi, m. ( ). midfusion: an adaptive middleware for information fusion in sensor network applications. information fusion, ( ), – . alexander, z., & matthias, j. ( ). implementing adaptive mobile gi services based on ontologies: examples from pedestrian navigation support. computers, environment and urban systems, ( ), – . anagnostopoulos, c. b., ntarladimas, y., & hadjiefthymiades, s. ( ). situational computing: an innovative architecture with imprecise reasoning. journal of systems and software, ( ), – . anagnostopoulos, c. b., tsounis, a., & hadjiefthymiades, s. ( ). context awareness in mobile computing environments. wireless personal communications, ( ), – . anhalt, j., smailagic, a., siewiorek, d. p., gemperle, f., salber, d., weber, s., et al. ( ). toward context-aware computing: experiences and lessons. ieee intelligent systems, ( ), – . arcos, j. l., & plaza, e. ( ). context aware personal information a gents. international journal of cooperative information systems, ( – ), – . augustin, i., yamin, a. c., da silva, l. c., real, r. a., frainer, g., & geyer, c. f. r. ( ). isamadapt: abstractions and tools for designing general-purpose pervasive applications. software: practice and experience, ( – ), – . avrahami, d., gergle, d., hudson, s. e., & kiesler, s. ( ). improving the match between callers and receivers: a study on the effect of contextual information on cell phone interruptions. behaviour and information technology, ( ), – . baek, s. h., lee, h. j., lim, s. y., & huh, j. d. ( ). managing mechanism for service compatibility and interaction issues in context-aware ubiquitous home. ieee transactions on consumer electronics, ( ), – . balasubramaniam, s., & indulska, j. ( ). vertical handover supporting pervasive computing in future wireless networks. computer communications, ( ), – . baldauf, m., dustdar, s., & rosenberg, f. ( ). a survey on context-aware systems. international journal of ad hoc and ubiquitous computing, ( ), – . barnard, l., yi, j. s., jacko, j. a., & sears, a. ( ). an empirical comparison of use-in- motion evaluation scenarios for mobile computing devices. international journal of human–computer studies, ( ), – . becerra-fernandez, i., cousins, k. c., & weber, r. o. ( ). nomadic context-aware knowledge management systems: applications, challenges and research problems. international journal of mobile learning and organisation, ( ), – . beigl, m., gellersen, h. w., & schmidt, a. ( ). media cups: experience with design and use of computer-augmented everyday artefacts. computer networks, ( ), – . bell, b., feiner, s., & höllerer, t. ( ). visualization viewpoints. ieee computer graphics and applications, ( ), – . bellavista, p., corradi, a., & foschini, l. ( ). context-aware handoff middleware for transparent service continuity in wireless networks. pervasive and mobile computing, ( ), – . bellavista, p., corradi, a., montanari, r., & stefanelli, c. ( a). context-aware middleware for resource management in the wireless internet. ieee transactions on software engineering, ( ), – . . bellavista, p., corradi, a., montanari, r., & stefanelli, c. ( b). dynamic binding in mobile applications: a middleware approach. ieee internet computing, ( ), – . . bellavista, p., corradi, a., montanari, r., & stefanelli, c. ( ). a mobile computing middleware for location- and context-aware internet data services. acm transactions on internet technology, ( ). bellavista, p., corradi, a., & stefanelli, c. ( ). the ubiquitous provisioning of internet services to portable devices. ieee pervasive computing, ( ), – . bellotti, f., berta, r., de gloria, a., & margarone, m. ( ). implementing tour guides for travelers. human factors and ergonomics in manufacturing, ( ), – . benedetto, m. d., & nardis, l. d. ( ). tuning uwb signals by pulse shaping: towards context-aware wireless networks. signal processing, ( ), – . berczuk, s. ( ). finding solutions through pattern languages. ieee computer, ( ), – . bhatti, r., bertino, e., & ghafoor, r. ( ). a trust-based context-aware access control model for web-services. distributed and parallel databases, ( ), – . bhatti, r., samuel, a., eltabakh, m. y., amjad, h., & ghafoor, a. ( ). engineering a policy-based system for federated healthcare databases. knowledge and data engineering, ieee, ( ), – . blake, m., kahan, d., & nowlan, m. ( ). context-aware agents for user-oriented web services discovery and execution. distributed and parallel databases, ( ), – . bolchini, c., schreiber, f. a., & tanca, l. ( ). a methodology for a very small data base design. information systems, ( ), – . bottazzi, d., corradi, a., & montanari, r. ( ). context-aware middleware solutions for anytime and anywhere emergency assistance to elderly people. communications magazine, ieee, ( ), – . boukerche, a., pazzi, r. w. n., & araujo, r. v. ( ). fault-tolerant wireless sensor network routing protocols for the supervision of context-aware physical environments. journal of parallel and distributed computing, ( ), – . bouvin, n. o., christensen, b., frank, k. g., & hansen, a. ( ). hycon: a framework for context-aware mobile hypermedia. new review of hypermedia and multimedia, , – . brilingaitė, a., & jensen, c. s. ( ). enabling routes of road network constrained movements as mobile service context. geoinformatica, ( ), – . broens, t., halteren, a. v., sinderen, m. v., & wac, k. ( ). towards an application framework for context-aware m-health applications. international journal of internet protocol technology, ( ), – . brown, b., & randel, r. ( ). building a context sensitive telephone: some hopes and pitfalls for context sensitive computing. computer supported cooperative work, ( – ), – . brunato, m., & battiti, r. ( ). statistical learning theory for location fingerprinting in wireless lans. computer networks, ( ), – . burrell, j., & gay, g. k. ( ). e-graffiti: evaluating real-world use of a context- aware system. interacting with computers, ( ), – . buschmann, f., meunier, r., rohnert, h., sommerlad, p., & stal, m. ( ). pattern- oriented software architecture – a system of patterns. wiley & sons ltd.. byun, h. e., & cheverst, k. ( ). utilizing context history to provide dynamic adaptations. applied artificial intelligence, ( ), – . cabri, g., leonardi, l., mamei, m., & zambonelli, f. ( ). location-dependent services for mobile users. ieee transactions on systems, ( ), – . calvary, g., coutaz, j., thevenin, d., limbourg, q., bouillon, l., & vanderdonckt, j. ( ). a unifying reference framework for multi-target user interfaces. interacting with computers, ( ), – . j.-y. hong et al. / expert systems with applications ( ) – cano, j. c., ferrandez-bell, d., & manzoni, p. ( ). evaluating bluetooth performance as the support for context-aware applications. telecommunication systems, ( – ), – . cano, j. c., manzoni, p., & toh, c. k. ( ). ubiqmuseum: a bluetooth and java based context-aware system for ubiquitous computing. wireless personal communications, ( ). capra, l., emmerich, w., & mascolo, c. ( ). carisma: context-aware reflective middleware system for mobile applications. ieee transactions on software engineering, ( ), – . casas, r., cuartielles, d., marco, a., gracia, h. j., & falco, j. l. ( ). hidden issues in deploying an indoor location system. ieee pervasive computing, ( ), – . castro, p., & munz, r. ( ). managing context data for smart spaces. ieee personal communications, ( ), – . celentano, a., & gaggi, o. ( ). context-aware design of adaptable multimodal documents. multimedia tools and applications, ( ). ceri, s., daniel, f., facca, m. f., & matera, m. ( ). model-driven engineering of active context-awareness. world wide web, ( ), – . chaari, t., ejigu, d., laforest, f., & scuturici, v. m. ( ). a comprehensive approach to model and use context for adapting applications in pervasive environments. journal of systems and software, ( ), – . chakraborty, d., dasgupta, k., mittal, s., misra, a., gupta, a., newmark, e., et al. ( ). businessfinder: harnessing presence to enable live yellow for small, medium and micro mobile businesses. communications magazine, ieee, ( ), – . chalmers, m. ( ). a historical view of context, a historical view of context. computer supported cooperative work (cscw), ( ), – . chan, a. t. s., & chuang, s. n. ( ). mobipads: a reflective middleware for context-aware mobile computing. ieee transactions on software engineering, ( ), – . chan, a. t. s., chuang, s. n., cao, j., & leong, h. v. ( ). an event-driven middleware for mobile context awareness. the computer journal, ( ), – . chen, h., & finin, t. ( ). an ontology for a context aware pervasive computing environment. in proceedings of ijcai workshop on ontologies and distributed systems, ijcai, . chen, g., & kotz, d. ( ). a survey of context-aware mobile computing research. dartmouth college technical report tr - . chen, h., finin, t., joshi, a., kagal, l., perich, f., & chakraborty, d. ( ). intelligent agents meet the semantic web in smart spaces. ieee internet computing, ( ), – . chen, h., & mohapatra, p. ( ). a context-aware html/xml document transmission process for mobile wireless clients. world wide web, ( ), – . chetan, s., ranganathan, a., & campbell, r. ( ). towards fault tolerant pervasive computing. ieee technology and society magazine, ( ), – . cheverst, k., mitchell, k., davies, n., & smith, g. ( ). exploiting context to support social awareness and social navigation. acm siggroup bulletin, ( ), – . cheverst, k., smith, g., mitchell, k., friday, a., & davies, n. ( ). the role of shared context in supporting cooperation between city visitors. computers & graphics, ( ), – . choi, j. h., shin, d. k., & shin, d. i. ( ). research and implementation of the context-aware middleware for controlling home appliances. ieee transactions on consumer electronics, ( ), – . curino, c., giani, m., giorgetta, m., giusti, a., murphy, a. l., & picco, g. p. ( ). mobile data collection in sensor networks: the tinylime middleware. pervasive and mobile computing, ( ), – . debaty, p., goddi, p., & vorbau, a. ( ). integrating the physical world with the web to enable context-enhanced mobile services. mobile networks and applications, ( ), – . demiris, a. m., vlahakis, v., makri, a., papaioannou, m., & ioannidis, n. ( ). intguide: a platform for context-aware services featuring augmented-reality, based on the outcome of european research projects. signal processing: image communication, ( – ), – . dey, a. k. ( ). understanding and using context. personal and ubiquitous computing, ( ), – . dix, a., rodden, t., davies, n., trevor, j., friday, a., & palfreyman, k. ( ). exploiting space and location as a design framework for interactive mobile systems. acm transactions on computer–human interaction (tochi), ( ), – . dixit, s. ( ). wireless ip and its challenges for the heterogeneous environment. wireless personal communications, ( ), – . doulkeridis, c., loutas, n., & vazirgiannis, m. ( ). a system architecture for context-aware service discovery. electronic notes in theoretical computer science, ( ), – . drakatos, s., pissinou, n., douligeris, c., & makki, k. ( ). a context-aware cache structure for mobile computing environments. journal of systems and software, ( ), – . falkovych, k., & nack, f. ( ). context aware guidance for multimedia authoring: harmonizing domain and discourse knowledge. multimedia systems, ( ). favela, j., rodríguez, m., preciado, a., & gonzález, v. m. ( ). integrating context- aware public displays into a mobile hospital information system. ieee transactions on information technology in biomedicine, ( ), – . favela, j., tentori, m., castro, l. a., gonzalez, v. m., moran, e. v., & martínez-garcía, a. i. ( ). activity recognition for context-aware hospital applications: issues and opportunities for the deployment of pervasive networks. mobile networks and applications, ( – ), – . fogarty, j., lai, j., & christensen, j. ( ). presence versus availability: the design and evaluation of a context-aware communication client. international journal of human–computer studies, ( ), – . friday, a., wu, m., finney, j., schmid, s., cheverst, k., & davies, n. ( ). network layer access control for context-aware ipv applications. wireless networks, ( ), – . gandon, f. l., & sadeh, n. m. ( ). semantic web technologies to reconcile privacy and context awareness. web semantics: science, services and agents on the world wide web, ( ), – . gellersen, h., kortuem, g., schmidt, a., & beigl, m. ( ). physical prototyping with smart-its. ieee pervasive computing, ( ), – . gellersen, h. w., schmidt, a., & beigl, m. ( ). multi-sensor context-awareness in mobile devices and smart artefacts. mobile networks and applications, ( ), – . genco, a., sorce, s., reina, g., & santoro, g. ( ). an agent-based service network for personal mobile devices. ieee pervasive computing, ( ), – . giaffreda, f., de carvalho, r., melia, t., giaffreda, f., & de carvalho, t. ( ). media network enablers for enhanced mobility management. bt technology journal, ( ), – . gonzalez-castano, f. j., garcia-reinoso, j., gil-castineira, f., costa-montenegro, e., & pousada-carballo, j. m. ( ). bluetooth-assisted context-awareness in educational data networks. computers education, ( ), – . goularte, r., pimentel, g. c. m., & moreira, e. d. s. ( ). context-aware support in structured documents for interactive-tv. multimedia systems, ( ). goulev, p., stead, l., mamdani, e., & evans, c. ( ). computer aided emotional fashion. computers & graphics, ( ), – . griswold, w. g., shanahan, p., brown, s. w., boyer, r., ratto, m., shapiro, r. b., et al. ( ). activecampus: experiments in community-oriented ubiquitous computing. ieee computer, ( ), – . gross, t., egla, t., & marquardt, n. ( ). sensation: a service-oriented platform for developing sensor-based infrastructures. international journal of internet protocol technology, ( ), – . gruber, t. ( ). a translation approach to portable ontology specifications. knowledge acquisition, ( ), – . gu, t., pung, h. k., & zhang, d. q. ( ). toward an osgi-based infrastructure for context-aware applications. ieee pervasive computing, ( ), – . gu, t., pung, h. k., & zhang, d. q. ( ). a service-oriented middleware for building context-aware services. journal of network and computer applications, ( ), – . harter, a., hopper, a., steggles, p., ward, a., & webster, p. ( ). the anatomy of a context-aware application. wireless networks, ( – ), – . hatala, m., & wakkary, r. ( ). ontology-based user modeling in an augmented audio reality system for museums. user modeling and user-adapted interaction, ( – ), – . hatala, m., wakkary, r., & kalantari, l. ( ). rules and ontologies in support of real-time ubiquitous application. web semantics: science, services and agents on the world wide web, ( ), – . hattori, m., cho, k., ohsuga, a., isshiki, m., & honiden, s. ( ). context-aware agent platform in ubiquitous environments and its verification tests. systems and computers in japan, ( ), – . henricksena, k., & indulska, j. ( ). developing context-aware pervasive computing applications: models and approach. pervasive and mobile computing, ( ), – . hill, r., al-muhtadi, j., campbell, r., kapadia, a., naldurg, p., & ranganathan, a. ( ). a middleware architecture for securing ubiquitous computing cyber infrastructures. ieee distributed systems online, ( ). hong, d., dickson, k. w., chiu, v. y., shen, s. c. c., & kafeza, e. ( ). ubiquitous enterprise service adaptations based on contextual user behavior. information systems frontiers, ( ), – . hong, j. i., & landay, j. a. ( ). an infrastructure approach to context-aware computing. human–computer interaction, , – . intille, s. s. ( ). designing a home of the future. ieee pervasive computing, ( ), – . intille, s. s. ( ). a new research challenge: persuasive technology to motivate healthy aging. ieee transactions on information technology in biomedicine, ( ), – . jiang, x., & landay, j. a. ( ). modeling privacy control in context-aware systems using decentralized information spaces. ieee pervasive computing, ( ), – . johnson, s. ( ). a framework for mobile context-aware applications. bt technology journal, ( ), – . jones, q., grandhi, s. a., terveen, l., & whittaker, s. ( ). people-to-people-to- geographical-places: the p framework for location-based community systems. computer supported cooperative work (cscw)( ), – . jonker, p., persa, s., caarls, j., de jong, f., & lagendijk, i. ( ). philosophies and technologies for ambient aware devices in wearable computing grids. computer communications, ( ), – . jorns, o., jung, o., & quirchmayr, g. ( ). transaction pseudonyms in mobile environments. journal in computer virology, ( ), – . julien, c., payton, j., & roman, g. c. ( ). reasoning about context-awareness in the presence of mobility. electronic notes in theoretical computer science, , – . julien, c., & roman, g. c. ( ). egospaces: facilitating rapid development of context-aware mobile applications. ieee transactions on software engineering, ( ), – . jutla, d. n., bodorik, p., & zhang, y. ( ). pecan: an architecture for users’ privacy-aware electronic commerce contexts on the semantic web. information systems, ( – ), – . http://www.cs.vu.nl/~heiner/ijcai- /papers/chen.pdf j.-y. hong et al. / expert systems with applications ( ) – kaasinen, e. ( ). user needs for location-aware mobile services. personal and ubiquitous computing, ( ), – . kanter, t. g. ( ). hottown, enabling context-aware and extensible mobile interactive spaces. ieee wireless communications, ( ), – . kanter, t. g. ( ). attaching context-aware services to moving locations. ieee internet computing, ( ), – . kao, t. h., & yuan, s. m. ( ). designing an xml-based context-aware transformation framework for mobile execution environments using cc/and xslt. computer standards interfaces, ( ), – . kao, t. h., & yuan, s. m. ( ). automatic adaptation of mobile applications to different user devices using modular mobile agents. software practice and experience, ( ), – . kellerer, w., wagner, m., balke, w. t., & schulzrinne, h. ( ). preference-based session management for ip-based mobile multimedia signaling. european transactions on telecommunications, ( ), – . khedr, m., & karmouch, a. ( ). negotiating context information in context- aware systems. ieee intelligent systems, ( ), – . khedr, m., & karmouch, a. ( ). acai: agent-based context-aware infrastructure for spontaneous applications. journal of network and computer applications, ( ), – . kjeldskov, j., & skov, m. b. ( ). exploring context-awareness for ubiquitous computing in the healthcare domain. personal and ubiquitous computing, ( ), – . knight, j. f., bristow, h. w., anastopoulou, s., baber, c., schwirtz, a., & arvanitis, t. n. ( ). uses of accelerometer data collected from a wearable system. personal and ubiquitous computing, ( ), – . kocaballı, a. b., & koçyiğitgranular, a. ( ). the best match algorithm for context- aware computing systems. journal of systems and software, ( ), – . korhonen, j., ojala, t., ristola, a., kesti, m., kilpelänaho, v., koskinen, m., et al. ( ). mobile fair diary: hybrid interface for taking, browsing and sharing context-aware notes. personal and ubiquitous computing, ( ), – . korpipää, p., mäntyjärvi, j., kela, j., keränen, h., & malm, e. j. ( ). managing context information in mobile devices. ieee pervasive computing, ( ), – . kouadri, s. m., & younas, m. ( ). context-oriented and transaction-based service provisioning. international journal of web and grid services, ( ), – . krause, a., smailagic, a., & siewiorek, d. p. ( ). context-aware mobile computing: learning context-dependent personal preferences from a wearable sensor array. ieee transactions on mobile computing, ( ), – . kumar, m., shirazi, b. a., das, s. k., sung, b. y., levine, d., & singhal, m. ( ). pico: a middleware framework for pervasive computing. ieee pervasive computing, ( ), – . kurvinen, e., lähteenmäki, m., salovaara, a., & lopez, f. ( ). are you alive? sensor data as a resource for social interaction. knowledge, technology, and policy, ( ), – . kwon, o. b. ( ). i know what you need to buy: context-aware multimedia-based recommendation system. expert systems with applications, ( ), – . kwon, o. b. ( ). modeling and generating context-aware agent-based applications with amended colored petri nets. expert systems with applications, ( ), – . kwon, o. b. ( a). multi-agent system approach to context-aware coordinated web services under general market mechanism. decision support systems, ( ), – . kwon, o. b. ( b). the potential roles of context-aware computing technology in optimization-based intelligent decision-making. expert systems with applications, ( ), – . kwon, o., choi, k., & kim, m. ( ). user acceptance of context-aware services: self-efficacy, user innovativeness and perceived sensitivity on contextual pressure. behaviour information technology, ( ), – . kwon, o. b., choi, s. c., & park, g. r. ( ). nama: a context-aware multi-agent based web service approach to proactive need identification for personalized reminder systems. expert systems with applications, ( ), – . kwon, o. b., & kim, m. y. ( ). mymessage: case-based reasoning and multicriteria decision making techniques for intelligent context-aware message filtering. expert systems with applications, ( ), – . kwon, o. b., & sadeh, n. m. ( ). applying case-based reasoning and multi-agent intelligent system to context-aware comparative shopping. decision support systems, ( ), – . kwon, o. b., shin, j. m., & kim, s. w. ( ). context-aware multi-agent approach to pervasive negotiation support systems. expert systems with applications, ( ), – . kwon, o. b., yoo, k. d., & suh, e. h. ( ). ubidss: a proactive intelligent decision support system as an expert system deploying ubiquitous computing technologies. expert systems with applications, ( ), – . kwon, o. b., yoo, k. d., & suh, e. h. ( ). ubies: applying ubiquitous computing technologies to an expert system for context-aware proactive services. electronic commerce research and applications, ( ), – . ladd, a. m., bekris, k. e., rudys, a., kavraki, l. e., & wallach, d. s. ( ). robotics- based location sensing using wireless ethernet. wireless networks, ( – ), – . lee, w. p. ( ). deploying personalized mobile services in an agent-based environment. expert systems with applications, ( ), – . liao, s. s., he, j. w., & tang, t. h. ( ). a framework for context information management. journal of information science, ( ), – . lieberman, h., & chu, a. ( ). an interface for mutual disambiguation of recognition errors in a multimodal navigational assistant. multimedia systems, ( – ), – . lieberman, h., & selker, t. ( ). out of context: computer systems that adapt to learn from, context. ibm systems journal, ( – ), – . liechti, o. ( ). awareness and the www: an overview. acm siggroup bulletin, ( ), – . lin, p., tsai, c. y., & lai, y. c. ( ). a sip-based mobility management platform for wlan location-aware broadcasting and multicasting application. wireless communications and mobile computing, , – . loke, s. w. ( ). context-aware artifacts: two development approaches. ieee pervasive computing, ( ), – . lum, w. y., & lau, f. c. m. ( ). a context-aware decision engine for content adaptation. ieee pervasive computing, ( ), – . maiden, n., omo, o., seyff, n., grunbacher, p., & mitteregger, k. ( ). determining stakeholder needs in the workplace: how mobile technologies can help. software, ieee, ( ), – . mandato, d., kovacs, e., hohl, f., & amir-alikhani, h. ( ). camp: a context- aware mobile portal. ieee communications magazine, ( ), – . mäntyjärvi, j., & seppänen, t. ( ). adapting applications in handheld devices using fuzzy context information. interacting with computers, ( ), – . marias, g. f., prigouris, n., papazafeiropoulos, g., hadjiefthymiades, s., & merakos, l. ( ). brokering positioning data from heterogeneous infrastructures. wireless personal communications, ( ), – . matsushita, s. ( ). a headset-based minimized wearable computer. ieee intelligent systems, ( ), – . mcnair, j., & zhu, f. ( ). vertical handoffs in fourth-generation multinetwork environments. ieee wireless communications, ( ), – . meeuwissen, e., reinold, p., & lieminferring, c. ( ). inffering and predicting context of mobile users. bell labs technical journal, ( ), – . michahelles, f., & samulowitz, m. ( ). smart caps for smart its – context detection for mobile users. personal and ubiquitous computing, ( ), – . minami, k., & kotz, d. ( ). secure context-sensitive authorization. pervasive and mobile computing, ( ), – . morla, r., & davies, n. ( ). evaluating a location-based application: a hybrid test and simulation environment. ieee pervasive computing, ( ), – . muñoz, m. a., rodriguez, m., favela, j., martinez-garcia, a. i., & gonzález, v. m. ( ). context-aware mobile communication in hospitals. ieee computer society, ( ), – . murphy, a. l., & picco, g. p. ( ). using coordination middleware for location- aware computing: a lime case study. lecture note in computer science, – . niemegeers, i. g., & heemstra de groot, s. m. ( ). fednets: context-aware ad- hoc network federations. wireless personal communications, ( – ), – . norrie, m. c., signer, b., grossniklaus, m., belotti, r., decurtins, c., & weibel, n. ( ). context-aware platform for mobile data management. wireless networks, ( ), – . o’grady, m. j., o’hare, g. m. p., & sas, c. ( ). mobile agents for mobile tourists: a user evaluation of gulliver’s genie. interacting with computers, ( ), – . o’hare, g. m. p., & o’grady, m. j. ( ). gulliver’s genie: a multi-agent system for ubiquitous and intelligent content delivery. computer communications, ( ), – . panagiotakis, s., & alonistioti, a. ( ). context-aware composition of mobile services. it professional, ( ), – . pashtan, a., heusser, a., & scheuermann, p. ( ). personal service areas for mobile web applications. ieee internet computing, ( ), – . pashtan, a., kollipara, s., & pearce, m. ( ). adapting content for wireless web services. ieee internet computing, ( ), – . pertselakis, m., ferles, c., tsiolis, k., & stafylopatis, a. ( ). wireless distributed implementation of a fuzzy neural classification system. international journal of artificial intelligence tools, ( ), – . pfeifer, t. ( ). redundant positioning architecture. computer communications, ( ), – . prekop, p., & burnett, m. ( ). activities, context and ubiquitous computing. computer communications, ( ), – . qi, h., & venkatasubramanian, n. ( ). timeliness-accuracy balanced collection of dynamic context data. parallel and distributed systems, ieee, ( ), – . qin, w., shi, y., & suo, y. ( ). ontology-based context-aware middleware for smart spaces. tsinghua science & technology, ( ), – . qing, w., károly, f., christian, p., paulo, m., & bernhard, p. ( ). context-aware handover using active network technology. computer networks, ( ), – . qiu, l., chang, l., lin, f., & shi, z. ( ). context optimization of ai planning for semantic web services composition. service oriented computing and applications, ( ), – . raento, m., oulasvirta, a., petit, o., & toivonen, h. ( ). context phone: a prototyping platform for context-aware mobile applications. ieee pervasive computing, ( ), – . ranganathan, a., al-muhtadi, j., & campbell, r. h. ( ). reasoning about uncertain contexts in pervasive computing environments. ieee pervasive computing, ( ), – . ranganathan, a., campbell, r. h., ravi, a., & mahajan, a. ( ). conchat: a context- aware chat program. ieee pervasive computing, ( ), – . rehman, k., stajano, f., & coulouris, g. ( ). an architecture for interactive context-aware applications. ieee pervasive computing, ( ), – . reignier, p., brdiczka, o., vaufreydaz, d., crowley, j. l., & maisonnasse, j. ( ). context-aware environments: from specification to implementation. expert systems, ( ), – . j.-y. hong et al. / expert systems with applications ( ) – riekki, j., huhtinen, j., ala-siuru, p., alahuhta, p., kaartinen, j., & röning, j. ( ). genie of the net, an agent platform for managing services on behalf of the user. computer communications, ( ), – . riva, o., nadeem, t., borcea, c., & iftode, l. ( ). context-aware migratory services in ad hoc networks. mobile computing, ieee, ( ), – . riva, o., & toivonen, s. ( ). the dynamos approach to support context-aware service provisioning in mobile environments. journal of systems and software, ( ), – . rocha, r. c. a., & endler, m. ( ). middleware: context management in heterogeneous, evolving ubiquitous environments. distributed systems online, ieee, ( ), . rodríguez, m. d., favela, j., martínez, e. a., & muñoz, m. a. ( ). location-aware access to hospital information and services. ieee transactions on information technology in biomedicine, ( ), – . rogers, y., & muller, h. ( ). a framework for designing sensor-based interactions to promote exploration and reflection in play. international journal of human- computer studies, ( ), – . román, m., hess, c., cerqueira, r., ranganathan, a., campbell, r. h., & nahrstedt, k. ( ). a middleware infrastructure for active spaces. ieee pervasive computing, ( ), – . roussaki, l. g., strimpakou, m., & anagnostou, m. e. ( ). designing next generation middleware for context-aware ubiquitous and pervasive computing. international journal of ad hoc and ubiquitous computing, ( ), – . . roussaki, i., strimpakou, m., & pils, c. ( ). distributed context retrieval and consistency control in pervasive computing. journal of network and systems management, ( ), – . roussos, g., marsh, a. j., & maglavera, s. ( ). enabling pervasive computing with smart phones. ieee pervasive computing, ( ), – . sacramento, v., endler, m., rubinsztejn, h. k., lima, l. s., gonçalves, k., nascimento, f. n., et al. ( ). moca: a middleware for developing collaborative applications for mobile users. ieee distributed systems online, ( ), – . sajal, k. d., nirmalya, r., & abhishek, r. ( ). context-aware resource management in multi-inhabitant smart homes: a framework based on nash h-learning. pervasive and mobile computing, ( ), – . samaan, n., & karmouch, a. ( ). a mobility prediction architecture based on contextual knowledge and spatial conceptual maps. ieee transactions on mobile computing, ( ), – . satoh, i. ( ). spatialagents: integrating user mobility and program mobility in ubiquitous computing environments. wireless communications and mobile computing, ( ), – . satoh, i. ( ). a location model for smart environments. pervasive and mobile computing, ( ), – . schilit, b. n, adams, n. i, & want, r. ( ). in context-aware computing applications (pp. – ). ieee computer society. schilit, b. n., hilbert, d. m., & trevor, j. ( ). context-aware communication. ieee wireless communications, ( ), – . schmidt, a., & laerhoven, k. v. ( ). how to build smart appliances? ieee personal communications, ( ), – . schmidt, a., takaluoma, a., & mantyjarvi, j. ( ). context-aware telephony over wap. personal technologies, ( ), – . schulzrinne, h., wu, x., sidiroglou, s., & berger, s. ( ). ubiquitous computing in home networks. ieee communications magazine, ( ), – . selker, t. ( ). visual attentive interfaces. bt technology journal, ( ), – . serrano, j. m., & serrat, j. ( ). information modeling and handling for context- aware multimedia services. ieee wireless communications, ( ), – . signer, b., grossniklaus, m., & norrie, m. c. ( ). interactive paper as a mobile client for a multi-channel web information system. world wide web, ( ), – . simons, c., & wirtz, g. ( ). modeling context in mobile distributed systems with the uml. journal of visual languages & computing, ( ), – . singh, a., & acharya, a. ( ). multiplayer networked gaming with the session initiation protocol. computer networks, ( ), – . skov, b., & høegh, t. ( ). supporting information access in a hospital ward by a context-aware mobile electronic patient record. personal and ubiquitous computing, ( ). smailagic, a., & kogan, d. ( ). location sensing and privacy in a context-aware computing environment. ieee wireless communications, ( ), – . smailagic, a., & siewiorek, d. ( ). application design: wearable and context- aware computers. ieee pervasive computing, ( ), – . smith, m. k., welty, c., & mcguinness, d. l. ( ). owl web ontology language reference. world wide web consortium (w c) recommendation, . smith, d., ma, l., & ryan, n. ( ). acoustic environment as an indicator of social and physical context. personal and ubiquitous computing, ( ), – . soldatos, j., dimakis, n., stamatis, k., & polymenakos, l. ( ). a breadboard architecture for pervasive context-aware services in smart spaces: middleware components and prototype applications. personal and ubiquitous computing, ( ), – . soldatos, j., stamatis, k., azodolmolky, s., pandis, i., & polymenakos, l. ( ). semantic web technologies for ubiquitous computing resource management in smart spaces. international journal of web engineering and technology, ( ), – . stylianos, d., niki, p., kia, m., & christos, d. ( ). a context-aware cache structure for mobile computing environments. journal of systems and software, ( ), – . sumi, y., & mase, k. ( ). supporting awareness of shared interests and experiences in community. acm siggroup bulletin, ( ), – . sumi, k., & nishida, t. ( ). telme: a personalized, context-aware communication support system. ieee intelligent systems, ( ), – . sung, m., gips, j., eagle, n., madan, a., caneel, r., devaul, r., et al. ( ). mobile-it education (mit. edu): m-learning applications for classroom settings. journal of computer assisted learning, ( ), – . syukur, e., & loke, s. w. ( ). implementing context-aware regulation of context- aware mobile services in pervasive computing environments. international journal of web and grid services, ( ), – . syukur, e., & loke, s. w. ( ). policy based control of context aware pervasive services. journal of ubiquitous computing and intelligence, ( ), – . tentori, m., favela, j., & rodriguez, m. d. ( ). privacy-aware autonomous agents for pervasive healthcare. intelligent systems, ieee, ( ), – . trumler, w., bagci, f., petzold, j., & ungerer, t. ( ). amun-autonomic middleware for ubiquitous environments applied to the smart doorplate project. advanced engineering informatics, ( ), – . tyan, j., & mahmoud, q. h. ( ). a comprehensive service discovery solution for mobile ad hoc networks. mobile networks and applications, ( ), – . udugama, a., kuladinithi, k., gorg, c., pittmann, f., & tionardi, l. ( ). netcape: enabling seamless ims service delivery across heterogeneous mobile networks. communications magazine, ieee, ( ), – . van kranenburg, h., bargh, m. s., iacob, s., & peddemors, a. ( ). a context management framework for supporting context-aware distributed applications. communications magazine, ieee, ( ), – . van sinderen, m. j., van halteren, a. t., wegdam, m., meeuwissen, h. b., & eertink, e. h. ( ). supporting context-aware mobile applications: an infrastructure approach. communications magazine, ieee, ( ), – . vassilis, k., loukas, h., dimitris, k., & stavros, k. ( ). a dynamic context-aware access control architecture for e-services. computers & security, ( ), – . vuković, m., kotsovinos, e., & robinson, p. ( ). an architecture for rapid, on-demand service composition. service oriented computing and applications, ( ), – . wac, k., halteren, a. v., bults, r., & broens, t. ( ). context-aware qos provisioning in an m-health service platform. international journal of internet protocol technology, ( ), – . wang, x., dong, j. s., chin, c. y., hettiarachchi, s. r., & zhang, d. ( ). semantic space: an infrastructure for smart spaces. ieee pervasive computing, ( ), – . weber, o., sorkine, o., lipman, y., & gotsman, c. ( ). context-aware skeletal shape deformation. computer graphics forum, ( ), – . wilson, d., doyle, j., weakliam, j., bertolotto, m., & lynch, d. ( ). personalised maps in multimodal mobile gis. international journal of web engineering and technology, ( ), – . wohltorf, j., cissée, r., & rieger, a. ( ). berlintainment: an agent-based context- aware entertainment planning system. ieee communications magazine, ( ), – . wu, z., wu, q., cheng, h., pan, g., zhao, m., & sun, j. ( ). scudware: a semantic and adaptive middleware platform for smart vehicle space. intelligent transportation systems, ieee, ( ), – . xiong, l., & hassan, a. k. ( ). location awareness through trajectory prediction. computers, environment and urban systems, ( ), – . xu, y., fu, t. y., lee, w. c., & winter, j. ( ). processing k nearest neighbor queries in location-aware sensor networks. signal processing, ( ), – . yang, k., wu, y., yang, y., & liu, e. ( ). policy-based service-driven dynamic planning of heterogeneous wireless networks. international journal of mobile network design and innovation, ( ), – . yau, s. s., & karim, f. ( ). an energy-efficient object discovery protocol for context-sensitive middleware for ubiquitous computing. ieee transaction on parallel and distributed systems, ( ), – . yau, s. s., & karim, f. ( a). a context-sensitive middleware for dynamic integration of mobile devices with network infrastructures. journal of parallel and distributed computing, ( ), – . yau, s. s., & karim, f. ( b). an adaptive middleware for context-sensitive communications for real-time applications in ubiquitous computing environments. real-time systems, ( ), – . yee, g., korba, l., lin, n. h., & shih, t. k. ( ). context-aware privacy and security agents for distance education. international journal of high performance computing and networking, ( – ), – . yu, z., zhou, z., yu, z., zhang, d., & chin, c. ( ). an osgi-based infrastructure for context-aware multimedia services. communications magazine, ieee, ( ), – . yu, z., zhou, x., zhang, d., chin, c., wang, x., & men, j. ( ). supporting context- aware media recommendations for smart phones. ieee pervasive computing, ( ), – . yuan, s. t., & chen, f. y. ( ). ubisrvint – a context-aware fault-tolerant approach toward wireless p p service provision. expert systems with applications, ( ), – . yuan, s. t., & peng, k. h. ( ). location based and customized voice information service for mobile community. information systems frontiers, ( ), – . zhang, q., qi, y., zhao, j., hou, d., & niu, y. ( ). fuzzy privacy decision for context-aware access personal information. wuhan university journal of natural sciences, ( ), – . zhong, y., & gilbert, j. e. ( ). a context-aware language model for spoken query retrieval. international journal of speech technology, ( ), – . zhu, f., mutka, m. w., & ni, l. m. ( ). service discovery in pervasive computing environments. ieee pervasive computing, ( ), – . zimmermann, a., specht, m., & lorenz, a. ( ). personalization and context management. user modeling and user-adapted interaction, ( – ), – . http://www.w .org/tr/owl-ref context-aware systems: a literature review and classification introduction procedures the selection criteria data sources and procedures to extract articles the general characteristics of the articles classification of articles by publication year classification of articles by online database classification of articles by journal classification framework abstract architecture of context-aware systems classification framework of context-aware systems result of article classification concept and research layer network layer middleware layer application and service layer user infrastructure layer discussions and suggestions how to extract and use the cognitive context in context-aware application? what are design patterns of context-aware systems? which is the best inferring algorithm to extract user context and provide service to user? how to deal with concurrently enormous data, information and knowledge having different format to serve suitable service to users? how to extract the best solution when the context of users is conflicted? how to reflect the preference of users for satisfying user needs? how to save users information in context-aware systems? how to evaluate performance of context-aware systems? conclusions references course generation based on htn planning carsten ullrich german research center for artificial intelligence, dfki gmbh stuhlsatzenhausweg , d- saarbrücken, germany cullrich@dfki.de abstract a course generator automatically assembles learning objects retrieved from one or several repositories to a greater unit. different frame- works for course generation exists, but only a few allow the representation of sophisticated peda- gogical knowledge. in this paper, we describe course generation based on pedagogical tasks and methods, formalized in a hierarchical task network (htn) planner. we describe the mod- erate constructivist learning strategy we imple- mented in this framework, problems that arouse from using htn planning and their solutions. a first evaluation supports the practical value of our approach. introduction course generation automatically assembles learning ob- jects retrieved from one or several repositories to a greater unit, a course. the learning objects are not assembled ran- domly, but support the user to reach a learning goal. such goals require a sophisticated representation and us- age of pedagogical knowledge. although course genera- tion has been a topic of research since long, only a few frameworks exist that offer a representation of pedagogical knowledge powerful enough to encode a variety of generic pedagogical approaches. in this paper, we will describe our approach to course generation, which is based on hierarchi- cal task network (htn) planning. we will start by review- ing related work (section ). then, section provides a short introduction to htn planning. the main part of the paper describes and exemplifies the course generation as developed for leactivemath. related work early approaches on course generation date back to the eighties [peachy and mccalla, ; murray, ]. with the generic tutoring environment (gte), van mar- cke [van marcke, ; van marcke, ] introduced the separation between instructional tasks (representations of pedagogical activities) and instructional methods (repre- sentations of different ways of achieving the activities). most of this instructional knowledge is independent of the learning domain. vassileva [vassileva, ] build on this approach and added several layers of rules that handled dif- ferent aspects of course generation (e.g., selection of the main strategy or of the appropriate media types, etc.). subsequent course generation frameworks (e.g., [specht and oppermann, ; steinacker et al., ]) did not reach such a high level of representing pedagogical knowl- edge (or at least never provided sufficiently detailed de- scriptions). today’s course generation focuses less on pedagogical knowledge, but more on semantic web and metadata, e.g., on using ontologies of the subject domain to automati- cally calculate the best path through the learning mate- rial [karampiperis and sampson, ], or on calculating a learner specific ranking of and trail through learning ob- jects retrieved according to his query [keenoy et al., ]. these approaches use rather simplified pedagogical knowl- edge, e.g., to select those learning objects with the lowest typical learning time. however, to generate a course which is adapted to the individual learner’s goals and needs and which is based on state of the art pedagogical strategies re- quires more elaborate expertise. in a former project, we used an expert system (jess, [friedman-hill, ]) to represent elaborated course gen- eration knowledge as rules [ullrich, ]. however, this approach had severe scaling problems: in case more than users used the system simultaneously, the latency time increased over an acceptable amount. while technical op- timizations might have alleviated this problem, using rules made it difficult to use the hierarchical structure inherent in pedagogical knowledge. therefore, we decided to investi- gate a new approach based on htn planning. hierarchical task network planning in htn planning a planning problem is represented by sets of tasks (task networks); methods decompose non- primitive tasks into sub-tasks until a level of primitive tasks is reached, which can be solved by operators [erol et al., ]. because htn incorporates heuristic knowl- edge in the form of the decomposition methods, it is a very efficient planning technique. it also offers a rela- tively straight-forward way for representing human expert knowledge. the current version of our course generator uses the htn planner jshop [ilghami and nau, ], a domain-independent planner that compiles domain de- scriptions into domain-specific planners. figure contains an example of a method (for the time being, we will focus on its htn related aspects, the un- derlying pedagogical strategy will be explained in sec- tion . ). the method is applicable in case an open task ex- ists that matches with teachconcept ?c, with ?c be- ing a variable (indicated by the prefix ?). then, because the method has no preconditions (the empty parentheses in line two), the task will be decomposed in seven subtasks. the subtasks annotated with the prefix ! are primitive tasks, i.e., tasks that will not be further decomposed. (:method (teachconcept ?c) () ((!startsection newbook ?c) (introduce ?c) (develop_concept ?c) (practice ?c) (connect ?c) (reflect ?c) (!endsection) ) figure : an htn method for teaching a concept . challenges of htn planning htn planning faces several difficulties if employed in a web-based context. however, today most learning en- vironments are web-based. as an example, the compo- nents of the learning environment activemath [melis et al., ], e.g., the learner model, the learning object databases, and learning supporting tools such as a concept map [melis et al., ], are distributed over the web and available as web services. such a setting poses the follow- ing challenges: distributed content in a distributed environment, a course generator should be able to integrate content from several databases. this involves the difficulty that tradi- tional ai planning requires evaluating a method’s precondi- tions against the planner’s world state. however, this would require mirroring all the content stored in the repositories (or its metadata) in the world state. this is simply infeasi- ble. dynamic user information similar problems arise with respect to user information. similar to most systems ac- tivemath uses information about the learner stored in a learner model. part of this information, e.g., the knowl- edge state, is frequently updated. thus, the naive solution of adding the learner information in the world state of the planner is unpractical, especially as the learner model can cover a vast range of content and it is unknown exactly which of the information will be relevant. learning services a vast range of services that support the learning process in various ways have been developed, for instance tools that stimulate meta cognition [white and shimoda, ] or perform assessment [conejo et al., ]. a course should integrate these services, not ar- bitrarily but in a pedagogically sensible way: during the learning process, at specific times the usage of a tool will be more beneficial than at some other time. for instance, reflecting over the learned content may be most effective at the end of a lesson. additionally, access to these services may vary, depending on the configuration of the learning environment and their general availability. therefore, the planning needs to take into account dynamic information about these services. in the following, we will describe how we represent ped- agogical knowledge in an htn planning framework and how we overcome the above limitations. representing course generation knowledge in an htn planner . pedagogical background in leactivemath, the university of augsburg and the dfki are jointly developing moderate constructivist ped- agogical strategies that rely on the “programme for inter- national student assessment” (pisa) framework, [oecd, ]. instead of exclusively assessing curriculum based knowledge, pisa focuses on mathematical literacy and competencies. mathematical competencies are general mathematical skills such as problem solving, the use of mathematical language and mathematical modeling. in leactivemath, authors annotate exercises and exam- ples with the competencies they train (in addition to a sub- set of standard lom metadata [ieee learning technology standards committee, ]). the pedagogical strategies take these competencies into account: first, they ensure that every course contains a diversity of competencies. sec- ondly, they primarily select exercises of competency levels just outside the learner’s current mastery state. we identified six different pedagogical scenarios, each corresponding to specific needs and high-level learning goals of the user: learnnew, rehearse, overview, train- competency, workbook, and examsimulation. for in- stance, the scenario learnnew provides the student with an in-depth course that teaches him all necessary material to fully understand the target concepts right from the start. for each scenario, we formalized the pedagogical knowl- edge required to generate courses supporting the learner to achieve the respective learning goals. . course generation by htn planning course generation knowledge lends itself to being repre- sented hierarchically. additionally, as van marcke [van marcke, ] showed, by representing the pedagogical objectives as tasks and the ways of achieving the objects as methods, sophisticated pedagogical strategies can be en- coded. in htn planning, tasks are a basic principle, too. therefore, in a first step, we will bring together the notion of pedagogical and htn task. a pedagogical task corresponds to the goals a learner wants to achieve. it combines the two dimensions domain (content) and educational goal. formally, a task t is de- fined as a tuple t = (l, c), where l is a pedagogical objec- tive and c a unique identifier of a learning object (content goal). while c specifies the concept the course will primar- ily target, the pedagogical objective determines the kinds of learning objects selected for c. in jshop, a task is defined as a predicate symbol fol- lowed by several arguments. hence, a pedagogical task can be mapped to a htn task by mapping the pedagogical ob- jective onto the predicate symbol, and the content goal onto the arguments. because of this straightforward mapping, in the following we will no longer distinguish between peda- gogical and htn tasks. a top-level task that serves as a starting point of course generation is teachconcept id. the goal of this task is to assemble a structured sequence of learning objects that help the learner to understand the content goal c. us- ing different collections of tasks and methods (i.e., tutorial strategies), this task can be planed differently. hence, the task-based approach can serve to represent a variety of ped- agogical strategies. figure illustrates a method developed in leactive- math based on pedagogical principles. it describes how the task of teaching a concept is decomposed in the sce- nario learnnew. this decomposition follows a course of action in a classroom which implements constructivist ele- ments and distinguishes between different stages that occur while learning a new concept. similar solutions were de- veloped for other high-level learning goals. !startsection and !endsection are primitive tasks, which serve to provide additional structure. they open and close sections within a course. the introduction (introduce ?c) makes the aims of the learning process explicit. it provides an overview to the learner and makes the learning process more transparent. the stage of devel- oping a concept serves to provide the essential information about the concept to the learner, and will be explained in more detail in the next paragraph. the fourth subtask of teaching a concept trains the application of the concept. the final stages, connect and reflect, serve metacognitive purposes. they support the learner to put the concept in its context with respect to the other learning material, and help him to get an overview of his knowledge state. as a detailed example, consider the two methods for de- veloping a concept shown in figure . the upper method uses information about the user that is retrieved from a learner model and is applicable if the learner exhibits a high math anxiety. in that case, first the concept is inserted in the course via the primitive task !insertelement ?c. then, the insertion of an explanation and an example illus- trating the application of the mathematical concept aim at alleviating the anxiety. the bottom method is applied in case the first method does not match, and just inserts the concept itself. (:method (develop_concept ?c) ((learnerproperty anxiety ?an) (>= ?an )) ((!insertelement ?c) (explain ?c) (examplify ?c))) (:method (develop_concept ?c) () ((!insertelement ?c))) figure : two methods for developing a concept . critical and optional tasks course generation has to distinguish between critical and optional tasks. critical tasks represent necessary elements of the pedagogical strategy that have to be fulfilled for the course generation not to fail. for instance, in a problem- based approach it is mandatory to present a learning ob- ject is of the type realworldproblem, hence the cor- responding task insertproblem! is marked as criti- cal (indicated by the suffix !). in contrast, optional tasks should be achieved if possible, but failing to do so will still result in a course. as an example, the subtask explain of the upper method in figure is optional, e.g., if there are no learning object that can serve to explain concept c the concept will still be considered as developed. . ideal and fallback methods in practice, it is impossible to encode course generation knowledge that covers every potential combination of in- formation about the learner. oversimplified, if a learner (:method (selectappropriateexercise ?c) ;; ideal method: if motivation is ;; high, then select exercise with ;; higher competency level. (;; preconditions: (learnerproperty field ?field) (learnerproperty educationallevel ?el) (learnerproperty motivation ?m) (>= ?m ) (learnerproperty competencylevel ?c ?cl) (try-all-assign ?exercise (call getelements ((class exercise) (relation for ?c) (property learningcontext ?el) (property competencylevel (+ ?cl)) (property field ?field))))) (;; sub-task: (insertelement ?exercise))) (:method (selectappropriateexercise ?c) ;; alternative method: ;; select exercise from any field ( (learnerproperty allowededuclevel ?el) (learnerproperty competencylevel ?c ?cl) (try-all-assign ?exercise (call getelements ((class exercise) (relation for ?c) (property learningcontext ?el) (property competencylevel ?cl) )))) ((insertelement ?exercise))) figure : selecting an exercise model represents n different kinds of information with m possible values each, then the number of possible combina- tions is nm. writing rules or methods for each case would keep a lot of teachers busy for a long time. therefore, in leactivemath, methods encode best possible and alternative ways of selecting learning objects. an ideal method typically takes a larger number of user properties into account and poses a larger number of con- straints on the learning object. as an example, consider the methods in figure . they encode the pedagogical knowl- edge about exercise selection, and are called as a sub-task of practice. the upper method checks whether the user is highly motivated. then, it searches for an exercise that would be slightly too difficult under normal circumstances (competency level plus ). additionally, the learning con- text of exercise should correspond to the educational level of the learner, and, similar, its field should correspond to the field of the learner. in case no ideal method is applicable, fallback methods come into play. they encode the least constraining con- ditions on the learning objects possible and serve to insert elements in case no ideal method can be fulfilled. the bot- tom method of figure matches if any exercise for concept ?c exists that is of the correct competency level and has a learning context equal or lower as the educational level of the learner. . the output of course generation the result of the planning is a sequence of learning objects called course structure. similar to the organization element of an ims content package [ims global learning consortium, ], it consists of nested sections with the leaves being pointers to learning objects. because tasks represent a vast range of pedagogical goals, the size of the generated courses ranges from a sin- gle element to a complete curriculum. for instance, while the task teachconcept results in sequences of several learning objects, other methods may select only a single element. frequently occurring examples are the methods for exercise selection shown in figure and for example selection, shown in figure . the example selection is additionally interesting because of its differences to exercise selection. examples serve to increase motivation by presenting the application of a con- cept. therefore, the primary requirement is that the exem- plary application is taken from the field of interest of the learner, even in case its assigned competence level is higher than the current competence level of the learner. (:method (selectappropriateexample ?c) ;; ideal method: if motivation is ;; low, select example with ;; matching field. ( (learnerproperty field ?field) (learnerproperty educationallevel ?el) (learnerproperty motivation ?m) (call < ?m ) (learnerproperty competencylevel ?c ?cl) (equivalent ?cl ?ex_cl) (try-all-assign ?example (call getelements ((class example) (relation for ?c) (property learningcontext ?el) (property competencylevel ?ex_cl) (property field ?field) )))) ((insertelement ?example))) (:method (selectappropriateexample ?c) ;; fallback method: select example ;; from any competency level. ( (learnerproperty allowededuclevel ?el) (learnerproperty field ?field) (try-all-assign ?example (call getelements ((class example) (relation for ?c) (property learningcontext ?el) (property field ?field) )))) ((insertelement ?example))) figure : selecting an example . solutions to the challenges of htn planning in section . we described the problem that traditional ai planning requires evaluating a method’s preconditions against the planner’s world state, which is practically im- possible in case the learning objects are stored in one or several databases. therefore, we extended the htn plan- ner so that queries about learning objects in a method’s pre- conditions result in a call to a mediator. mediators are well known services used in information integration (for a recent overview on this topic, see [doan et al., ]). they act as links between the knowledge processing and knowledge storing components, and offer a uniform query interface to a multitude of autonomous data sources. the leactivemath mediator passes incoming queries to the available databases and combines the results. jshop ’s call command allows integrating exter- nal function calls during planning. the function call getelements conditions (e.g., in figure ) queries the mediator to search the repositories for elements that fulfill the given conditions. try-all-assign ?p ?identifiers generates all bindings of ?p to the el- ements of the returned identifier list. thereby subsequently all possible values can be tried when backtracking. similarly, preconditions involving information about the learner result in queries to the learner model. these queries are encapsulated in a planning axiom as shown in figure . axioms are used to infer preconditions not directly asserted in the world state. in the figure, the result of the query is bound to the variable value using the additional ax- iom same. the axiom same returns true in case both provided parameters are instantiated and equal, and false otherwise. if one parameter is uninstantiated, the axiom assigns the value of the instantiated parameter to the unin- stantiated parameter. (:- (learnerproperty ?property ?value) (same ?value (call querylm ?property))) figure : querying the learner model the integration of learning services happens by service- adding methods. these methods encode the pedagogical knowledge at what time in the course the learner should use a tool and insert calls to the tool at the appropriate place in the course structure. later, when the learner navigates through the course, an invitation to use the service is pre- sented to her. checking for the availability of a service is done by introducing an external function, too. thereby a method’s preconditions can test for the availability of a ser- vice. in case the service is available, its call is inserted in the course. otherwise, an alternative method is applied. this way, the pedagogical knowledge remains reusable, re- gardless of the actual configuration of the learning environ- ment. (:method (assess) ;; if siette is available, ;; then use it for assessment ((call serviceavailable siette)) ((!insertservicecall siette))) (:method (assess) ;; fallback: insert some exercises () ((inserttext assessment) (insertexercises ))) figure : integration of an assessment service figure illustrates the integration of the assessment tool siette [conejo et al., ]. the upper, ideal method checks whether siette is available and if so inserts a call to the service in the course structure. in the fallback case, the assessment will be simulated manually. a short text is added in the course that describes that the following ex- ercises will assess his knowledge, and five exercises are inserted. first results and conclusions the results reported in this section focus on technical as- pects. evaluations regarding learning gains are underway and will be reported at a later time. we performed a first comparison between the old, expert system based (exscg) and the new course generator (ht- ncg). the tests and all necessary components (active- math server and a database) ran on a . gh intel pentium cpu with gb ram. both course generators repeatedly generated an average length course about the mathematical concept derivation. exscg showed a slightly slower performance for sin- gle user course generation (average time of and sec- onds). however, the exscg course generation did not ter- minate in case of the simultaneous generation of more than users. htncg slowed down, but still terminated (aver- age of minutes). we suspect the differences to be caused by the on- demand retrieval of the metadata. in exscg, the meta- data of the complete content was loaded in the fact-based of the expert system and then the rules were evaluated. in htncg, only content relevant for the current plan was re- trieved if necessary, i.e., when queried. it is interesting that the new approach is faster even though an additional indirect step via the mediator was in- serted. in the future, we plan to extend the mediator with caching possibilities. first of all, it can optimize queries and secondly avoid queries over the web. an additional evaluation compared an xml-rpc based connection and a direct connection between the mediator and one of our database (based on lucene [foundation, ]). because of the overhead of processing xml-rpc, we expected the direct connection to outperform xml- rpc. surprisingly, the direct connection proved not to be much faster. a possible cause may be the limited number of queries (about ). we plan to further investigate on this topic. however, a significant speed-up was achieved by using a cached version of lucene database. there, in case the very same query is posed more than once, the result elements were not retrieved from index but directly from memory. course generation for a single user decreased to the average time of three seconds, which is about factor faster than compared to the non-cached version. the speedup holds in case of user planning, which took about seconds. this speedup is even more surprising in regard of the fact that due the early stage of implementation the cache was local to a single user, which means instead of one “global” cache, the system used up to “local” caches. we expect the speedup to be caused mainly be the large amount of similar queries, which occur even in a single course genera- tion. for instance, if several examples for the same concept are inserted in a course, a large number of the same queries are processed. in this paper, we described how htn planning can be used for representing and executing course generation knowledge that is based on tasks and methods. besides the described advantages, we plan to use tasks to describe the service offered by a course generator web service. addi- tionally, they may be used for communication between user and system, for instance by providing the user the possibil- ity of executing tasks on demand with respect to a given course and thereby intelligently extending the course. in- vestigating these possibilities will be our next work. acknowledgments the work described in this publication is implemented in the learning environment activemath . activemath serves as the technical basis used in the leactivemath project , where a consortium of eight european partners is developing an innovative third generation elearning sys- tem for high school, college, and university level class- rooms. the system adapts to the learner and learning con- text and comprises personalization, tutorial dialogues, open student modeling and interactivity that is tool-supported for active and exploratory learning. the leactivemath project is funded under the th framework programm of the european community - (con- tract nr ist- - ). the author is solely respon- sible for its content, it does not represent the opinion of the european community and the community is not re- sponsible for any use that might be made of data appearing therein. references [conejo et al., ] r. conejo, e. guzman, e. millan, m. trella, j. l. perez de-la cruz, and a. rios. siette: a web-based tool for adaptive teaching. international journal of artificial intelligence in education (ijaied ), : – , . [doan et al., ] anhai doan, natalya f. noy, and alon y. halevy. introduction to the special issue on se- mantic integration. sigmod rec., ( ): – , . [erol et al., ] k. erol, j. hendler, and d. s. nau. htn planning: complexity and expressivity. in proceedings of the twelfth national conference on artificial intelli- gence (aaai- ), volume , pages – , seattle, washington, usa, . aaai press/mit press. [foundation, ] the apache software foundation. jakarta lucene, july . [friedman-hill, ] e. friedman-hill. jess, the java ex- pert system shell. technical report sand - , sandia national laboratories, . [ieee learning technology standards committee, ] ieee learning technology standards committee. . . - ieee standard for learning object metadata, . [ilghami and nau, ] okhtay ilghami and dana s. nau. a general approach to synthesize problem-specific planners. technical report cs-tr- , department of computer science, university of maryland, october . [ims global learning consortium, ] ims global learning consortium. ims content packaging informa- tion model, june . [karampiperis and sampson, ] p. karampiperis and d. sampson. adaptive instructional planning using on- tologies. in proc. of the th ieee international confer- ence on advanced learning technologies, icalt , pages – , . [keenoy et al., ] k. keenoy, a. poulovassilis, v. christophides, p. rigaux, g. papamarkos, a. magka- naraki, m. stratakis, n. spyratos, and p. wood. personalisation services for self e-learning networks. in http://www.activemath.org http://www.leactivemath.org http://www.activemath.org http://www.leactivemath.org n. koch, p. fraternali, and m. wirsing, editors, proc. of th international conference on web engineering, volume of lecture notes in computer science, pages – . springer, . [melis et al., ] e. melis, e. andrès, j. büdenbender, a. frischauf, g. goguadze, p. libbrecht, m. pollet, and c. ullrich. activemath: a generic and adaptive web-based learning environment. international journal of artificial intelligence in education, ( ): – , . [melis et al., ] e. melis, p. kärger, and m. homik. in- teractive concept mapping in activemath. in delfi , . [murray, ] w. r. murray. control for intelligent tutor- ing systems: a blackboard-based dynamic instructional planner. in d. bierman, j. breuker, and j. sandberg, editors, proc. th international conference of ai and education, pages – , amsterdam, springfield va, tokyo, . ios. [oecd, ] oecd, editor. organisation for economic co-operation and development: measuring student knowledge and skills – a new framework for assess- ment. . [peachy and mccalla, ] d. r. peachy and g. i. mc- calla. using planning techniques in intelligent tutoring systems. international journal of man-machine stud- ies, : – , . [specht and oppermann, ] m. specht and r. opper- mann. ace - adaptive courseware environment. the new review of hypermedia and multimedia, : – , . [steinacker et al., ] a. steinacker, c. seeberg, k. re- ichenberger, s. fischer, and r. steinmetz. dynamically generated tables of contents as guided tours in adaptive hypermedia systems. in proceedings of the edmedia & edtelecom, pages – , june . [ullrich, ] c. ullrich. pedagogical rules in active- math and their pedagogical foundations. seki report sr- - , universität des saarlandes, fb informatik, . [van marcke, ] k. van marcke. instructional exper- tise. in c. frasson, g. gauthier, and g.i. mccalla, editors, proceedings of the second international con- ference on intelligent tutoring systems, number in lecture notes in computer science, pages – . springer, . [van marcke, ] k. van marcke. gte: an epistemo- logical approach to instructional modelling. instruc- tional science, : – , . [vassileva, ] j. vassileva. dynamic courseware gen- eration: at the cross point of cal, its and authoring. in proceedings of icce’ , singapore, pages – , december . [white and shimoda, ] b.y. white and t.a. shi- moda. enabling students to construct theories of collab- orative inquiry and reflective learning: computer sup- port for metacognitive development. international jour- nal of artificial intelligence in education, : – , . introduction related work hierarchical task network planning challenges of htn planning representing course generation knowledge in an htn planner pedagogical background course generation by htn planning critical and optional tasks ideal and fallback methods the output of course generation solutions to the challenges of htn planning first results and conclusions wp-p m- .ebi.ac.uk params is empty sys_ exception wp-p m- .ebi.ac.uk no params is empty exception params is empty / / - : : if (typeof jquery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/ . . /js/jig.min.js"][/script]'.replace(/\[/g,string.fromcharcode( )).replace(/\]/g,string.fromcharcode( ))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} page not available reason: the web page address (url) that you used may be incorrect. message id: (wp-p m- .ebi.ac.uk) time: / / : : if you need further help, please send an email to pmc. include the information from the box above in your message. otherwise, click on one of the following links to continue using pmc: search the complete pmc archive. browse the contents of a specific journal in pmc. find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/med/ role of expert systems in identification and overcoming of dengue fever (ijacsa) international journal of advanced computer science and applications, vol. , no. , | p a g e www.ijacsa.thesai.org role of expert systems in identification and overcoming of dengue fever dr. nadeem ahmed department of computer science superior university lahore, pakistan muhammad shoaib department of computer science sharif college of engineering and technology lahore, pakistan adeed ishaq department of computer science & it the university of lahore lahore, pakistan abdul wahab department of computer science superior university lahore, pakistan abstract—this paper presents a systematic literature review on expert systems which are used for identification and overcoming of dengue fever. dengue is a viral disease produced by flavivirus. the expansion of dengue fever is because of uncontrolled population and urbanization without suitable water administration. with the quick technological enhancement, we can identify and overcome dengue fever by using expert systems. these expert systems require knowledge of the relevant problem and techniques to infer the result in order to make decisions. the literature review provides a comparison of techniques, methodologies, limitations and advantages of different dengue expert systems. these expert systems facilitate both doctors and patients in dengue detection. multiple risk factors can be eliminated by the detection of dengue fever through expert systems at early stages of dengue. furthermore, we find that enhancement of knowledge base improves accuracy of expert systems. keywords—expert system; rule based; dengue; health care; disease; fever i. introduction dengue fever is one of the viral diseases around the world that is caused by blood feeding mosquito. dengue fever happens due to a single stranded rna flavivirus, which is transmitted by mosquito bite [ ], [ ]. it is one of the most emerging diseases of the world [ ]. there are about - million cases annually and majority of these are diagnosed as flu that causes about , deaths every year. due to lack of proper treatment around . % of the infection results to be fatal. the symptoms of this fever include joint pain, headache, muscle-pain and skin rashes [ ]. dengue ranks as the most important mosquito-borne viral disease in the world [ ]. countries which are most effected by dengue are more than in number [ ]. mainly this is due to bad sanitary conditions in underdeveloped countries. the experts provided a number of solutions regarding dengue diagnosis but due to inexperienced doctors it is still misclassified [ ]. thus, throughout the world the experts are facing the problem of diagnosis of dengue due to its misclassifications. the diagnosis process is quite time consuming which make things more complex. much time is required for the identification of its correct classification. there is no problem in diagnosis if the number of patients is less but if they are huge in numbers and the country has limited resources then it becomes problematic to correctly diagnose this fever. the dengue fever issue motivates computer experts to devise an automated and time efficient solution that correctly classify the dengue. these automated systems for dengue diagnosis can be deployed in faraway places as well as in epidemic situations to save precious lives. in computer science those systems that can solve a specific problem and can reason rationally are known as expert systems. expert systems require knowledge of the regarding problem and techniques to infer the result in order to make decision [ ]. today there exists a number of expert systems for disease diagnosis as for bacterial infection mycin, for lung disease puff, for angiograms angy, for edema pip, for meningitis diagnosis neomycin, for chest pain med , for bacterial infection guidon and for hematological disease heme [ ]- [ ]. there are several types of expert systems based on knowledge base, i.e. rules, cases, hybrid approach (rules and cases) and hybrid approach with ann like medical expert solution (mes) system [ ]. the mes uses a knowledge base which consists of two knowledge structures, named as symptom and diseases. the mes provides medical facilities to those users who don’t have access for it. further, we will discourse research methodology, different mosquito borne diseases, related work, discussion, conclusion and future work. ii. research methodology a. rsearch question rq : does expert systems correctly identify dengue fever? rq : does knowledge base of expert system fulfill the requirements? rq : does interfaces of expert systems are user friendly for all type of users? (ijacsa) - international journal of advanced computer science and applications vol. , no. , | p a g e www.ijacsa.thesai.org b. search strategy after observing and viewing all aspects of expert systems a research has been carried out on expert systems for dengue. the field of computer science plays a major role in this perspective because it provides us the tools and techniques for gathering data from massive data stores/sources. ) identification of search terms: from research question extract major terms and check their synonyms for search related research papers. these terms verified in relevant papers. or operator use for concatenation of synonyms words. and operator use for concatenation of major terms. ) trial search: trial search conducted by using the following search string in science direct, ieeexplore, springerlink and acm digital library. table contains complete information of digital libraries with search items and search conduction date. (“dengue” or “fever” or “malaria" or “rule based” or “artificial neural network”) and (“identification” or “overcoming” or “expert system” or “task” or “factors” or “challenges” and or risks or problems or “issues”) table i. trail search results s. no. digital library search conduction date ieee explore january acm digital library january google scholar january science direct january site seer digital library january springer link january c. study selection criteria some of the papers are relevant to our research and properly defining our research scope. these papers have been included here that helped in the research and those papers which are not relevant to our research, not properly defining our research scope and their language is different, are excluded. ) inclusive criteria: after viewing different papers some papers are, giving the answer of our research questions, defining our search terms, their language is english and within our scope, these papers are included. ) exclusion criteria: some papers are not within a scope or not written in the english language which is prescribed in this paper, or not giving proper answers of our research questions, such type of papers is excluded. d. data extraction and assessment of study quality as our inclusive criteria is defined, we have gathered the data which is relevant to our research, fulfilling our requirements and defining the scope of our research, also it will guide us for quality study. e. electronic data stores we search related data from the following electronic data sources:  www.ieeexplore.ieee.org  www.acm.org  www.sciencedirect.com  www.citeseer.ist.psu.edu  www.scholar.google.com  www.springerlink.com fig. . research methodology. fig. shows complete research process and number of research paper included in our study. duplicate publication and non-primary research studies exclude. iii. different mosquito borne diseases here we describe the origin of different mosquito born disease. these diseases have some common symptoms with others. some are viral while other are non-viral. viral diseases are either due to mosquitos like yellow fever, malaria, dengue and influenza [ ], whereas non-viral are natural diseases such as heart diseases, diabetes and cancer. unlike non-viral diseases, the viral diseases can spread in short spans of time and are difficult to handle. fig. expresses hierarchical structure of viral and non-viral diseases. (ijacsa) - international journal of advanced computer science and applications vol. , no. , | p a g e www.ijacsa.thesai.org fig. . viral and non-viral diseases. a. diseases due to mosquitos these diseases are caused by viruses that are transmitted by mosquitos. which includes dengue, yellow fever and malaria [ ]. in past years’ research shows that there is leap in diseases caused by mosquitos. this motivates researchers and scientists to work to reduce these diseases [ ]. fig. presents complete detail of mosquitos borne diseases. ) yellow fever: it is one of the infectious diseases which is caused by aedes mosquito that is infected by flavivirus. this mosquito is usually found in sub-tropical or tropical regions [ ]. signs and symptoms: yellow fever happens due to a virus cause by mosquito. the symptoms appear in to days after mosquito bite. there are three stages of this disease.  stage (infection): the infected person shows following symptoms in first to days, fever, loss of appetite, headache, joint pain, vomiting and jaundice. after day these symptoms become brief.  stage (remission): in this stage, most of the symptoms are gone and people can be recovered but due to carelessness the situation can get even worst.  stage (intoxication): in organs such as liver, heart and kidney different problems may occur [ ], [ ]. ) malaria: malaria is also a disease that happens due to virus by a female mosquito plasmodium of genus anopheles carry. if this disease is not treated in time it may lead to complications towards death. there are two types of malaria in general: simple and severe. simple malaria is curable if treated within time while severe malaria may lead to death. fig. . mosquitos borne diseases. signs and symptoms:  simple malaria: this disease remains for to hours. its symptoms are, cold stage that has sense of shivering, hot stage with headache and vomiting and sweating stage. attacking schedule is with tertain parasite, it occurs every second day & with quartan parasite it occurs every third day. general symptoms are: fever, headache, sweating, vomiting, body ache and chills. the physical impact of malaria can be seen on the infected body that includes, getting sweaty, temperature rise up, mild jaundice, respiration increased and weakness in body.  severe malaria: during simple malaria if there arise a complication of infection in any part of body by failure of any organ it may create severe malaria. ) dengue: aedes aegypti is certain type of mosquitos that causes of dengue. it is a viral disease and the virus is single-positive strand rna virus and commonly found in sub- tropical and tropical regions [ ]. signs and symtoms: major symptoms of dengue are high-fever and combination of at least two which are, pain in joints, bleeding gums, rashes on skin, very low white-blood cells, pain in bones and severe pain behind eyes. to get better a proper lookout is very necessary because after to days following things can happen, red spots on body, pale skin, drowsiness, pain in abdomen, breathing in difficult. dengue fever that stays for to days is normally called dengue hemorrhagic fever (dhf). dhf warning appears when after to hours fever suddenly drops and at this time bleeding is started, due to this circulatory system is disturbed and it may lead to death [ ]. world health organization (who) has worked on the dengue risk factors that play the pivotal role in the disease diagnosis [ ]. similarly, resources have showed that dengue is assorted disease, its symptoms are mixed with other diseases that make its detection and diagnosis more difficult [ ]. iv. related work a. expert system using ann in this paper, som (self organization map) was used to distinguish between healthy and dengue infected patients. the research consists of two maps, one that contain bia (bio impedance analysis) factors and other contains disease symptoms. the research resulted these major bia factors; abdominal pain, anorexia, bleeding and rashes that were present in dengue patients. after bia analysis, they found -unit best map size with error of . and quantization error of . [ ]. goal was to predict occurrence of dengue by comparing weights. the learning rate varied from . to . [ ]. table demonstrates advantages and disadvantages of expert systems by using artificial neural network. (ijacsa) - international journal of advanced computer science and applications vol. , no. , | p a g e www.ijacsa.thesai.org table ii. advantages and disadvantages of expert system using ann advantages disadvantages reprogramming is not needed proper training is needed to operate tasks are completed by adjusting weights large data, high processing time learning is by hit and trail complexity is large if data is big easy to use and learn by examples lack in history maintenance perform functions that a linear system can’t perform highly dependent on data b. emerge this is rule-based expert system which was developed to assist in emergency room. this system uses weighting factors determined by neural network. it is composed of input & output block and a hidden layer block that communicates input to output. neural network learns from example & predict an output based on this knowledge. it uses if-then-unless instead of if-then that make more precise decision [ ]. c. expert system using rule based reasoning here, dengue was detected using rule based classifier that are: decision tree, naïve bayes & associative classifiers. the research proved that result from more than one classifier is better than results by using one classifier. the result jumped up to % by using more than one classifier thus giving better result than single classifier [ ]. in this paper, a dengue diagnosis system was proposed that observe medical record procedures. it uses tree like model rather than df and dhf. out of the record of infected were classified df, were dhf-i and were dhf-ii [ ]. table provides different advantages and disadvantages of expert systems by using rule based reason. table iii. advantages and disadvantages of expert system using rule based reason advantages disadvantages representation of knowledge is natural knowledge range is narrow knowledge is kept separately cost effective expert system training is used to increase skills no creative solutions there is no overhead no proper rules additional help is involved can’t learn itself d. mycin mycin was one of the earliest diagnosis systems that used to diagnose infectious diseases caused by bacteria. first it describes the infection produced by bacteria then, it suggests antibiotic to give the infected person based on the decision [ ]. e. pneumoconiosis x-ray diagnosis expert system this expert system was developed by miriam kubiska and julie herzner in . the system has inference engine to examine the shadows on the x-ray. these shadows are used to determine the degree and type of pneumoconiosis that is a lung disease. system includes the following:  the knowledge base that contain data of x-ray representation.  the interface that details the conclusions.  knowledge acquisition mode that allow user to add or change information. f. xdis xids had information of more than - internal diseases and syndromes that are frequently met in practice which help experts in diagnosis. the system has instructions for each disease, if disease is curable the system will prescribe medicine but if it is unable to identify the disease it will refer to a specialist. this process takes no more than minutes [ ]. g. expert system using integration of rule base and ann there is also a hybrid technique in which inference engine is updated and maintained whenever a new rule come. the purpose of new rule is to refining of time complexity and efficiency of symbolic rule. the authors explained about neural composition and how they can be shaped directly from results of experiments. also, they explained the pattern based methodologies in comparison with rule based cases [ ]. the expert system was also available on mobile application. the data represented in form of decision tree & cases were stored in table in symbolic form. they maintained database for cases and rules which show that hybrid approach make system faster and always available due to web [ ]. h. role of ultrasound in dengue detection in this paper, the authors perform a study using ultrasound rays to find that can be used to diagnose dengue fever and also to be useful in predicting the severity of this disease. to test the study people were taken as a subject that suffer dengue disease. people were identified to be positive case of dengue, remaining underwent ultrasound again until those who were infected were differentiated from those who were not infected by dengue virus. they found that after to days of ultrasound effect their gall bladder wall that got thick and had pleural effusion and ascites. they concluded that in epidemic of dengue ultrasound can be used for diagnosis of dengue fever [ ]. table presents detail comparison of various expert systems and also express future guidelines for new researches. (ijacsa) - international journal of advanced computer science and applications vol. , no. , | p a g e www.ijacsa.thesai.org table iv. comparison of different expert systems proposed by different researchers expert system targeted diseases expert system technique expert system methodology future guidelines advantage limitations medical expert solution (mes) system [ ] tropical diseases like dengue fever , malaria . inference engine . knowledge base mes consist of two main sections: . user-interface: this module used for displaying information and perform user interaction with system . expert system: it consists of inference engine and knowledge base. knowledge base consist of two elements, diseases and their symptoms. mes use this data for diagnosis the diseases . enhance knowledge base for getting accurate results. . beneficial for those people who don’t have excess to medical facilities. . provides first aid facility before going to medical consultant. . reduce physicians work. . it can be used in epidemics and war environment. . training required to use the system. adaptive neuro-fuzzy inference system (anfis) [ ] dengue , dengue fever , malaria . fuzzy inference system . hybrid learning algorithm for anis . optimization of anis structure . optimization of hybrid learning training parameters . evaluation of optimal models performance . develop such kind of system which can be used by physicians in hospitals or clinics. . diagnose risk in dengue patients. . it has better performance as compare to noninvasive system which is developed for diagnosis risks in dengue patients [ ]. . diagnostic accuracy achieved up to . %. . more robust and reliable than ibrahim et al. [ ] . not practically implemented yet. . no module for physicians. adaptive- network-based fuzzy inference system (anfis)[ ] dengue , dengue fever . fuzzy if-then rules or fuzzy conditional statements . fuzzy-rule- based systems . architectures and learning algorithms . architecture and basic learning rule . batch (off-line) learning . pattern (on-line) learning . adaptive-network-based fuzzy inference system . anfis architecture . hybrid learning algorithm . fuzzy inference systems with simplified fuzzy if-then rules . the system proposed new fuzzy models for data classification and feature extraction. .fixed structure of anfis. . training issue regarding overlapping between adjacent membership functions and minimal uncertainty. an application of expert system (aexs) [ ] varicella (chickenpox), dengue fever hands-foot- and-mouth disease epidemic parotitis (mumps) .forward chaining . rule-base system . user-interface . inference mechanism . explanation facility . knowledge acquisition facility . construct aexs for large computer program. . web portal design for easy excess of user anywhere. . it is important for diagnosis of viral infections. . provide feedback in explanation component and inference engine. . complete system with experimental results. . user can only select predefined symptoms. . lack of self- inference ability. web-based patient support system[ ] dengue, diabetes . artificial intelligence in medicine . fuzzy logic . neural network .centralized databases and www . web-based medical diagnosis and prediction . . user-interface . . predication modules . . diagnosis modules . . database . developed distributed databases for improving quality of treatment and providing better diagnosis. . artificial intelligence used to make consultancy more interactive. . fuzzy logic used in uncertainty conditions. . ann produce better results. . provide tele-health care. . not deployed in working in environment. . lack of experimental study. rule based expert system[ ] dengue infections .bioelectrical impedance analysis (bia) . development of a rule based expert system . identifying the problem . tree diagram . rule based expert system . architecture of the expert . applying nonlinear modeling technique for improving classification . perform risk classification in df and dhf. . do not recognize complex patterns. (ijacsa) - international journal of advanced computer science and applications vol. , no. , | p a g e www.ijacsa.thesai.org system . structure of the patient database accuracy like ann. predicting arboviral disease emergence using bayesian networks [ ] dengue virus . bayesian belief network (bbn) . framework and process of bbn risk modelling and mapping . input data processing and classification . climatic parameters . road and railroad density . seaports and airports . human population density . frequency of denv introduction . inclusion of endemicity risk node . infection risk under current summer and winter climates . scenario modelling . sensitivity analysis of risk node . case study done on a large region and get better results for predicating dengue virus risk. . by using this model other diseases like chikungunya, yellow fever, and zika virus can be predicted. . identify dengue risk in different environments. . dengue virus risk model developed which predict dengue risk in particular condition. . case study perform only for specific region. the integrated management of health care strategies and differential diagnosis [ ] malaria, typhoid fever, cough . combination of integrated management of child illnesses (imci) and health information systems (his). . action and disease oriented differential diagnosis of malaria and typhoid fever . knowledge engineering . imci/his expert system interface architecture . system evaluation . adaptive interfaces can be designed for urban and rural area users. . provide support for distributed environments. . used for diagnosis malaria and typhoid fever. . used in rural area health care where maximum people are illiterate. . provide user-interface for user interaction. . reduce child mortality rate in rural areas. . it is stand- alone application which cannot be run in distributed environments. v. discussion and conclusion dengue is a viral disease produce by flavivirus. it has two forms: one is dengue fever and second is dengue hemorrhagic fever. dhf is more severe form of dengue. it has become worldwide major problem in recent years. the expansion of dengue fever is because of uncontrolled population and urbanization without suitable water administration. if infected female aedes mosquitos bite human than dengue virus transmitted in human blood. in computer science, systems that can solve a specific problem and can reason rationally are known as expert systems. these systems require knowledge of the relevant problem and techniques to infer the result in order to make decision. in recent years, different researchers have proposed different expert systems for identification, prescribe medicine and overcoming dengue disease. these expert systems use different technique for getting better result like bayesian belief network (bbn), artificial neural networks (ann), fuzzy inference and hybrid learning algorithm for anis. for implementing these techniques different researchers have proposed different methodologies for determining dengue patient’s problems. the main common strategies in these methodologies are user-interface, knowledge base engineering process and system evaluation process. user- interface works as bridge between system and user. during designing of user interfaces careful analysis of user tasks and context must be done [ ]. we suggest that user interfaces should be well formed to convey knowledge according to user mental model so that a friendly communication may be done between user and system. mostly knowledge base consist of two elements, diseases and their symptoms which can be used for determining disease and predict medical suggestions related to particular disease. we find during study that knowledge base has limited expressive power. knowledge base should update itself according to environment as if it faces a new symptom which is not added in knowledge base than it should automatically be added in the knowledge base. knowledge base should be supported for distributed environments. system evaluation used for checking the performance of expert system. a generic mechanism required for evaluation which can be followed by every researcher during designing of expert systems. globally define certain set of rules to evaluate the expert system. vi. future work furthermore, we can enhance knowledge base of expert systems for getting accurate results. such kind of system should be developed which can be used by doctors and patients in web portal. distributed databases should also be developed for improvement of quality of treatment and to provide better diagnosis. the expert system should be updated as it may be used for other diseases like chikungunya, yellow fever, and zika. adaptive interfaces should be designed for both literate and illiterate users which would help them to use these systems more easily and efficiently. online assistance provided to semiliterate users during use of expert system can enhance their performance [ ]. ahmad, n. [ ] introduced a new approach which use a virtual character for providing assistance to barely literate or deaf people. by using such kind of support, we can enhance expert system utility by offering additional services like automatic translation from text to sign language. (ijacsa) - international journal of advanced computer science and applications vol. , no. , | p a g e www.ijacsa.thesai.org finally, autistic people are very sensitive and information technology is very useful for resolving their problem [ ]. therefore, expert system can be designed for autistic people because they also don’t know how to react in this situation. references [ ] gubler, d. j. ( ). dengue and dengue hemorrhagic fever. clinical microbiology reviews, ( ), - . [ ] gubler, d. j., ooi, e. e., vasudevan, s., & farrar, j. (eds.). ( ). dengue and dengue hemorrhagic fever. cabi.. [ ] fajardo-dolci, g., meljem-moctezuma, j., vicente-gonzález, e., venegas-páez, f. v., mazón-gonzález, b., & aguirre-gas, h. g. ( ). [the dengue fever in mexico. knowledge for improving the quality in health]. revista médica del instituto mexicano del seguro social, ( ), - . [ ] sulehri, m. a., hussain, r., & gill, n. i. ( ). dengue fever its diagnosis, treatment, prevention and control. gomal journal of medical science, , - . [ ] world health organization. ( ). global strategy for dengue prevention and control - . [ ] ranjit, s., & kissoon, n. ( ). dengue hemorrhagic fever and shock syndromes. pediatric critical care medicine, ( ), - .` [ ] salman, a., lina, y., & simon, c. ( ). computational intelligence method for early diagnosis dengue haemorrhagic fever using fuzzy on mobile device. in epj web of conferences (vol. , pp. - ). [ ] ashwinkumar, u. m., & anandakumar, k. r. ( , january). a web- based patient support system using artificial intelligence to improve health monitoring and quality of life. in second international conference on advanced computing & communication technologies (pp. - ). ieee. [ ] kumar, k. a., singh, y., & sanyal, s. ( ). hybrid approach using case-based reasoning and rule-based reasoning for domain independent clinical decision support in icu. expert systems with applications, ( ), - . [ ] tomar, p. p., & saxena, p. k. ( ). architecture for medical diagnosis using rule-based technique. in first int. conf. on interdisciplinary research & development, thailand (vol. , pp. - ). [ ] national pesticide information center, diseases transmitted by mosquitoes,npic. http://npic.orst.edu/pest/mosquito/diseases.html [ ] caraballo, h., & king, k. ( ). emergency department management of mosquito-borne illness: malaria, dengue, and west nile virus. emergency medicine practice, ( ), . [ ] world health organization. ( ). global strategy for dengue prevention and control - . [ ] yellow fever: causes, symptoms & diagnosis - healthline http://www.healthline.com/health/yellow-fever [ ] yellow fever: nursing - wolters kluwer health vol. , no. , p. , . [ ] ward, a. m., gunaratne, j., & garcia-blanco, m. a. ( ). identification of dengue rna binding proteins using rna chromatography and quantitative mass spectrometry. dengue: methods and protocols, - . [ ] simmons, c. p., farrar, j. j., van vinh chau, n., & wills, b. ( ). dengue. new england journal of medicine, ( ), - . [ ] hadinegoro, s. r. s. ( ). the revised who dengue case classification: does the system need to be modified? paediatrics and international child health, (sup ), - . [ ] farrar, j. j., hien, t. t., horstick, o., hung, n. t., jaenisch, t., junghanns, t., & simmons, c. p. ( ). dogma in classifying dengue disease. the american journal of tropical medicine and hygiene, ( ), - . [ ] faisal, t., ibrahim, f., & taib, m. n. ( , august). analysis of significant factors for dengue infection prognosis using the self- organizing map. in th annual international conference of the ieee engineering in medicine and biology society (pp. - ). ieee. [ ] husin, n. a., mustapha, n., sulaiman, m. n., & yaakob, r. ( , september). a hybrid model using genetic algorithm and neural network for predicting dengue outbreak. in th conference on data mining and optimization (dmo) (pp. - ). ieee. [ ] bakar, a. a., kefli, z., abdullah, s., & sahani, m. ( , july). predictive models for dengue outbreak using multiple rulebase classifiers. in electrical engineering and informatics (iceei), international conference on (pp. - ). ieee. [ ] ibrahim, f., taib, m. n., sulaiman, s., & abas, w. w. ( ). dengue fever (df) and dengue haemorrhagic fever (dhf) symptoms analysis from an expert system perspective. in multi topic conference, . ieee inmic . technology for the st century. proceedings. ieee international (pp. - ). ieee. [ ] suwa, m., scott, a. c., & shortliffe, e. h. ( ). an approach to verifying completeness and consistency in a rule-based expert system. ai magazine, ( ), . [ ] adekoya, a. f.@l,akinwale,a. t. and oke, o. e. . a medical expert system for managing tropical diseases. department of computer science, university of agriculture, abeokuta, nigeria adekoyaaf@unaab.edu.ng www.journal.unaab.edu.ng/index.php/colnas/article/download/ / [ ] prentzas, j., & hatzilygeroudis, i. ( ). rule-based update methods for a hybrid rule base. data & knowledge engineering, ( ), - . [ ] sulistiarini, e. b. ( ). representing knowledge base into database for wap and web-based expert system. arxiv preprint arxiv: . . [ ] sai, p. v., dev, b., & krishnan, r. ( ). role of ultrasound in dengue fever. the british journal of radiology. [ ] adekoya, a. f., akinwale, a. t., & oke, o. e. ( ). a medical expert system for managing tropical diseases. college of natural sciences proceedings, - . [ ] faisal, t., taib, m. n., & ibrahim, f. ( ). adaptive neuro-fuzzy inference system for diagnosis risk in dengue patients. expert systems with applications, ( ), - . [ ] jang, j. s. ( ). anfis: adaptive-network-based fuzzy inference system. ieee transactions on systems, man, and cybernetics, ( ), - . [ ] faisal, t., ibrahim, f., & taib, m. n. ( ). a noninvasive intelligent approach for predicting the risk in dengue patients. expert systems with applications, ( ), - . [ ] chuang, l. y. ( ). an application of expert system for diagnosing fever caused by viral infection. journal of life sciences and technologies vol, ( ). [ ] shortliffe, e. (ed.). ( ). computer-based medical consultations: mycin (vol. ). elsevier. [ ] ashwinkumar, u. m., & anandakumar, k. r. ( , january). a web- based patient support system using artificial intelligence to improve health monitoring and quality of life. in second international conference on advanced computing & communication technologies (pp. - ). ieee. [ ] ibrahim, f., mohamad, m. i., makhtar, s. n., & ibrahim, j. ( ). classification of risk in dengue fever and dengue haemorrhagic fever using rule based expert system. in rd kuala lumpur international conference on biomedical engineering (pp. - ). springer berlin heidelberg. [ ] ho, s. h., speldewinde, p., & cook, a. ( ). predicting arboviral disease emergence using bayesian networks: a case study of dengue virus in western australia. epidemiology and infection, ( ), . [ ] adehor, a. b., & burrell, p. r. ( ). the integrated management of health care strategies and differential diagnosis by expert system technology: a single-dimensional approach. world academy of science, engineering and technology, , - . (ijacsa) - international journal of advanced computer science and applications vol. , no. , | p a g e www.ijacsa.thesai.org [ ] mcleod, r., & schell, g. p. ( ). management information systems. upper saddle river, nj: prentice hall [ ] iqbal, m. w., ahmad, n., & shahzad, s. k. ( ). usability evaluation of adaptive features in smartphones. procedia computer science, , - . [ ] ahmad, n. ( ). people centered hmi's for deaf and functionally illiterate users (doctoral dissertation, politecnico di torino-universität potsdam, germany). [ ] ahmad, n., shoaib, u., & prinetto, p. ( ). usability of online assistance from semiliterate users’ perspective. international journal of human-computer interaction, ( ), - . [ ] shoaib, m., hussain, i., mirza, h. t., & tayyab, m. ( , july). the role of information and innovative technology for rehabilitation of children with autism: a systematic literature review. in computational science and its applications (iccsa), th international conference on ieee (pp. - ). open data for anomaly detection in maritime surveillance i master’s thesis computer science september school of computing blekinge institute of technology se – karlskrona sweden open data for anomaly detection in maritime surveillance shahrooz abghari samira kazemi this thesis is submitted to the school of computing at blekinge institute of technology in partial fulfillment of the requirements for the degree of master of science in computer science. the thesis is equivalent to weeks of full time studies. contact information: authors: shahrooz abghari e-mail: shahroozabghari@gmail.com samira kazemi e-mail: kazemi.samira@gmail.com university advisors: dr. henric johnson, school of computing blekinge institute of technology dr. niklas lavesson, school of computing blekinge institute of technology school of computing blekinge institute of technology se – karlskrona sweden internet : www.bth.se/com phone : + fax : + mailto:shahroozabghari@gmail.com� mailto:kazemi.samira@gmail.com� i abstract context maritime surveillance (ms) has received increased attention from a civilian perspective in recent years. anomaly detection (ad) is one of the many techniques available for improving the safety and security in the ms domain. maritime authorities utilize various confidential data sources for monitoring the maritime activities; however, a paradigm shift on the internet has created new sources of data for ms. these newly identified data sources, which provide publicly accessible data, are the open data sources. taking advantage of the open data sources in addition to the traditional sources of data in the ad process will increase the accuracy of the ms systems. objectives the goal is to investigate the potential open data as a complementary resource for ad in the ms domain. to achieve this goal, the first step is to identify the applicable open data sources for ad. then, a framework for ad based on the integration of open and closed data sources is proposed. finally, according to the proposed framework, an ad system with the ability of using open data sources is developed and the accuracy of the system and the validity of its results are evaluated. methods in order to measure the system accuracy, an experiment is performed by means of a two stage random sampling on the vessel traffic data and the number of true/false positive and negative alarms in the system is verified. to evaluate the validity of the system results, the system is used for a period of time by the subject matter experts from the swedish coastguard. the experts check the detected anomalies against the available data at the coastguard in order to obtain the number of true and false alarms. results the experimental outcomes indicate that the accuracy of the system is %. in addition, the coastguard validation results show that among the evaluated anomalies, . % are true alarms, . % are false and . % belong to the vessels that remain unchecked due to the lack of corresponding data in the coastguard data sources. conclusions this thesis concludes that using open data as a complementary resource for detecting anomalous behavior in the ms domain is not only feasible but also will improve the efficiency of the surveillance systems by increasing the accuracy and covering some unseen aspects of maritime activities. keywords: open data, anomaly detection, maritime security, maritime domain awareness ii acknowledgment first and foremost, we would like to thank our supervisors dr. niklas lavesson and dr. henric johnson for their patient guidance, encouragement and advice throughout this thesis. we would also like to acknowledge the invaluable support of the swedish coastguard in this thesis. in particular, we are grateful to peter ryman, the law enforcement officer at the coastguard, without whose knowledge and kind support this study would not have been successful. special thanks are due to martin boldt for all his help and support during the tough time of system installation. last but not least, we would like to express our heartfelt thanks to our beloved families, for their understanding and endless love through the duration of our studies. iii table of contents . introduction ....................................................................................................................... . problem statement ........................................................................................................... . research questions .......................................................................................................... . aims and objectives .......................................................................................................... . contribution ...................................................................................................................... . outline ................................................................................................................................ . background .......................................................................................................................... . terminology ....................................................................................................................... . jdl data fusion model .................................................................................................... . related work .................................................................................................................... . research methodology ............................................................................................... . open data in the ms domain ........................................................................................... . identification of maritime anomalies .......................................................................... . framework design .......................................................................................................... . implementation ................................................................................................................ . . data description ............................................................................................................ . . data collector module ................................................................................................... . . database ......................................................................................................................... . . anomaly detector module .............................................................................................. . . display client ................................................................................................................. . experimental design ....................................................................................................... . validation design ........................................................................................................... . validity threats ............................................................................................................. . verification ....................................................................................................................... . experimental results .................................................................................................. . validation results ........................................................................................................ . discussion ............................................................................................................................. . conclusion and future work ................................................................................. appendix a: open and closed data sources ............................................................. appendix b: ports regions ................................................................................................... appendix c: algorithms ....................................................................................................... appendix d: user interface ................................................................................................ appendix e: acronyms ........................................................................................................... references ................................................................................................................................... . introduction ms is the effective understanding of all maritime activities that could impact the security, safety, economy or environment in addition to having a complete rmp, the way that the ms systems are used by human operators plays an important role in the efficiency of the surveillance operations. monitoring vast sea areas and trying to establish maritime domain awareness (mda) for human operators is a difficult and time-consuming task (riveiro, falkman, & ziemke, a). this is due in part to the large amounts of heterogeneous data from multiple sources but also to the difficulties in detecting anomalous behavior from normal maritime activities. therefore, having an automatic detector of unusual activities would help decision makers to efficiently monitor the ongoing activities in the surveillance area. . in recent years, the ms domain has received increased attention because of terrorism, smuggling activities and illegal immigration. an efficient ms system requires a complete recognized maritime picture (rmp), which can be defined as a composite picture of maritime activities over an area of interest (lefebvre & helleur, ). for national maritime sovereignty, the rmp should include all activities within the nautical miles exclusive economic zone (eez). however, for some purposes such as the detection of illegal vessel transits, the rmp could extend beyond this region (ponsford, d’souza, & kirubarajan, ). using today’s technology, continuous tracking of all maritime activities by a single sensor data is not sufficient since it cannot monitor everything that happens in the surveillance area. on the other hand, there are large amounts of data in the ms domain that are gathered from a variety of sensors, databases and information systems. therefore, by taking advantage of all the available data sources it would be possible to obtain a complete rmp. according to the department of homeland security as well as the ad techniques, the use of different data sources will highly influence the detection of suspicious activities. usually, only the data received from sensors are used for ad but there are a number of additional data sources regarding maritime activities that can be useful for this purpose. these sources of data consist of open and closed data about vessels, cargos, crew, etc. the closed data are only accessible to the maritime authorities, such as the coastguard, but, by contrast, the open data are available online and freely accessible and reusable to the public. for instance, there are different organizations such as ports that publish their vessel traffic data or their facilities information online. in addition to the organizations, there are different online communities such as blogs, forums and social , ad is one of the enabling techniques for mda. however, there are various ad techniques available and it is essential to choose the appropriate techniques that accomplish the ms goals. data-driven ad approaches find the anomalous behavior by constructing a model from normal data and calculating the deviation from that model. however, relying only on data-driven approaches for surveillance systems is not sufficient due to the lack of user involvement in the detection process (riveiro & falkman, ). furthermore, because of the diverse and complicated nature of the activities in the surveillance area, some of the suspicious behaviors are not directly observable. therefore, finding all types of anomalies by using data-driven approaches seems to be impossible. on the other hand, maritime domain experts have the required knowledge and experience for finding maritime anomalies. including the expert’s knowledge about suspicious activities in the detection process can result in improved ad. integrating maritime surveillance, common information sharing environment (cise). retrieved from http://ec.europa.eu/maritimeaffairs/policy/integrated_maritime_surveillance/documents/integrating_maritime_su rveillance_en.pdf national plan to achieve maritime domain awareness for the national strategy for maritime security. retrieved from www.dhs.gov/xlibrary/assets/hspd_mdaplan.pdf http://ec.europa.eu/maritimeaffairs/policy/integrated_maritime_surveillance/documents/integrating_maritime_surveillance_en.pdf� http://ec.europa.eu/maritimeaffairs/policy/integrated_maritime_surveillance/documents/integrating_maritime_surveillance_en.pdf� http://www.dhs.gov/xlibrary/assets/hspd_mdaplan.pdf� networks which provide the possibility of sharing information about maritime events. some of the advantages of open data in addition to the availability for the public and free accessibility are: first, open data can reveal some facts that are not reported to the maritime authorities or available in their databases and second, open data can be used in the global context and are not suffered from legitimate limitations of exchanging data between different countries. by applying the open data to the detection process, the ad can be done more wisely and the results can have more facts of interests for the maritime experts. . problem statement in the maritime domain, there are different kinds of data sources that provide heterogeneous data regarding maritime activities. the majority of data, which are used by the ms systems, belong to the surveillance area of each country and are obtained from a variety of sensors and databases that are only accessible by the countries authorities. for detecting some of the anomalous activities such as smuggling, the maritime data beyond the surveillance area of each country are required. in order to assure security, maritime organizations in different countries need to exchange their privileged data and for this purpose they should deal with the diverse regulation of the data protection in each land. exchanging data among countries is difficult, time-consuming and in some cases impossible because of the legislative issues. moreover, there are activities that are neither reported to the maritime organizations, nor recorded in their data sources but they can be useful for the surveillance purpose. the publicly accessible and reusable data that are free from the legislative issues and revealing the unseen aspects of maritime activities are referred as open data. consequently, employing the open data along with other confidential data sources would be beneficial for the ms systems to achieve their goals. . research questions given the context of available data sources and ms systems at the swedish coastguard and the types of anomalies that the subject matter experts at the coastguard are interested in, the first research question is: . how accurate and valid are the results of an ad system that exploits open data as a complement to the available closed data? in addition, by considering the accuracy as the degree to which the aforementioned ad system is able to distinguish between the normal and anomalous activities, and the validity as the degree to which the system results are true in real life, the next question is: . what is the performance difference between the system accuracy and the validity of results? . aims and objectives the aim is to investigate the potential open data as a complementary resource for ad in the ms domain. objectives: • identify existing open data sources in the maritime domain. • identify those open data sources that are suitable for being used for ad in the ms domain. • propose a framework for ad in the ms domain based on using the open data. • develop an ad system based on the proposed framework and evaluate the accuracy of the system. • validate the implemented ad system in real life. . contribution this thesis contributes with a deeper understanding of open data as a complementary resource for effectively establishing the ms operations. it provides a framework for using the open data sources together with other sources of data for ad in the ms domain. according to the framework, an ad system is developed which employs a number of algorithms to implement the expert rules for detecting anomalies. the final contribution is the evaluation of the implemented ad system via the coastguard validation in real life and also an experiment. . outline the remainder of this work is organized as follows: section reviews the background and related work regarding the open data, ad and data fusion in the ms domain. section presents the research methodology. validity threats and verification are described in sections and , respectively. section presents the experiment results and the validation results are shown in section . section features a detailed discussion about the obtained results. finally, section concludes the research with a discussion on the possible directions for future work. . background the idea behind open data has been established for a long time. open data can be used in a variety of domains and can be obtained from any resource. the two major sources of open data are the open data in science and the open data in government according to the estimation by dedijer and jéquier ( ), % of all information is open source, % is grey information (preprints of scientific articles, rumors in business circles, project proposals submitted to a research-funding agency, discussions with well- informed specialists, etc.), . % is secret and . % is non-existent information (i.e. the information you have, but you are not aware of it). considering the large ratio of the open data sources, there should be a great value in using them in different domains. in the ms systems, the majority of the exploited data are obtained from the confidential sources. however, in recent years the new concept of the web, which takes the network as a platform for information sharing, interoperability and collaboration, has created new sources of data for ms. there are organizations and communities that provide their maritime related data online and make them accessible for the public. therefore, it would be beneficial for the ms . the longstanding concept of open data in science tries to overcome the difficulties in the current system of scientific publishing such as the inability to access data or usage limitation that is applied by the publishers or data providers (molloy, ). different groups, individuals and organizations are gathered to participate in a movement toward reforming the process of scientific publication (molloy, ). one of the outcomes of the open data movement in science is the online availability of large number of scientific datasets for the public by different organizations. as well as the open data movement in science, governments for over a decade attempt to publish government data online and make it publicly accessible, readily available, understandable and usable (alonso et al., ). sharing the government data with the public provides openness and transparency with citizens. it can also improve the degree of participation in the society activities and the efficiency and effectiveness of the government services and the operations within and between the governments (dietrich et al., ). open definition. retrieved from opendefinition.org/okd/ http://opendefinition.org/okd/� systems if they can take advantage of the open data to increase the safety and security in their surveillance area. . terminology according to the department of homeland security data fusion involves the process of combining data from multiple sources or sensors and making inferences that may not be possible from a single source or sensor (hall & mcmullen, ). , mda would be achieved by monitoring the maritime activities, fusing and analyzing the data in a way that normal activities can be identified and anomalies differentiated. therefore, ad and data fusion techniques are important technologies for mda. ad is widely used in the areas such as video surveillance, network security and military surveillance. chandola, banerjee and kumar ( ) define ad as: the problem of finding patterns in data that do not conform to expected behavior. depending on the domain of study, the non-conforming patterns are called by different names such as anomalies, outliers, exceptions, etc. in the ms domain, these non-conforming patterns are referred as anomalies. defense r&d canada (roy, ) provides the following definition for the term anomaly in the context of the ms domain: something peculiar (odd, curious, weird, bizarre, atypical) because it is inconsistent with or deviating from what is usual, normal, or expected, or because it is not conforming to rules, laws or customs. the term open data refers to the idea of making data freely available to use, reuse or redistribute without any restriction. the open data movement follows the other open movements such as open access and open source. according to the open knowledge foundation . jdl data fusion model , a community based organization that promote open knowledge (whether it is content, data or information-based), an open work should be available as a whole, with a reasonable reproduction cost, preferably downloading via the internet without any charge and in a convenient and modifiable form. furthermore, it should be possible to modify and distribute the work without any discrimination against persons, groups, fields or endeavor. in the scope of this thesis, the open data term refers to the publicly available data that may or may not require free registration. one of the most widely used data fusion models in the literature is the jdl data fusion model. it was developed by the us joint directors of laboratories (jdl) data fusion sub- panel in . a current version of the model consists of six different levels. table shows the description of these six levels (hall & mcmullen, ). in this work the main focus is on ad which is done in the level (situation assessment) of the jdl model. the input of this level is the identified entities and their related information regarding each other or the environment and the output would be the assessment of a situation as normal or anomalous (brax, niklasson, & smedberg, ). national plan to achieve maritime domain awareness for the national strategy for maritime security. retrieved from www.dhs.gov/xlibrary/assets/hspd_mdaplan.pdf open definition. retrieved from opendefinition.org/okd/ http://www.dhs.gov/xlibrary/assets/hspd_mdaplan.pdf� http://opendefinition.org/okd/� table summary of the jdl data fusion model components jdl model component description level processing (source preprocessing) at this level, preprocessing of data from sensors and databases would be done by means of image processing, signal processing and conditioning, unit conversions, bias corrections or feature extractions, etc. level processing (object refinement) this level focuses on combining data from sensors and databases in order to obtain the most accurate and reliable estimates of an entity’s position, movement, attributes, characteristics and identity. level processing (situation assessment) according to the obtained result of previous level, a description of current relationships among entities and their relationship to the environment would be developed in order to determine the interpretation of the situation. level processing (impact assessment) this level focuses on the estimation and prediction of alternative futures and hypotheses concerning the current situation to determine the potential impacts or threats. level processing (process refinement) this level is a meta-process that monitors the whole data fusion process to optimize the utilization of data sources and algorithms and improve the performance of the ongoing data fusion. level processing (cognitive refinement) level focuses on transforming the result of data fusion in to the displays and understandable information for the user and improvement of human/computer effectiveness. . related work data fusion techniques have been used for a long time in the ms domain. the majority of studies focused on target tracking, tactical situation awareness and threat assessment (akselrod, tharmarasa, kirubarajan, zhen ding, & ponsford, ; bick & barock, ; danu, sinha, kirubarajan, farooq, & brookes, ; di lallo et al., ; gad, ; giompapa, farina, gini, graziano, & di stefano, ; giompapa et al., ; hatch, kaina, mahler, & myre, ; henrich, kausch, & opitz, ; jouan, valin, gagnon, & bosse, ; lefebvre & helleur, ; maresca et al., ; vespe, sciotti, & battistello, ). typically, in these works the combination of two or more sensors such as: automatic identification system (ais), infrared (ir), video camera, synthetic aperture radar (sar), vessel traffic service (vts) radar, over the horizon (oth) radar, high frequency surface wave radar (hfswr) and microwave radar is used and the surveillance area is limited to the coastal regions. in recent years, the number of studies that address the use of ad in the ms domain is increasingly growing. ad techniques are divided into two groups, namely data-driven and knowledge-driven approaches. there are a couple of works that proposed knowledge-based systems with different representation techniques and reasoning paradigms such as rule- based, description logic and case-based reasoning (guyard, roy, & defence r&d canada- valcartier, ; nilsson, van laere, ziemke, & edlund, ; roy & davenport, ). a prototype for a rule-based expert system based on the maritime domain ontologies was developed (edlund, gronkvist, lingvall, & sviestins, ) that could detect some of the anomalies regarding the spatial and kinematic relation between objects such as simple scenarios for hijacking, piloting and smuggling. another rule-based prototype was developed by defense r&d canada (roy, , ). the aforementioned prototype employed various maritime situational facts about both the kinematic and static data in the domain to make a rule-based automated reasoning engine for finding anomalies. one of the popular data-driven ad approaches is the bayesian network (fooladvandi, brax, gustavsson, & fredin, ; johansson & falkman, ; lane, nevell, hayward, & beaney, ). johansson and falkman ( ) used the kinematic data for creating the network; however, in the work that was done by fooladvandi et al. ( ) expert’s knowledge as well as the kinematic data was utilized in the detection process. moreover, lane et al. ( ) presented five unusual vessel behaviors and the way of formulating them in an ad system that the estimation of the overall threat was performed by using a bayesian network. unsupervised learning techniques have been widely used for data-driven ad such as trajectory clustering (dahlbom & niklasson, ), self organizing map (riveiro, johansson, falkman, & ziemke, ) and fuzzy artmap neural network (rhodes, bomberger, seibert, & waxmamn, ). some statistical approaches, such as gaussian mixture model (laxhammar, ), hidden markov model (andersson & johansson, ), adaptive kernel density estimator (ristic, la scala, morelande, & gordon, ) and precise/imprecise state- based anomaly detection (dahlbom & niklasson, ) have been used in this context. the majority of the works that have been done in the context of ad only used the ais data. there are a number of studies that employed data fusion techniques to fuse data from different sensors in ad systems (carthel et al., ; guerriero, willett, coraluppi, & carthel, ; rhodes, bomberger, seibert, & waxman, ; vespe, sciotti, burro, battistello, & sorge, ). in these studies, the surveillance area was restricted to the coastal regions and the combination of data from ais, sar, ir, video and radar was used in the fusion process to obtain the vessel tracks. furthermore, there are some other works that focused on the fusion of both sensor and non-sensor data (andler et al., ; ding, kannappan, benameur, kirubarajan, & farooq, ; fooladvandi et al., ; lefebvre & helleur, ; mano, georgé, & gleizes, ; riveiro & falkman, ). for example, lefebvre and helleur ( ) and riveiro and falkman ( ) treated the expert’s knowledge as the non-sensor data. riveiro and falkman ( ) introduced a normal model of vessel behavior based on ais data by using self organizing map and a gaussian mixture model. according to the model, the expert’s knowledge about the common characteristic of the maritime traffic was captured as if-then rules and the ad procedure was supposed to find the deviation from the expected value in the data. lefebvre and helleur ( ) fused radar data with user’s knowledge about the vessels of interests. the sensor data were modeled as track and the non-sensor data were modeled as templates. the track-template association was done by defining mathematical models for tracks and using fuzzy membership functions for association possibilities. mano et al. ( ) proposed a prototype for the ms system that could collect data from different types of sensors and databases and regroup them for each vessel. sensors like ais, hfswr and classical radars and databases such as environmental database, lloyd’s insurance and tf vessel db were included in this prototype. by using multi-agent technology an agent was assigned to each vessel and anomalies could be detected by employing a rule-based inference engine. when the combination of anomalies exceeded a threshold, vessel status was informed to the user as an anomaly. the work presented by ding, kannappan, benameur, kirubarajan and farooq ( ), proposed the architecture of a centralized integrated maritime surveillance system for the canadian coasts. sensors and databases included in this architecture were: hfswr, ads (automatic dependant surveillance) reports, visual reports, information sources, microwave radar and radar sat. a common data structure was defined for storing data that were collected from different sensors. andler et al. ( ) also described a conceptual ms system that integrated all available information such as databases and sensor systems (ais, lrit, intelligence reports, registers/databases of vessels, harbors, and crews) to help user to detect and visualize anomalies in the vessel traffic data in a worldwide scale. furthermore, the authors suggested using open data in addition to other resources in the fusion process. in conclusion, the main focus of the projects that have been done in the context of ad in the ms domain was related to using sensors data and mainly the ais data to find the anomalies in the coastal region. detection of some suspicious activities such as smuggling requires vessel traffic data beyond the coastal region. maritime authorities in each country have overall information of maritime activities in their surveillance area. but exchanging information among different countries is a complicated procedure because of the diverse regulation of data protection in each land. therefore, using data sources that are free from legislative procedures can be a good solution for providing information that belongs to the regions outside the land territory. furthermore, all the information about maritime activities is not recorded in the authorities’ databases or reported to them. on the other hand, there are numerous open data sources consists of different websites, blogs and social networks that can be useful for observing the hidden aspects of maritime activities. hence, this thesis will investigate the potential open data sources for maritime activities and exploit them to build an ad system. . research methodology this thesis starts by investigating the applicable open data sources for ad in the ms domain. then, potential maritime anomalies that can be detected by the obtained open data sources and the way that the open data can be applied to the ad process are identified. the next step is to design and implement an anomaly detector system that can use open data in the detection process. for this purpose, a general framework for ad based on using both open and closed data sources in the ms domain is proposed. then, an anomaly detector system is developed according to the proposed framework. in order to understand to what extent the system results are valid, the system is evaluated in real life by the subject matter experts from the swedish coastguard. however, the system accuracy should be evaluated before using the system in real life. therefore, an experiment is conducted to measure the system accuracy. the following sections describe these steps in detail. . open data in the ms domain applicable data sources for ad in the ms domain can be divided into three categories. the first and main category consists of sensors. sensors provide kinematic data for each object in their coverage area and can be categorized as passive and active. active sensors do not require cooperation from objects and collect the data by active probing the environment such as radar and sonar (a. m. ponsford et al., ). however, passive sensors rely on the data that are broadcasted by objects intentionally such as ais or unintentionally such as electronic intelligence (elint) and signal intelligence (sigint) systems (a. m. ponsford et al., ). more information about the main maritime sensors can be found in (İnce, topuz, & panayirci, ; vespe, sciotti, & battistello, ). the second category of data sources includes the authorized databases which contain information about vessels, cargos, crew, etc integrated maritime policy for the eu : working document iii on maritime surveillance systems. retrieved from . the first and second categories are only accessible to the maritime authorities such as the coastguard and can be referred as closed data sources. the third category belongs to the open data sources which are publicly available via the internet and are free to access or reuse. these data sources consists of vessels traffic data and reports or news that are related to the maritime domain and can be found in different blogs, websites or social networks. epub.sub.uni-hamburg.de/epub/volltexte/ / /pdf/maritime_surveillance_en.pdf http://epub.sub.uni-hamburg.de/epub/volltexte/ / /pdf/maritime_surveillance_en.pdf� to obtain the applicable open data for ad, the first step is started by looking through the information resources document . identification of maritime anomalies provided by the international maritime organization (imo). imo is the united nations specialized agency with responsibility for the safety and security of shipping and the prevention of marine pollution by vessels. the document introduces governmental and intergovernmental organizations that work in different fields related to the ms domain such as maritime safety, prevention of pollution from vessels, liability and insurance issues, shipping information, etc. all these organizations’ websites and the links provided by each of them are investigated and a list of online data sources is prepared. the prepared list contains data sources that have information that can be qualified for the ms purpose such as information related to ais, vessel characteristics, ports, maritime companies, suppliers’ information, weather, etc. table a. of appendix a, presents the obtained data sources. these sources of data are available online but having access to some of them needs non-free registration. moreover, in the process of finding open data sources, it is attempted to obtain sources of data that are related to the baltic region and mostly sweden by use of the previously observed data sources and also google search engine. an important aspect of the literature review is to find out more about the potential maritime anomalies. the main sources of information about maritime anomalies are reports of the two workshops that were held in canada (roy, ) and sweden (andler et al., ; van laere & nilsson, ). in these two workshops attendees were experts in the maritime domain and a variety of maritime anomalies were identified. the outcome of the swedish workshop was the identification of the most desired anomalous behaviors for the swedish stakeholders. the identified anomalies belong to the following categories: tampering, owner/crew, history, rendezvous (object, location), movement and cargo. the outcome of the canadian workshop was a taxonomy of maritime anomalies that categorized anomalies as static and dynamic. static anomalies were related to the vessel characteristics such as name, imo number, etc. dynamic anomalies were divided into two groups, kinematic and non- kinematic anomalies. the kinematic anomalies included anomalies related to location, course, speed, reporting and maneuver of vessels. the non-kinematic anomalies were related to passengers, crew list, cargo list, last port of call and next port of call. according to the identified anomalies by the two workshops, a list of some potential maritime anomalies that can be detected by use of the available open data sources is prepared. then, in a meeting with representatives of the swedish coastguard the types of anomalies that are interesting for them and the possibility of using open data for ad are discussed. during the meeting the prepared list of anomalies is presented and they are asked about the possibility of occurrence and their degree of interest for each anomaly. as an outcome of the meeting a number of scenarios are created and based on them rules are defined. the first scenario refers to the anomalies related to the vessel static information such as name, owner, imo number, dimensions, type and the status (in service or laid up). for example, sailing a vessel with a draught of meters over an area with a meters depth or observing a vessel that should be laid up or changing the name or the owner of a vessel during its voyage indicate the existence of suspicious activities. the second scenario is related to the prior arrival notification for vessels. vessels should inform their arrival time to the ports at least hours in advance. each port also provides an online timetable for the incoming vessels. therefore, any mismatch between the reported ais data regarding the destination or the arrival time of a vessel and the destination port timetable needs to be checked by coastguards. the third scenario is related to ordering pilots. usually, large vessels because of their size and weight need to be guided by pilots through dangerous and congested waters. therefore, vessels need to submit their request for a pilot and also inform information resources on maritime security and isps code. retrieved from www.imo.org/knowledgecentre/informationresourcesoncurrenttopics/maritimesecurityandispscode/documents/inf ormation% resources% on% maritime% security% and% isps% code.pdf http://www.imo.org/knowledgecentre/informationresourcesoncurrenttopics/maritimesecurityandispscode/documents/information% resources% on% maritime% security% and% isps% code.pdf� http://www.imo.org/knowledgecentre/informationresourcesoncurrenttopics/maritimesecurityandispscode/documents/information% resources% on% maritime% security% and% isps% code.pdf� the destination port. however, in some cases vessels order a pilot without informing the port. such situations should be investigated. in the next meeting, the scenarios and the rules are presented to the representatives of the swedish coastguard and they are asked to comment or suggest new scenarios or rules. by getting the final approval from the coastguard experts, one new rule (rule number ) is added to the list. table shows the admitted rules by the experts. these identified maritime anomalies can be detected by use of ais data, vessel traffic timetables in ports and pilots websites and the vessel characteristic data that are available in data sources such as lloyd’s. a name is given to each anomaly that can be detected by the rules, and for the rest of this thesis anomalies will be referred by their names. table the identified anomalies that can be detected by open data (confirmed by the swedish coastguard) no. expert rules anomaly if a vessel destination does not exist in the port schedule then anomaly. vessel_not_informed_port (a ) if a vessel eta does not match with the port eta for the vessel then anomaly. arrival_time_mismatched (a ) if a vessel entered a port without informing the port then anomaly. vessel_entered_port_without_ notice (a ) if a vessel has requested a pilot but has not used the service then anomaly. vessel_not_used_pilot (a ) if vessel a which normally travels between ports x and y, suddenly goes to port z then anomaly. unusual_trip_pattern (a ) if a vessel has not left a port according to the port schedule then anomaly. vessel_not_left_port (a ) if a vessel exists in a port schedule but it has not entered the port then anomaly. vessel_not_entered_port (a ) if a vessel does not exist in the port schedule and the vessel has requested a pilot then anomaly. vessel_ordered_pilot_and_ not_informed_port (a ) if a vessel has moored in a port and has been observed somewhere else then anomaly. vessel_moored_in_port (a ) if vessel a has not entered a port according to the port schedule instead vessel b enters the port at the same time slot then anomaly. wrong_vessel_entered (a ) if a vessel with the laid up status has been observed somewhere else then anomaly. vessel_laid_up (a ) note. eta = estimated time of arrival . framework design the open data anomaly detection system (odads) is designed for traffic monitoring and detecting anomalies in the ms domain by using open and closed data sources. figure depicts the odads architecture. odads consists of three modules: ) data collector, ) anomaly detector and ) display client. the data collector module is responsible for collecting open data from the internet, preprocessing and storing the data in a database. the data can be related to vessel traffic (such as ais reports, ports and pilots timetables), vessel characteristics, ports equipments and facilities, companies that are involved in maritime activities, news or reports about maritime events and activities available in different social media platforms (such as blogs and social networks), etc. the data store comprises a set of databases that contain data belong to different types of sensors, authorized databases and open data sources. the data in the data store can be fused or integrated before being used in the detection process. when the data collector completes its task, anomaly detector starts to work. the anomaly detector module analyzes the available data (open and closed data) and detects possible anomalies by utilizing both knowledge-driven and data-driven techniques. different ad techniques are employed due to the distinct nature of anomalies and the complexity of the environment in the ms domain. previously known anomalies can be detected by knowledge based techniques such as rule-based, but in real life an ad system must be able to detect the unseen anomalies, too. this is one of the benefits of using data- driven methods such as machine learning techniques. therefore, detecting different types of anomalies seems to be possible by exploiting different techniques. the display client module is the user interface of the system. this module represents the cognitive refinement level (level ) of the jdl model. m. j. hall, hall and tate ( ), argued that the effectiveness of a system can be affected by the way that the system produced information is comprehended by the human user. the cognitive refinement process involves traditional human-computer interaction (hci) utilities such as geographical display or advanced methods that support functionalities such as cognitive aids, negative reasoning enhancement, focus/defocus of attention and representing uncertainty. section . describes the implementation details for each module. figure . the open data anomaly detection system (odads) architecture. the data collector module collects data from the internet and stores them in the database. the anomaly detector module detects anomalies by taking advantage of different techniques. the display client module displays the detected anomalies to the user and enables system- user interaction. display client data collector anomaly detector knowledge-driven approaches data-driven approaches internet .... weather information vessel traffic vessels, ports and companies information social networks data store open data closed data . implementation odads is implemented by taking advantage of the identified maritime anomalies and the obtained open data sources that were discussed in the previous sections. to limit the scope, first of all only four types of vessels: passenger, ferry, cargo and tanker are considered and other types of vessels such as fishing and sailing vessels, which ports usually do not provide any information about them, are omitted. secondly, the rule related to the vessel static information is ignored. furthermore, the wrong_vessel_entered anomaly is excluded due to its complexity. as well as the a -a anomalies, in further collaboration with the coastguard representatives during the implementation phase, a new type of anomaly is proposed. this anomaly is called under_surveillance_vessel and occurs when a vessel of interest has any of the a -a anomalies and the vessel exists in the vessels blacklist. . . data description the required vessel traffic data can be obtained from ais reports and ports and pilots timetables. the surveillance area is restricted to the north of the baltic sea and a part of the gulf of finland, the regional area between three european countries sweden, finland and estonia. figure shows the surveillance area where the geographic coordinates lie between latitudes . º - . º n and longitudes . º – . º e. this region is one of the high- traffic regions in the baltic sea and is surrounded by the four highly used ports. more information regarding each port can be found in appendix b. figure . the area of interest is restricted to the north of the baltic sea and a part of the gulf of finland. ports from left to right are norrköping, stockholm group (nynäshamn, stockholm and kapellskär), helsinki and tallinn (the image is adapted from google earth). . . . ais data due to inaccessibility to the raw ais data in this thesis, the ais reports that are provided by the marinetraffic.com website are exploited. these reports consist of both static and dynamic types of data for each vessel during its voyage. vessel static data include name, type, built year, size, draught, flag, call sign, maritime mobile service identity (mmsi), imo identification number, origin (last known port), destination, estimated time of arrival (eta) and dynamic data are related to speed (max and average), position (longitude and latitude), course over ground (cog), heading. www.marinetraffic .aegean.gr/ais/getkml.aspx http://www.marinetraffic .aegean.gr/ais/getkml.aspx� data collector ais logger web scraper database ais reports table port and pilot schedule table . . . ports and pilots data ports timetables are obtained from the websites of the high-traffic ports in the area of interest: stockholm group (stockholm, kapellskär and nynäshamn) and norrkoping ports in sweden, helsinki port in finland and tallinn port in estonia. the pilots timetables belong to the vessel traffic in stockholm pilotage area in sweden. table a. of appendix a, presents each data source in detail. . . data collector module the data collector module is responsible for extracting data from different open data sources via the internet, data preprocessing and data storage. this module consists of two parts. the first part is the ais logger, a shell script program, which downloads and extracts the processed vessel ais reports and tracks figure . the data collector module consists of two parts, ais logger that provides a list of vessels in the area of interest and the web scraper, a program which gathers vessels details information and ports and pilots timetables from the internet. data from marinetraffic.com website. since the marinetraffic.com website updates the data at -minute intervals, data extraction is done every minutes and the collected data are stored in a file. when the file is prepared, the second part of the module, the web scraper, starts to work. it is a java program that harvests data from different websites and stores them into the database. these data include ports and pilots timetables from their websites and the vessels details information from the marinetraffic.com. figure illustrates the different parts of the data collector module. . . . preprocessing data preprocessing can be performed by a number of techniques such as data cleaning to remove noise, data integration to merge multiple data sources into an understandable data, data transformation such as normalization to improve the accuracy and data reduction to eliminate the redundant features (han et al., ). in the first step of preprocessing, text values such as vessel name, origin, destination and company name are transformed to the same format (lowercase characters) and all special characters and whitespaces are removed. since the data are collected from different sources in different countries, time values are converted to central european time (cet). every data source has its own format of data representation. for instance, vessel arrival time in ports can be stored as two separated parts of date and time, a combination of date and time values or just as date. moreover, some of the ports provide additional vessel information such as weight and length as well as the timetables. this diversity makes some parts of the provided data remain unused. therefore, a common data representation format for ports and pilots data is defined that contains vessel name, vessel type, origin, destination, company name, vessel status and arrival/departure time. www.marinetraffic.com/ais/gettrackxml.aspx?mmsi=xxxxxxxx http://www.marinetraffic.com/ais/gettrackxml.aspx?mmsi=xxxxxxxx� database ais reports table port and pilot schedule table history anomaly table anomaly detector data-driven approaches statistical techniques knowledge-driven approaches expert rules . . database the database contains four tables to store the extracted open data: ) ais reports table, ) ports and pilots schedule table, ) vessel trip history table and ) detected anomaly table. there is also a table for storing vessels that are kept under surveillance because of their previous involvement in the criminal activities. the under surveillance vessels data are provided by the coastguard officer via the display client module and stored in the blacklist table. the ais reports table contains the processed ais data such as vessel name, type, flag, origin, destination, etc. the ports and pilots schedule table is used for storing the vessel data including arrival and departure time gathered from ports and pilots data sources. in the vessel trip history table, data related to the frequency of vessels trips between different ports are stored and the detected anomaly table contains the history of all types of detected anomalies. . . anomaly detector module the anomaly detector module is the main part of the system which employs different techniques to detect anomalies. after investigating the nature of the anomalies and the potential techniques, it is determined that except for the unusual_trip_pattern anomaly, the other types can be detected by exploiting search techniques. detection of the unusual_trip_pattern anomaly requires data-driven approaches such as machine learning or statistical techniques and also a history of the vessel traffic data for training the system. in case of using machine learning techniques, the system will detect the anomaly by learning the normal vessel trip pattern and highlighting the unusual trip and by using statistical approaches the system will make decision based on the frequency or the possibility of a vessel trip between two different ports. the next section describes the detection algorithm for each anomaly in detail. this module is depicted in figure . figure . the anomaly detector module . . . algorithms to define the appropriate detection algorithms, the first step is to specify what types of data are required for detecting each anomaly (table a. of appendix a). detection of each individual anomaly (except the unusual_trip_pattern anomaly) can be done by performing a search in the specified data for finding the desired match. if the match is not found then the vessel would be marked as an anomaly. appendix c gives a detailed description about each algorithm. algorithm shows the main procedure of the anomaly detector module. algorithms - and - present the described search process in detail. for the unusual_trip_pattern anomaly, data related to six months (september , - march , ) of vessel traffic in the surveillance area are gathered. by monitoring the activities during this period, the system will be able to find the normal pattern of vessels trips in the area of interest. for detecting this anomaly a simple statistical approach is used. a look up table is created and for each vessel the number of times that the vessel travels between two different ports is stored. for each vessel, if the frequency of travelling between its origin and destination is less than a predefined threshold then the vessel will be reported as an anomaly. algorithms and describe the process in detail. after executing all the algorithms, the final type of anomaly for each vessel should be determined. there are some situations that multiple anomalies can occur in the same time for a specific vessel. these situations are presented in table . for the combination of anomalies that do not have any feature in common new types of anomalies are defined. table the combined anomalies anomaly condition anomaly type a ,a vessel did not inform its arrival to the destination port and its trip to the port is not common. unusual_trip_pattern_and_ vessel_not_informed_port a ,a vessel did not inform its arrival to the destination port and it ordered a pilot. vessel_ordered_pilot_and_ not_informed_port a ,a ,a vessel did not inform its arrival to the destination port, it ordered a pilot and its trip to the port is not common. unusual_trip_pattern_and_ vessel_ordered_pilot_and_ not_informed_port_ informed_port a ,a vessel has delay and it ordered a pilot but did not use it. vessel_arrival_time_ mismatched_and_vessel_not_ used_pilot a ,a vessel has delay and its trip to the port is not common. unusual_trip_pattern_and_ vessel_arrival_time_ mismatched a ,a ,a vessel has delay and it ordered a pilot but did not use it and its trip to the port is not common. unusual_trip_pattern_and_ vessel_not_used_pilot_and_ vessel_arrival_time_ mismatched a ,a vessel entered a port without prior notification and has not left the port according to the plan. vessel_entered_port_ without_notice_and_not_ left_port_on_time a ,a vessel ordered a pilot and informed to the port however it did not used the pilot and its trip to the port is not common. unusual_trip_pattern_and_ vessel_not_used_pilot a ,a vessel ordered a pilot but did not use it and it has not entered the port. vessel_not_entered_port_ and_not_used_pilot a ,a vessel ordered a pilot but did not inform the port and it did not use the pilot. vessel_ordered_pilot_and_ not_informed_port_and_ vessel_not_used_pilot (continued) anomaly condition anomaly type a ,a vessel trip to the destination port is not common and it has not entered the port. unusual_trip_pattern_and_ vessel_not_entered_port a ,a ,a vessel ordered a pilot but did not use it and has an unusual trip and it has not entered the port. unusual_trip_pattern_and_ vessel_not_entered_port_ and_vessel_not_used_pilot a ,a vessel ordered a pilot but did not inform the port and the trip to the destination port is not common. unusual_trip_pattern_and_ vessel_ordered_pilot_and_ not_informed_port a ,[a -a ] vessel uses the information that belongs to a moored vessel in a port; since the vessel provides unreal information, it is normal that other anomalies are set for this vessel. vessel_moored_in_port a ,a ,a vessel ordered a pilot but did not inform the port and the trip to the destination port is not common and the vessel also did not use the pilot. unusual_trip_pattern_and_ vessel_ordered_pilot_and_ not_informed_port_and_ vessel_not_used_pilot . . . string matching techniques using exact string matching techniques for comparing vessel information from different data sources is inapplicable due to the potential errors that might occur because of different notations or even human operator mistakes during data entry. therefore, a metric should be used for measuring the degree of similarity between two vessels from different sources. for this purpose, some string comparison methods are evaluated such as damerau-levenshtein (d-l) also known as edit distance (damerau, ; levenshtein, ), the n-gram technique (p. a. v. hall & dowling, ), jaro (jaro, , ) and jarowinkler (winkler, ). d-l or edit distance is equal to the minimum number of edits (substitution, delete, insert and transposition) required to change one string to the other. the use of n-gram for string matching is performed by first constructing the n-grams (contiguous sequences of n items) for each string and then comparing the n-grams of both strings in order to find the number of common n-grams. jaro and jarowinkler (a variant of jaro) measure the number and order of common characters in two strings and also the number of transposition that is needed to change one of the strings to the other. jaro metrics are more similar to the human decision making compared to the d-l distance (denk & hackl, ). to evaluate these metrics, the d-l distance is implemented and for the other metrics simmetrics the collected data related to one day of vessel traffic in the area of interest are taken with distinct vessels names, origins and destinations and stored in separate groups. then, the string matching methods are applied to each group by considering all pair wise combination of the elements in each group. to evaluate the results, the obtained similarities are sorted in ascending order and it is verified whether or not the values in each pair can be considered similar by a human operator. two values are considered equal if they are exactly the same or a bit different because of misspellings or shortening. finally, based on the results it is , an open source library that contains a set of similarity metrics, is used. regarding the n-gram technique, choosing the length of n-grams is an important factor since it can affect the processing time and accuracy of the method. according to salton and mcgill ( ) and zamora, pollock, and zamora ( ), trigram is one of the n-gram methods (when n= ) that can achieve the best results in retrieving similar words. www.aktors.org/technologies/simmetrics/index.html http://www.aktors.org/technologies/simmetrics/index.html� concluded that for the available data, jarowinkler has the best performance among all the other metrics and its results are closer to reality. jaro and d-l are the second and the third best respectively and trigram is the worst. according to the degrees of similarity that are obtained from the result of jarowinkler and the data characteristics, two similarity thresholds, one for the vessel name and the other for the origin/destination are defined. the vessel name similarity threshold is considered as . and value for the trip threshold is . (given p as the similarity value, ≤ p ≤ , p= shows the exact match and p= indicates that the two strings are dissimilar). these thresholds are referred as vessel_name_similarity_threshold and trip_ similarity_threshold in appendix c. . . display client the display client module is the graphical user interface of odads. it is a web-based application and supports the functionalities of the jdl level processing such as hci utilities, cognitive aids and focus/defocus attention. while designing the user interface, the six principles of the user interface design that are based on the usage-centered design approach are considered. according to constantine and lockwood ( ), these principles are: structure, simplicity, visibility, feedback, tolerance, and reuse. figure d. of appendix d presents a snapshot of the user interface. when a user logs in to the system, the main view of the odads user interface is displayed. it consists of a geographical display and four tabs, namely anomaly, dataset, reports and setting. the geographical display utilizes the google maps api features to enable the vessel traffic monitoring in the area of interest. vessels are represented on the map by different colors according to their types. ferries and passenger vessels are green, cargo vessels are cyan and tankers are yellow. in order to give a better perspective of the surveillance area, tug and pilot vessels are also displayed on the map. the tug and pilot vessels are presented in gray and dark blue, respectively. by clicking on each vessel, an information window appears which contains the vessel information such as name, mmsi number, imo number, flag, heading, origin, destination, eat and anomaly type. the ad process is performed every minutes and consequently, the map is updated and if any new anomaly is detected, the anomaly tab becomes red in order to inform the user about the incoming suspicious activities. the newly detected anomalous vessels are displayed in red and the previously detected anomalies are shown in light red. the anomaly tab presents the overall information about the number of detected anomalies, total number of vessels and the last data collection time. moreover, anomalous vessels are shown in a table which has the search, pagination and zoom to map utilities. the dataset tab shows the ports and pilots timetables within a time interval and in a common representation format. this common data representation format contains vessel information such as name, type, company name, origin, destination, time, status and the name of the data source. the reports tab consists of two sections. the first section provides reports regarding the total number of each anomaly and detailed information about anomalous vessels in a time interval in the html or excel formats. the second section presents graphical reports about daily statistics of the detected anomalies in a stacked chart and overall statistics of all detected anomalies in a pie chart. the setting tab is designed for getting any kind of information or system settings from the user. in the current implementation this tab is responsible for presenting the information about vessels that should be kept under surveillance and providing the features to add, remove, edit or disable/enable a vessel. any changes in this tab would be considered in the next period of the ad process. the technologies that are used for implementing this module are java, spring framework, jsp, javascript, jquery and google maps api. . experimental design before using odads in real life, it is important to figure out to what extent the results of the system are accurate. accuracy is the degree to which the estimates or measurements of a quantity correctly describe the exact value of that quantity. in other words, accuracy is the proportion of true results in the population. to evaluate the system accuracy, the number of true positives (tp), false positives (fp), true negatives (tn) and false negatives (fn) are needed. accuracy is calculated by the following formula (han, kamber, & pei, ): 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 the first step in designing the experiment is to identify the population. the population consists of the vessel traffic data in the surveillance area. since the population is too large and it is impossible to look into all members manually to count the number of tp, fp, tn and fn, a sample should be taken from the population. the second step is to identify the sampling procedure. the important factors in sampling are the sampling frame, method and the sample size. the sampling frame consists of all members of the population that have the chance to be included in a sample and must be representative of the population. in this thesis the sampling frame is the vessel traffic data related to ais, ports and pilots in the surveillance area, which are provided by odads. due to the high volume of traffic through the surveillance area, it is expected that the majority of anomalies can be observed in one week execution of the system. therefore, one week of vessel traffic data in april is used as the sample frame. for choosing a sampling method, it is important to decide whether or not to use probability sampling. in probability sampling, all members of the population have the chance of being selected in the sample. if the members have an equal probability of selection, the sample can be unbiased. in this thesis it is possible to perform a random sampling with equal probabilities. however, because of the odads periodic data collection (every minutes), the data regarding to a specific vessel can be selected multiple times. to omit the effects of repeated data in a sample, a multistage sampling is performed. in the first stage, a simple random sampling without replacement is done for selecting the time slots that odads attempts to collect and analyze the data. after selecting the time slots, the corresponding data for each time slot will be selected by a stratified sampling. three strata are defined according to the type of vessels: ferry and passenger, cargo and tanker vessels. selection of vessels is also limited to the vessels that are originated from or targeted to the four particular ports. the total number of time slots in the sample frame is . this means that on average, odads collects data times a day. in the first stage of sampling a random timeslot is selected for each day, which results in time slots for one week. then, by considering the described limitation in the selection process, the average number of entire data in a selected time slot is about records. among these records, records are selected by stratification. almost % of the vessels in each time slot are moored. since the majority of the anomalies are related to the vessels trips, a limitation on the number of moored vessels in the samples is defined. in this way, it can be possible to check more anomalies in the evaluation process. the second stage of sampling is repeated by taking into consideration that the number of moored vessel in the sample cannot exceed from the half of the sample size (in this case ). . validation design after completing the implementation of odads, a meeting is held with the coastguard representatives. during the meeting the system is presented and it is determined that odads will be used by the coastguard officers for a period of time to verify the validity of the detected anomalies and evaluate the usefulness of such system for the coastguard. therefore, odads is installed on an internet server and another meeting is arranged with the coastguard to define the validation process. as the outcome of the meeting, it is decided the validation will take place at the coastguard headquarters office in karlskrona, sweden for four weeks (april , to may , ), at any time during working hours ( : to : ). the officers are supposed to evaluate the detected anomalies by checking them against the available data in the systems and data sources that are used during the normal operational activities at the coastguard. they are asked to provide weekly report about their evaluation results in order to decrease the possible malfunctioning of the system and the validation process. an excel form is provided to the officers in order to receive the reports in a common format. while investigating the validity of an alarm for a specific vessel, the following information should be reported: detection time, vessel information (such as name, mmsi and imo), trip information, type of the alarm (true/false), systems or data sources that are used to investigate the detected anomaly. the coastguard officers are also asked for providing any information about vessels that are marked as anomaly according to the available systems at the coastguard, but odads is unable to recognize them. although having such information is useful for analyzing the validation results, it is impossible for the coastguard to provide this information. a detected anomaly for a vessel is true if it can be confirmed by the available data sources at the coastguard and consequently it is false if the authorized data sources provide any information that declines the detected anomaly. no further assessment is done regarding the classification of the detected anomalies to true and false alarms. . validity threats in this thesis there are some issues that can threat the validity of the results and they should be considered before developing the system and performing the evaluation and validation. construct validity refers to the extent to which the results of a study reflect the theory or the concept behind (shadish, cook, & campbell, ). the main issue that may threat the construct validity is the design and reliability of implementation. results can be affected by the potential faults that may happen in the implementation either because of programming faults or lack of tuning. in addition, the inaccurate nature of the open data that are used may have some effect on the results. the open data can have errors due to the human operator mistakes. they do not follow a similar format and can be unavailable for a while or are not updated immediately. to diminish the undesirable effect of the data, creating a common data representation format, filtering of the data and using approximate string matching techniques would be helpful. for decreasing the effect of programming faults, the validity of implemented application should be tested different times with both real and manipulated data that contain anomalies. there are also some parameters in the application that are needed to be selected correctly. for instance, in order to match the vessel information, string matching techniques are required and to determine that two strings are matched with each other, a similarity threshold should be used. choosing an inappropriate value for such parameters will lead to incorrect results. the other threat of validity, which can occur while performing the evaluation and validation, targets both the internal and external validities. internal validity ensures that the observed relationship between the treatment and outcome is due to a casual relationship and it is not because of an uncontrolled factor. external validity refers to the ability of generalizing the result of the study to other domains, times or places (shadish et al., ). the threat to internal and consequently the external validity may occur if the data that are used in the evaluation process are biased and not representative of the population. in such situation generalizing the results of the treatments to the whole system is unrealistic. to prevent this issue, the samples, which are taken from the population in the experiment, should be real representative of the population. as well as the experiment, while performing the validation by the subject matter experts the same procedure for selecting and checking the detected anomalies should be followed. . verification to ensure that odads works properly, the system is tested manually with both real and manipulated data. the tests are performed during the implementation and also after completing the system. at first, a number of vessels with different types of anomalies are inserted to the real collected data to check whether all types of anomalies can be detected by odads. then, the system is run for a period of time and the detected anomalies are checked manually against the available data to make sure about their correctness. during the test phase the anomaly detector module is updated and some of the detection conditions are narrowed down. the process is repeated until the system can detect all the anomalies correctly. before starting the validation process, the system is used by the subject matter experts from the coastguard and according to their comments some of the algorithms in the anomaly detector module are updated. for example, ferries usually provide their schedule monthly or once in a couple of months; therefore, their arrival is not always available in daily schedule of the ports and the vessel_not_informed_port and unusual_ trip_pattern anomalies are often set for them. for this reason, these two anomalies will not be checked for ferries. . experimental results the results of the system execution during the specified week are provided. to provide an overview of the detected anomalies by odads, table illustrates the total number of vessels in the surveillance area. table shows the number of detected types of anomaly for each day and finally table provides the total percentage of each anomaly type during that week. table total number of vessels in the surveillance area days avg total number of vessels . cargo, tanker, passenger and ferry vessels . cargo, tanker, passenger and ferry vessels that are originated from or targeted to the specified ports . table total and the average number of detected anomalies during one week of execution anomaly days avg arrival_time_mismatched . vessel_not_informed_port . (continued) anomaly days avg vessel_not_left_port . unusual_trip_pattern . vessel_not_used_pilot . unusual_trip_pattern_and_vessel_ arrival_time_mismatched . vessel_moored_in_port . unusual_trip_pattern_and_ vessel_not_informed_port . vessel_entered_port_without_notie . vessel_not_entered_port . under_surveillance_vessel . vessel_arrival_time_mismatched_ and_vessel_not_used_pilot . vessel_ordered_pilot_and_not_ informed_port . unusual_trip_pattern_and_vessel_ not_used_pilot . vessel_ordered_pilot_and_not_ informed_port_and_vessel_not_ used_pilot . unusual_trip_pattern_and_ vessel_ordered_pilot_and_not_ informed_port . unusual_trip_pattern_and_vessel_ not_used_pilot_and_vessel_arrival_ time_mismatched . vessel_entered_port_without_ notice_and_not_left_port_on_time . vessel_not_entered_port_and_not_ used_pilot . unusual_trip_pattern_and_vessel_ not_entered_port . unusual_trip_pattern_and_ vessel_not_entered_port_and_ vessel_not_used_pilot . unusual_trip_pattern_and_vessel_ ordered_pilot_and_not_informed_ port_and_vessel_not_used_pilot . table the percentage of each anomaly during one week of execution anomaly count (%) arrival_time_mismatched . vessel_not_informed_port . vessel_not_left_port . unusual_trip_pattern . vessel_not_used_pilot . unusual_trip_pattern_and_vessel_arrival_time_mismatched . vessel_moored_port . unusual_trip_pattern_and_vessel_not_informed_port . vessel_entered_port_without_notice . vessel_not_entered_port . under_surveillance_vessel . vessel_arrival_time_mismatched_and_vessel_not_used_pilot . vessel_ordered_pilot_and_not_informed_port . unusual_trip_pattern_and_vessel_not_used_pilot . vessel_ordered_pilot_and_not_informed_port_and_ vessel_not_used_pilot . unusual_trip_pattern_and_vessel_not_entered_port . unusual_trip_pattern_and_vessel_ordered_pilot_and_not_ informed_port . vessel_entered_port_without_notice_and_not_left_port_on_ time . vessel_not_entered_port_and_not_used_pilot . unusual_trip_pattern_and_vessel_not_used_pilot_and_ vessel_arrival_time_mismatched . unusual_trip_pattern_and_vessel_not_entered_ port_and_vessel_not_used_pilot . unusual_trip_pattern_and_vessel_ordered_pilot_and_not_ informed_port_and_vessel_not_used_pilot . after carrying out the sampling, all the samples are checked against the primary identified anomalies (a -a ). to compute the number of tp, fp, tn and fn, a confusion matrix is created based on the nine classes of anomalies and the normal class (table ). the accuracy of the system is calculated as follow: 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 = + + + + = . the existing fn for the arrival_time_mismatched anomaly is due to the wrong provided ais data by the vessel or possibly the marinetraffic.com website and also the limitation of the system for considering all conditions. in this case, the vessel arrival time belongs to a couple of days before the current date and for this reason it is ignored by the system. however, it is quite possible to handle such situations if additional sources of ais data are available. table confusion matrix for the nine classes of anomalies and the normal class predicted class a a a a a a a a a normal total actual class a - - - - - - - - - a - - - - - - - - a - - - - - - - - - - a - - - - - - - - - - a - - - - - - - - - a - - - - - - - - - a - - - - - - - - - - a - - - - - - - - - - a - - - - - - - - - - normal - - - - - - - - - total . validation results the sea monitoring system that is used by the coastguard officer is called sjÖbasis • safeseanet . sjÖbasis aggregates the maritime data from different systems and agencies with the aim of improving the efficiency of ms. in sjÖbasis, the required data that contain vessel position, speed, heading, arrival/departure time and trip, are obtained from the following sources: : a vessel traffic monitoring and information system, which is the centralized european platform for maritime data exchange and linking maritime authorities across europe. it enables the member countries to provide and receive information about vessels, vessels movements and hazardous cargoes. main sources of information include ais-based position reports and notification messages that are sent by designated authorities in the participating countries. this system is available via the swedish maritime administration. www.kustbevakningen.se/sv/granslos-samverkan/sjoovervakningsuppdraget/samverkan-sjoinformation/ www.emsa.europa.eu/operations/maritime-surveillance/safeseanet.html http://www.kustbevakningen.se/sv/granslos-samverkan/sjoovervakningsuppdraget/samverkan-sjoinformation/� http://www.emsa.europa.eu/operations/maritime-surveillance/safeseanet.html� • sjöc: the swedish naval forces surveillance center, which uses military sensors for correlating positions with ais. • ais: the international system for presenting the vessels position and identity. • helcom ais the officer checks the validity of anomalies according to the priority that each anomaly has for him. table provides the priority of the anomalies for an officer with an interest in law enforcement. : the national ais data that is consolidated in cooperation of countries around the baltic sea. table prioritization of the identified anomalies from a law enforcement officer’s point of view anomaly priority arrival_time_mismatched vessel_entered_port_without_notice vessel_not_entered_port under_surveillance_vessel vessel_not_left_port vessel_moored_in_port unusual_trip_pattern vessel_not_informed_port vessel_not_used_pilot vessel_ordered_pilot_and_not_informed_port note. shows the highest priority and shows the lowest. during the four-week validation period, odads is used at the coastguard for working days and in total of the detected anomalies are evaluated. table presents the validation results. among the evaluated anomalies, there are a number of anomalous vessels that remain unchecked due to a lack of corresponding data in the coastguard systems. a large number of detected anomalies are related to the arrival_time_mismatched anomaly that in many cases can be due to the inconsistent time formats (utc, cet and eet either with or without daylight saving) in different data sources and various settings in the ais transmitters. in these cases, the detected anomalies by odads are logically true, however, in actual fact they are not. for example, a vessel that is going from helsinki to stockholm is reporting its arrival time according to the finland local time (eet format) instead of using the utc format. therefore, this time difference causes that the arrival time reported by the vessel does not match with the expected arrival time of the vessel at the destination port, which leads to set the arrival_time_mismatched anomaly for that vessel. www.helcom.fi/bsap/actionplan/en_gb/segmentsummary/ http://www.helcom.fi/bsap/actionplan/en_gb/segmentsummary/� table validation result of the coastguard alarms anomaly true false not checked vessel_not_informed_port arrival_time_mismatched vessel_not_used_pilot - - unusual_trip_pattern - vessel_not_left_port - vessel_ordered_pilot_and_not_informed_port - - vessel_moored_in_port - - under_surveillance_vessel - - unusual_trip_pattern_and_vessel_not_ informed_port - unusual_trip_pattern_and_vessel_ arrival_ time_ mismatched - - unusual_trip_pattern_and_vessel_ordered_ pilot_and_not_informed_port - - vessel_ordered_pilot_and_not_informed_ port_and_vessel_not_used_pilot - - total count total count (%) . . . in addition to the comparative analysis in the validation process, modus operandi of using an anomaly detector that takes advantage of open data is investigated. the coastguard officer uses an analysis tool it is also possible to look into the relation between the types of vessels and the detected anomalies. such assessment can be used for strategic and risk analysis. for tanker vessels the to analyze the odads excel reports and draw conclusions regarding the modus operandi of the system in emergency situations that can have impact on the ms operations. one of the possible analyses of the system reports can be the investigation of vessels with multiple anomalies. here are some examples of the vessels with multiple anomalies. a cargo vessel has recurring anomalies related to the arrival/departure time, trip and notification to the port. from the types of anomalies that are detected for this vessel, it can be concluded that the vessel behavior points to a higher threat concerning customs and border. for a passenger vessel the following anomalies are often detected: vessel_not_informed_port, arrival_time_mismatched, vessel_not_ entered_port and vessel_not_left_port. these anomalies may happen because of inaccurate or wrong provided data by the vessel. the conclusion and modus operandi for this vessel is that the duty officer will contact the vessel to highlight the importance of submitting accurate information. in real life, occurring anomalies such as vessel_not_entered_port or vessel_not_left_port for a passenger vessel can be related to serious issues such as an accident and it will result in increased scrutiny for that vessel. ibm i analyst's notebook available at www.i group.com/us/products/analysis-product-line/ibm-i -analysts- notebook http://www.i group.com/us/products/analysis-product-line/ibm-i -analysts-notebook� http://www.i group.com/us/products/analysis-product-line/ibm-i -analysts-notebook� most popular anomaly is vessel_not_informed_port. this anomaly has a high priority for emergency preparedness for accidents involving tankers. tankers with unusual_trip_pattern anomaly are a potential risk to other vessels and can cause accidents. from a risk assessment point of view the combination of this anomaly with the vessel_not_entered_port or vessel_not_used_pilot anomalies can lead to high-risk situations. the most occurring anomaly for ferries and passenger vessels is arrival_time_mismatched. in several cases this anomaly is detected incorrectly because of the wrong reported arrival time; however, this anomaly has great importance for the authorities to plan their operations regarding ferries and passenger vessels arrivals effectively. the most serious anomaly for ferries and passenger vessels is vessel_not_ left_port and the authorities should suspect that some form of accident or difficulty is arising regarding to the departure of the vessels. for cargo vessels, the most recurring anomalies are vessel_not_informed_port and arrival_time_mismatched. however, some of the arrival_time_mismatched anomalies are false alarms because of the incorrect data. the vessel_not_informed_port anomaly is important for ports security and safety. the prior notification to the ports is obligatory for vessels, but the fine for breaking this rule is negligible which lets the vessels that are involved in illegal activities such as smuggling disobey this rule. the most serious anomalies for the cargo vessels are the vessel_not_entered_port and under_surveillance_ vessel anomalies. furthermore, looking into the most frequent anomalies for different ports will assists the maritime authorities to make their decisions more efficiently. according to the report analysis, the most frequent anomaly in stockholm and nynäshamn ports is arrival_ time_mismatched, in port of kapellskär is vessel_not_informed_port and in norrköping port is unusual_trip_pattern. one possible conclusion from the most popular anomalies for the ports is that the ports authorities should be informed of the divergence in the traffic flow and the operational management functions in order to plan and allocate resources efficiently. on the other hand, in some cases the anomalies are too common for a port because of the inaccurate provided data and they can be disregarded by the officer. the received feedback from the coastguard representatives during the validation process shows that odads is implemented in a way that it can assist the operator to have a better understanding of the ongoing maritime activities. they believe that the odads results are reliable and the quality of the open data that are used is good and can be used in real life. the functionality, usability and visualization of odads are satisfying. in addition to illustrating the vessel traffic data in a simple, clear and informative way, odads can provide clear statement about the anomalies and its statistical reports are beneficial when the authorities and freight companies conduct strategic analysis of maritime traffic and risk assessment. finally, the capability of automated detection of anomalies by utilizing open data and the form of data representation could prove to be a valuable asset to the coastguard. . discussion taking advantage of ad systems will assist authorities to tighten security in the ms domain. there are a number of studies which focused on developing ad systems by using knowledge-driven and data-driven approaches. for instance, defense r&d canada (roy, , ) developed a rule-based prototype for ad by exploiting maritime situational facts about both kinematic and static data of the domain. edlund et al ( ) developed another prototype for a rule-based expert system to detect the anomalies regarding spatial and kinematic relation between objects. riveiro and falkman ( ) proposed using a combination of data-driven and knowledge-driven approaches to detect anomalies by use of a normal model of vessel behavior based on ais data and experts rules. in the majority of studies that addressed ad, the exploited data for the ad process were obtained from closed data sources and there is a lack of investigation on using open data sources for ad in the ms domain. therefore, in this thesis odads is implemented by employing expert rules to investigate the potential open data as a complement to the closed data for ad in the ms domain. the accuracy of odads is evaluated via an experiment and the experimental results indicate an accuracy of % for the system. in addition, validity of the system results is evaluated in real life by the experts from the swedish coastguard. despite the inaccurate nature of open data and by considering the fact that only open data sources are used in the system, the high number of true alarms ( . %) in the validation process admits the validity of the system outcomes. furthermore, there are no corresponding data in the authorized databases for . % of the evaluated anomalies by the coastguard. this fact refers to a potential information gap in the closed data sources. however, the considerable number of false alarms ( . %) for a surveillance system is unsatisfying. the number of false alarms indicates the difference between the accuracy of the system and the validity of the results. even though the data that are used in this thesis are obtained from relatively trusted data sources such as ports, the false alarms occur mostly because of the inaccurate data. the open data that are exploited by odads suffer from the errors due to human operator mistakes, irregular data update, data update latency and incompatible data format. in odads, there are situations that a detected anomaly is disappeared in the next periods of the system execution because of new arrival of the correct data. frequent occurrence of false alarms distracts operator’s attention from the real anomalies in the surveillance area. to decrease the false alarms in odads, the main solution is to integrate the open data with the closed data which can cover the lack of information or inaccuracy in the open data. in addition, defining a probability for detected anomalies can decrease the number of false alarms. this would be possible by analyzing the history of vessels behavior as well as the current situation and defining a probability threshold to omit the anomalies that have a lower probability than the threshold. furthermore, having extra information regarding vessels such as crew and cargo information, can affect the probability of being a real anomaly for a specific vessel. for example, if a vessel has the arrival_time_mismatched anomaly and it has a crew member with a criminal record or a special cargo, then there is a possibility that the vessel is stopped somewhere to exchange something. therefore, in such situation, the probability of being a true anomaly is high. according to the validation results, the unusual_trip_pattern anomaly creates the majority of false alarms. this is due in part to the statistical approach that is used for detecting this anomaly and also the wrong origin and destination information that the vessels provide. the lookup table that is created for storing the frequency of the trips between different places is not updated periodically. while populating the table, the ports timetables are used which can be incomplete. an alternative detection approach can be using of machine learning techniques which attempt to detect the anomaly according to the pattern of movements for individual vessels instead of the reported trip data by the vessels or ports. in local areas such as the surveillance area in this thesis, mainly because of large amount of quality assured data and the limited size of the surveillance area, it is easier for the maritime authorities to track and control the vessels activities. therefore, use of open ais data in this region is not required and it should be prohibited to decrease the negative impacts of open data on the system results. on the other hand, when the vessel information beyond the eez is required, the value of open data becomes more obvious. . conclusion and future work this thesis investigated the potential open data as a complementary resource for anomaly detection (ad) in the maritime surveillance (ms) domain. a framework for ad was proposed based on the usage of open data sources along with other traditional sources of data. according to the proposed ad framework and the algorithms for implementing the expert rules, the open data anomaly detection system (odads) was developed. to evaluate the accuracy of the system, an experiment on the vessel traffic data was conducted and an accuracy of % was obtained for the system. there was a false negative case in the system results that decreased the accuracy. it was due to incorrect ais data in a special situation that was not possible to be handled by the detection rules in the scope of this thesis. the validity of the results was investigated by the subject matter experts from the swedish coastguard. the validation results showed that the majority of the odads evaluated anomalies were true alarms. moreover, a potential information gap in the closed data sources was observed during the validation process. despite the high number of true alarms, the number of false alarms was also considerable that was mainly because of the inaccurate open data. this thesis provided insights into the open data as a complement to the common data sources in the ms domain and is concluded that using open data will improve the efficiency of the surveillance systems by increasing the accuracy and covering some unseen aspects of maritime activities. in the future, it is important to investigate how the open data sources in the maritime domain can be used in a global perspective. in this thesis, the surveillance area was limited to a local area which is fully covered by the authorities’ data sources. when the data beyond the exclusive economic zone are needed, it is more valuable to use open data sources. by taking advantage of the subject matter experts’ knowledge about ms, it would be possible to figure out how the global open data should be exploited for the surveillance purpose. integration of the open data with maritime confidential data can improve the efficiency of ms and should be considered as a further improvement of the system. another improvement can be considering a probability for each detected anomaly according to the history of the vessels behavior and the current situation. moreover, further investigation on the other sources of open data such as social data, which is created and shared through social media platforms, and online videos from the ports activities in the high risk regions, will be useful. the data that are used in odads are relatively trusted, but in case of using other open data sources in the ms domain for ad, the quality assurance of the data should be investigated. as well as using knowledge based systems, taking advantage of data-driven approaches such as machine learning techniques can increase the efficiency of the ms systems. finally, the next step for improving the ms systems after being equipped with the ad functionality is to predict the future threats or incoming anomalies based on the analysis of the current situation. appendix a: open and closed data sources table a. maritime open and closed data sources that are available via the internet ar organization name categories nar fr nfr provider baltic and international maritime council www.bimco.org/ shipping information consulting services ● denmark association of ship brokers and agents, inc. www.asba.org/ ship brokers information ● usa bureau international des containers www.bic-code.org/ freight containers-coding identification marking ● france european maritime safety agency www.emsa.europa.eu/oil-recovery-vessels/vessel-technical- specifications.html maritime safety prevention of pollution from ships ● portugal icc commercial crime services www.icc-ccs.org/ fraud in international trade ● ● uk international association of classification societies www.iacs.org.uk/shipdata/default.aspx maritime safety regulation ● ● uk international group of p&i clubs www.igpandi.org/home liability and insurance issues ● ● uk international association of independent tanker owners www.intertanko.com/ transportation safety prevention of pollution from ships free competition ● uk (continued) http://www.bimco.org/� http://www.asba.org/� http://www.bic-code.org/� http://www.emsa.europa.eu/oil-recovery-vessels/vessel-technical-specifications.html� http://www.emsa.europa.eu/oil-recovery-vessels/vessel-technical-specifications.html� http://www.iacs.org.uk/shipdata/default.aspx� http://www.igpandi.org/home� http://www.intertanko.com/� ar organization name categories nar fr nfr provider oil companies international marine forum www.ocimf.com/home transportation safety prevention of pollution from ships ● uk internet ships register www.ships-register.com/ ships information ● uk world shipping register www.world-register.org/ ships information ports information companies information ● — lloyd’s register ships in class www.lrshipsinclass.lrfairplay.com/default.aspx ships information ● uk international telecommunication union www.itu.int/itu- r/index.asp?category=terrestrial&rlink=mars&lang=en ships information coasts information addresses of accounting authorities, administrations which notify information mmsi assigned to search and rescue aircraft mmsi assigned to ais aids to navigation ● ● switzerland international maritime consultancy specialists in vts www.maritime-vts.co.uk/ vessel traffic services and maritime organization ● uk port directory www.port-directory.com/ ports information ● uk equasis www.equasis.org/equasisweb/public/homepage?fs=homepage ships information companies information ● portugal q .com www.q .com/home.aspx?c= questionnaire generator ships information ● ● usa informare www.informare.it/indexuk.htm shipping information ● italy (continued) http://www.ocimf.com/home� http://www.ships-register.com/� http://www.world-register.org/� http://www.lrshipsinclass.lrfairplay.com/default.aspx� http://www.itu.int/itu-r/index.asp?category=terrestrial&rlink=mars&lang=en� http://www.itu.int/itu-r/index.asp?category=terrestrial&rlink=mars&lang=en� http://www.maritime-vts.co.uk/� http://www.port-directory.com/� http://www.equasis.org/equasisweb/public/homepage?fs=homepage� http://www.q .com/home.aspx?c= � http://www.informare.it/indexuk.htm� ar organization name categories nar fr nfr provider paris mou www.parismou.org/ port state control ● netherlands tokyo mou www.tokyo-mou.org/ port state control ● japan australian government bureau of meteorology www.bom.gov.au/ weather, climate and water ● ● australia finnish meteorological institute en.ilmatieteenlaitos.fi/home weather, climate and water ● finland meteo france france.meteofrance.com/ weather, climate and water ● ● france earth science office nasa weather.msfc.nasa.gov/goes/ weather, climate and water ● — the weather channel www.weather.com/ weather, climate and water ● — ocean color web nasa oceancolor.gsfc.nasa.gov/ weather, climate and water ● ● — earth european space agency earth.esa.int/ers/eeo . /atsr_med.html weather, climate and water ● ● italy weather bbc www.bbc.co.uk/weather/ weather, climate and water ● uk sailwx.info www.sailwx.info/ live marin information ● usa ship.gr www.ship.gr/ ship brokers information ship suppliers companies information ● — (continued) http://www.parismou.org/� http://www.tokyo-mou.org/� http://www.bom.gov.au/� http://france.meteofrance.com/� http://weather.msfc.nasa.gov/goes/� http://www.weather.com/� http://oceancolor.gsfc.nasa.gov/� http://earth.esa.int/ers/eeo . /atsr_med.html� http://www.bbc.co.uk/weather/� http://www.sailwx.info/� http://www.ship.gr/� ar organization name categories nar fr nfr provider american bureau of shipping www.eagle.org/eagleexternalportalweb/appmanager/abseagle/absea gledesktop?_nfpb=true&_pagelabel=abs_eagle_portal_home_page classification societies ● usa det norske veritas www.dnv.com/ classification societies ● ● norway bureau veritas groups www.bureauveritas.com/wps/wcm/connect/bv_com/group/footer/ho me/ classification societies ● ● — china classifiaction societies www.ccs.org.cn/en/index.htm classification societies ● china hellenic register of shipping www.hrs.gr/index.htm classification societies ● greece nippon kaiji kyokai www.classnk.or.jp/hp/en/index.aspx classification societies ● japan vesseltracker.com www.vesseltracker.com/en/vesselarchive.html ais data ships information ● ● germany digital seas www.digital-seas.com/start.html ais data ships information ● ● germany marinetraffic.com www.marinetraffic.com/ais/ ais data ships information ● greece shipspotting.com www.shipspotting.com/ ais data ships information ● ● — international maritime organization www .imo.org/sharepoint/mainframe.asp?topic_id= &offset piracy reports ● uk copenhagen malmö port www.cmport.com/ port authorities ● denmark (continued) http://www.eagle.org/eagleexternalportalweb/appmanager/abseagle/abseagledesktop?_nfpb=true&_pagelabel=abs_eagle_portal_home_page� http://www.eagle.org/eagleexternalportalweb/appmanager/abseagle/abseagledesktop?_nfpb=true&_pagelabel=abs_eagle_portal_home_page� http://www.dnv.com/� http://www.bureauveritas.com/wps/wcm/connect/bv_com/group/footer/home/� http://www.bureauveritas.com/wps/wcm/connect/bv_com/group/footer/home/� http://www.ccs.org.cn/en/index.htm� http://www.hrs.gr/index.htm� http://www.classnk.or.jp/hp/en/index.aspx� http://www.vesseltracker.com/en/vesselarchive.html� http://www.digital-seas.com/start.html� http://www.marinetraffic.com/ais/� http://www.shipspotting.com/� http://www .imo.org/sharepoint/mainframe.asp?topic_id= &offset� http://www.cmport.com/� ar organization name categories nar fr nfr provider port of gothenburg www.portgot.se/prod/hamnen/ghab/dalis b.nsf port authorities ● sweden genoa port authority www.porto.genova.it/index.php/en port authorities ● italy port of klaipeda www.portofklaipeda.lt/en.php port authorities ● lithuania philippine ports authority www.ppa.com.ph/ port authorities ● philippine panama canal authority www.pancanal.com/eng/index.html port authorities ● usa uk p&i club www.ukpandi.com/ liability and insurance issues ● ● uk the american club www.american-club.com/ liability and insurance issues ● ● usa steamship mutual www.simsl.com/ liability and insurance issues ● ● bermuda skuld www.skuld.com/ liability and insurance issues ● norway north of england p&i association www.nepia.com/home/ liability and insurance issues ● uk the standard club www.standard-club.com/ liability and insurance issues ● ● uk baltic ports organization www.bpoports.com/ port coordinator ● denmark note. nar = no authorization required; ar= authorization required; fr = free registration; nfr = non-free registration; dashes indicate undisclosed information. http://www.portgot.se/prod/hamnen/ghab/dalis b.nsf� http://www.porto.genova.it/index.php/en� http://www.portofklaipeda.lt/en.php� http://www.ppa.com.ph/� http://www.pancanal.com/eng/index.html� http://www.ukpandi.com/� http://www.american-club.com/� http://www.simsl.com/� http://www.skuld.com/� http://www.nepia.com/home/� http://www.standard-club.com/� http://www.bpoports.com/� table a. data sources that are used in the implementation website name data type website url marinetraffic.com real time information based on ais systems www.marinetraffic.com/ais/ swedish maritime administration (sjofartsverket) stockholm pilotage area www.sjofartsverket.se/sv/infrastruktur- amp-sjotrafik/lotsning/lotsinfo/ ports of stockholm, kapellskär and nynäshamn vessels in port and expected arrival www.stockholmshamnar.se/en/karta/vessel -calls/ port of norrköping vessels in port www.norrkoping- port.se/anlop.php?page=snabb_fih&link= | expected vessels arrival www.norrkoping- port.se/anlop.php?page=snabb_fih_ank&lin k= | port of helsinki cargo vessels in port www.portofhelsinki.fi/cargo_traffic/vessels _in_ports expected cargo vessels arrival www.portofhelsinki.fi/cargo_traffic/arrival_ ships expected passenger vessels departure www.portofhelsinki.fi/passengers/departure _times_and_terminals expected passenger vessels arrival www.portofhelsinki.fi/passengers/arrival_ti mes_and_terminals passenger vessels have visited the port before www.portofhelsinki.fi/passengers/cruise_shi ps_that_have_visited_the_port port of tallinn vessels in port www.ts.ee/?op=ships_in_port&lang=eng expected cargo vessels arrival www.ts.ee/?op=cargo_ships_arrivals&lang= eng expected passenger vessels arrival www.ts.ee/?op=passenger_ship_arrivals&la ng=eng expected passenger vessels departure www.ts.ee/?op=passenger_ship_departures &lang=eng http://www.marinetraffic.com/ais/� http://www.sjofartsverket.se/sv/infrastruktur-amp-sjotrafik/lotsning/lotsinfo/� http://www.sjofartsverket.se/sv/infrastruktur-amp-sjotrafik/lotsning/lotsinfo/� http://www.stockholmshamnar.se/en/karta/vessel-calls/� http://www.stockholmshamnar.se/en/karta/vessel-calls/� http://www.norrkoping-port.se/anlop.php?page=snabb_fih&link= | � http://www.norrkoping-port.se/anlop.php?page=snabb_fih&link= | � http://www.norrkoping-port.se/anlop.php?page=snabb_fih&link= | � http://www.norrkoping-port.se/anlop.php?page=snabb_fih_ank&link= | � http://www.norrkoping-port.se/anlop.php?page=snabb_fih_ank&link= | � http://www.norrkoping-port.se/anlop.php?page=snabb_fih_ank&link= | � http://www.portofhelsinki.fi/cargo_traffic/vessels_in_ports� http://www.portofhelsinki.fi/cargo_traffic/vessels_in_ports� http://www.portofhelsinki.fi/cargo_traffic/arrival_ships� http://www.portofhelsinki.fi/cargo_traffic/arrival_ships� http://www.portofhelsinki.fi/passengers/departure_times_and_terminals� http://www.portofhelsinki.fi/passengers/departure_times_and_terminals� http://www.portofhelsinki.fi/passengers/arrival_times_and_terminals� http://www.portofhelsinki.fi/passengers/arrival_times_and_terminals� http://www.portofhelsinki.fi/passengers/cruise_ships_that_have_visited_the_port� http://www.portofhelsinki.fi/passengers/cruise_ships_that_have_visited_the_port� http://www.ts.ee/?op=ships_in_port&lang=eng� http://www.ts.ee/?op=cargo_ships_arrivals&lang=eng� http://www.ts.ee/?op=cargo_ships_arrivals&lang=eng� http://www.ts.ee/?op=passenger_ship_arrivals&lang=eng� http://www.ts.ee/?op=passenger_ship_arrivals&lang=eng� http://www.ts.ee/?op=passenger_ship_departures&lang=eng� http://www.ts.ee/?op=passenger_ship_departures&lang=eng� table a. required data for detection of each anomaly anomaly data vessel_not_informed_port ais - port vessel arrival data arrival_time_mismatched ais - port vessel arrival data vessel_entered_port_ without_notice ais - port vessel arrival data - port vessel departure/ in-port data vessel_not_used_pilot ais - pilot vessels service data unusual_trip_pattern ais - port vessel arrival data – port vessel departure/ in-port data vessel_not_left_port ais - port vessel departure data vessel_not_entered_port ais - port vessel arrival data – port vessel departure/ in-port data vessel_ordered_pilot_and_ not_informed_port ais – port vessel arrival data – pilot vessels service data vessel_moored_in_port ais – port vessel departure/ in-port data appendix b: ports regions this section provides the geographical information about each port and its harbors. table b. ports areas and their geographical positions port number of areaa minimum longitude maximum longitude minimum latitude maximum latitude helsinki . . . . kapellskär . . . . norrköping . . . . . . . . . . . . nynäshamn . . . . stockholm . . . . . . . . . . . . . . . . . . . . . . . . tallinn . . . . . . . . . . . . . . . . . . . . . . . . note. a in order to have a more precise ad, harbors and areas around each port that are used by vessels are identified. figure b. . helsinki port area (the image is adapted from google earth). figure b. . kapellskär port area, stockholm group (the image is adapted from google earth). figure b. . norrköping port area (the image is adapted from google earth). figure b. . nynäshamn port area, stockholm group (the image is adapted from google earth). figure b. .stockholm port areas, stockholm group (the image is adapted from google earth). figure b. . tallinn port areas (the image is adapted from google earth). appendix c: algorithms this section provides a list of all algorithms that are used for implementing anomaly detection module. algorithm main procedure for the anomaly detector module while true . extract ais, ports and pilot data from database regarding to the latest time of data collection by the data collector module . call methods regarding to the following algorithms: a. algorithm b. algorithm c. algorithm d. algorithm e. algorithm f. algorithm g. algorithm . determine the final type of anomaly for vessels (algorithm ) . store detected anomalies in database . wait until the new data are collected by the data collector module end while algorithm detect vessel_not_informed_port and arrival_time_mis- matched anomalies arrival_time_delay_threshold ← minutes for each vessel ∈ ais data do destination ← vessel destination if destination ≠ null and isrelatedtoports(destination) and not isinportsarea(vessel) then vessel_found ← false port_schedule ← vessels that are expected to arrive at destination port // for vessels that have not updated their trip information at the beginning of their trip if destination contains vessel origin and vesselleftportrecently(destination port, vessel) and not istargetedport(destination port, vessel) then continue end if for each port_vessel ∈ port_schedule do if (vessel arrival date = port_vessel arrival date or current date = port_vessel arrival date or // to consider an interval for changing dates |vessel arrival time - port_vessel arrival time| < five_hours ) and isvesselsimilar(vessel, port_vessel) then vessel_found ← true if vessel arrival time ≠ port_vessel arrival time and |vessel arrival time - port_vessel arrival time| > arrival_time_delay_threshold then add(vessel anomaly, arrival_time_mismatched) algorithm detect vessel_not_informed_port and arrival_time_mis- matched anomalies end if break end if end for if vessel_found ≠ true then add(vessel anomaly, vessel_not_informed_port) end if end if end for algorithm detect vessel_entered_port_without_notice anomaly for each port ∈ available ports port_vessels ← vessels that are currently available in port and arrived one hour ago for each port_vessel ∈ port_vessels do expected_vessels ← vessels that were expected to enter the port at vessel arrival time vessel_found ← false for each exp_vessel ∈ expected_vessels do if isvesselsimilar(port_vessel, exp_vessel) then vessel_found ← true break end if end for if vessel_found ≠ true then add(port_vessel anomaly, vessel_entered_port_without_notice) (port_vessel) end if end for end for algorithm detect vessel_not_used_pilot and vessel_ordered_pilot_and_not_informed_port anomaly pilot_start_time_delay_threshold ← minutes for each vessel ∈ ais data do for each pilot_vessel ∈ pilot schedule do if current date = pilot_vessel date and isvesselsimilar(vessel, pilot_vessel) then if vessel anomaly contains vessel_not_informed_port then add(vessel anomaly,vessel_ordered_pilot_and_not_informed_port) end if if pilot_vessel has not received service and pilot_vessel start_operation_time + pilot_start_time_delay_threshold < current time then add(vessel anomaly , vessel_not_used_pilot) end if // to solve the issue for duplicate records with different status in pilot data if pilot_vessel has received service and vessel anomaly contains vessel_not_used_pilot then remove(vessel anomaly, vessel_not_used_pilot) break algorithm detect vessel_not_used_pilot and vessel_ordered_pilot_and_not_informed_port anomaly end if end if end for end for algorithm detect unusual_trip_pattern anomaly trip_count_threshold ← for each vessel ∈ ais data do if vessel origin ≠ null and vessel destination ≠ null and vessel type ≠ ferry and (isrelatedtoports(vessel origin) or isrelatedtoports(vessel destination)) and not isinportsarea(vessel) then count ← gettriphistoryforvessel(vessel) if count ≠ - and count < trip_count_threshold then add(vessel anomaly , unusual_trip_pattern) end if end if end for algorithm detect vessel_not_entered_port anomaly for each port ∈ available ports expected_vessels ← vessels that were expected to enter the port since one hour ago port_vessels ← vessels that are currently available in port for each exp_vessel ∈ expected_vessels do vessel_found ← false for each port_vessel ∈ port_vessels do if port_vessel time ≥ exp_vessel time and isvesselsimilar(exp_vessel, port_vessel) then vessel_found ← true break end if end for if vessel_found ≠ true then add(port_vessel anomaly, vessel_not_entered_port) findvesselmatchinaisdata(port_vessel) end if end for end for algorithm detect vessel_not_left_port anomaly for each port ∈ available ports port_vessels ← vessels that were expected to leave the port since one hour ago for each port_vessel ∈ port_vessels do for each vessel ∈ ais data do if isvesselsimilar(vessel, port_vessel) and ismooredinportsarea (vessel) then add(vessel anomaly , vessel_not_left_port) algorithm detect vessel_not_left_port anomaly break end if end for end for end for algorithm detect vessel_moored_in_port anomaly for each vessel ∈ ais data do if not isinportsarea(vessel) then vessel_found ← false for each port ∈ available ports do if vessel_found ≠ true then port_vessels ← vessels that are available in port for at least minutes later for each port_vessel ∈ port_vessels do if isvesselsimilar(vessel, port_vessel) and vessel type = port_vessel type and isothersimilarvesselinport(vessel) then add(vessel anomaly , vessel_moored_in_port) vessel_found ← true break end if end for end if end for end if end for algorithm isinportsarea(vessel) in_area ← false for each port ∈ available ports do rectangle ← get the pre-defined port area if vessel position is within the rectangle then in_area ← true break; end if end for return in_area algorithm ismooredinportsarea(vessel) in_area ← false for each port ∈ available ports do create a rectangle from pre-defined coordinates for the port area if vessel is not moving and vessel position is within the rectangle then in_area ← true break; algorithm ismooredinportsarea(vessel) end if end for return in_area algorithm isvesselsimilar(vessel_ , vessel_ ) vessel_name_similarity_threshold ← . similarity ← calculate the similarity between the vessel_ name and vessel_ name using a string matching technique is_similar ← false if similarity < vessel_name_similarity_threshold then is_similar ← true end if return is_similar algorithm isrelatedtoports(name) is_related ← false for each port ∈ available ports do if name = port name or name contains port name or istripsimilar(name, port name) then is_ related ← true break end if end for return is_related algorithm istripsimilar (trip , trip ) trip_similarity_threshold ← . similarity ← calculate the similarity between the two trips using a string matching technique is_similar ← false if similarity < trip_similarity_threshold then is_similar ← true end if return is_similar algorithm vesselleftportrecently(port, vessel) left_port ← false port_vessels ← vessels that were expected to leave the port since twelve hours ago for each port ∈ available ports do if isvesselsimilar(vessel, port_vessel) then left_port ← true break end if end for algorithm vesselleftportrecently(port, vessel) return left_port algorithm istargetedport(port, vessel) is_targeted ← false horizontal_line ← get the pre-defined horizontal line for the port vertical_line ← get the pre-defined vertical line for the port if vessel position is on the left side of vertical_line and vessel position is below the horizontal_line and vessel heading is to the north east then is_targeted ← true else if vessel position is on the right side of vertical_line and vessel position is below the horizontal_line and vessel heading is to the north west then is_targeted ← true else if vessel position is on the right side of vertical_line and vessel position is above the horizontal_line and vessel heading is to the south west then is_targeted ← true else if vessel position is on the left side of vertical_line and vessel position is above the horizontal_line and vessel heading is to the south east then is_targeted ← true end if return is_targeted algorithm gettriphistoryforvessel(vessel) vessel_found ← false count ← for each history_vessel ∈ vessel trip history do if isvesselsimilar(vessel, history_vessel) then vessel_found ← true trip ← vessel origin and destination if istripsimilar(trip, history_vessel trip) then count ← history_vessel trip_count break end if end if end for if vessel_found ≠ true then count ← - return count algorithm isothersimilarvesselinport(ais_vessel) is_similar ← false for each vessel ∈ ais data do if (vessel name = ais_vessel name and vessel type = ais_vessel type and algorithm isothersimilarvesselinport(ais_vessel) vessel mmsi ≠ ais_vessel mmsi)) and isinportarea(vessel) then is_similar ← true end if end for return is_similar algorithm findvesselmatchinaisdata(port_vessel) for each vessel ∈ ais data do if isvesselsimilar(vessel, port_vessel) and vessel type = port_vessel type and vessel type ≠ ferry then if port_vessel anomaly = vessel_entered_port_without_notice and not isinportsarea(vessel) then continue else if port_vessel anomaly = vessel_not_entered_port and (vessel has arrival_time_mismatched or isinportsarea(vessel)) then continue else add(vessel anomaly , port_vessel anomaly) end if end if end for algorithm determine the final type of anomaly for vessels for each vessel ∈ ais data do //a if vessel ∈ smuggling list and vessel anomaly count ≠ then vessel anomaly ← under_surveillance_vessel continue end if if vessel anomaly count > then //a if vessel anomaly contains vessel_moored_in_port then anomaly ← vessel_moored_in_port //a ,a else if vessel anomaly contains vessel_not_informed_port and vessel anomaly contains unusual_trip_pattern then anomaly ← unusual_trip_pattern_and_vessel_not_informed _port //a ,a else if vessel anomaly contains arrival_time_mismatched and vessel anomaly contains unusual_trip_pattern then anomaly ← unusual_trip_vessel_arrival_time_mismatched //a ,a else if vessel anomaly contains vessel_not_entered_port and algorithm determine the final type of anomaly for vessels vessel anomaly contains unusual_trip_pattern then anomaly ← unusual_trip_pattern_and_vessel_not_entered _port //a ,a else if vessel anomaly contains vessel_ordered_pilot_and_not_ informed_port and vessel anomaly contains unusual_trip _pattern then anomaly ←unusual_trip_pattern_and_vessel_ordered_pilot _and_not_informed_port //a ,a else if vessel anomaly contains vessel_entered_port_without_notice and vessel anomaly contains vessel_not_left_port then anomaly ← vessel_entered_without_information_not_left _port_ on_time //a ,a else if vessel anomaly contains arrival_time_mismatched and vessel anomaly contains vessel_not_used_pilot then anomaly ← vessel_arrival_time_mismatched_and _vessel_not_used_pilot //a ,a ,a else if vessel anomaly contains arrival_time_mismatched and vessel anomaly contains vessel_not_used_pilot and vessel anomaly contains unusual_trip_pattern then anomaly ← unusual_trip_pattern_and_vessel_not_used_pilot _and_vessel_arrival_time_mismatched //a ,a else if vessel anomaly contains vessel_not_used_pilot and vessel anomaly contains unusual_trip_pattern then anomaly ← unusual_trip_pattern_and_vessel_not_used_pilot //a ,a else if vessel anomaly contains vessel_not_entered_port and vessel anomaly contains vessel_not_used_pilot then anomaly ← vessel_not_entered_port_and_not_used_pilot //a ,a ,a else if vessel anomaly contains vessel_not_entered_port and vessel anomaly contains vessel_not_used_pilot and vessel anomaly contains unusual_trip_pattern then anomaly ← unusual_trip_pattern_and_vessel_not_entered _ port_ and_vessel_not_used_pilot //a ,a else if vessel anomaly contains vessel_ordered_pilot_and_not_ informed_port and vessel anomaly contains vessel_not_used _pilot then anomaly ← vessel_ordered_pilot_and_not_informed_port _and_vessel_not_used_pilot algorithm determine the final type of anomaly for vessels //a ,a ,a else if vessel anomaly contains vessel_ordered_pilot_and_not_ informed_port and vessel anomaly contains vessel_not_used _pilot and vessel anomaly contains unusual_trip_pattern then anomaly ← unusual_trip_pattern_and_vessel_ordered_pilot _and_not_informed_port_and_vessel_not_used _pilot else // this part is a default state and executes when the combination of anomalies is not // logical and it is due to the wrong and inconsistent data anomaly ← select one of the anomalies randomly end if vessel anomaly ← anomaly end if end for appendix d: user interface figure d. . the main view of the odads user interface, the map displays the surveillance area, green vessels are ferry and passenger, cargo vessels are cyan and tankers are yellow. grey and dark blue colors belong to tug and pilot vessels, respectively. the four tabs on the left side contain the data related to detected anomalies, ports and pilot timetables, system reports and settings. in order to inform operators about newly detected anomalies, the anomaly tab on the left side and the anomalous vessels become red. moreover, previously detected vessels are shown in light red. by clicking on each vessel it is possible to view more information about the vessel and its track. appendix e: acronyms ad anomaly detection ads automatic dependant surveillance ais automatic identification system cet central european time cog course over ground d-l damerau-levenshtein eet eastern european time eez exclusive economic zone elint electronic intelligence eta estimated time of arrival fn false negatives fp false positives hci human-computer interaction hfswr high frequency surface wave radar imo international maritime organization ir infrared jdl joint directors of laboratories mda maritime domain awareness mmsi maritime mobile service identity ms maritime surveillance odads open data anomaly detection system oth over the horizon radar radio detection and ranging rmp recognized maritime picture sar synthetic aperture radar sigint signal intelligence tn true negatives tp true positives utc coordinated universal time vts vessel traffic service references akselrod, d., tharmarasa, r., kirubarajan, t., zhen ding, & ponsford, t. ( ). multisensor-multitarget tracking testbed. proceedings of the ieee symposium: computational intelligence for security and defence applications, canada, - . doi: . /cisda. . alonso, j. m., ambur, o., amutio, m. a., azañón, o., bennett, d., flagg, r., mcallister, d., et al. ( , may ). improving access to government through better use of the web. retrieved from world wide web consortium website: www.w .org/tr/ egov-improving/ andersson, m., & johansson, r. ( , november). multiple sensor fusion for effective abnormal behaviour detection in counter-piracy operations. paper presented at the international waterside security conference, carrara, italy. doi: . /wssc. . andler, s. f., fredin, m., gustavsson, p. m., van laere, j., nilsson, m., & svenson, p. ( ). smartracin: a concept for spoof resistant tracking of vessels and detection of adverse intentions. proceedings of the spie - the international society for optical engineering, usa, , g- - g- . doi: . / . avdic, a., hedström, k., rose, j., & grönlund, Å. (eds.). ( ). understanding eparticipation : contemporary phd eparticipation research in europe. Örebro: Örebro universitet. bick, e. t., & barock, r. t. ( ). centurion harbor surveillance test bed. proceedings of mts/ieee oceans, usa, , - . doi: . /oceans. . brax, c., niklasson, l., & smedberg, m. ( ). finding behavioural anomalies in public areas using video surveillance data. proceedings of the eleventh international conference on information fusion, germany. doi: . /icif. . carthel, c., coraluppi, s., & grignan, p. ( ). multisensor tracking and fusion for maritime surveillance. proceedings of the tenth international conference on information fusion, canada. doi: . /icif. . chandola, v., banerjee, a., & kumar, v. ( ). anomaly detection: a survey. acm computing surveys, ( ), : – : . doi: . / . constantine, l. l., & lockwood, l. a. d. ( ). software for use: a practical guide to the models and methods of usage-centered design. addison wesley. dahlbom, a., & niklasson, l. ( ). trajectory clustering for coastal surveillance. proceedings of the tenth international conference on information fusion, canada. doi: . /icif . . damerau, f. j. ( ). a technique for computer detection and correction of spelling errors. communications of the acm, ( ), – . doi: . / . danu, d., sinha, a., kirubarajan, t., farooq, m., & brookes, d. ( ). fusion of over-the- horizon radar and automatic identification systems for overall maritime picture. proceedings of the tenth international conference on information fusion, canada. doi: . /icif. . dedijer, s., & jéquier, n. ( ). intelligence for economic development : an inquiry into the role of the knowledge industry. oxford: berg. di lallo, a., farina, a., fulconi, r., stile, a., timmoneri, l., & vigilante, d. ( , october). a real time test bed for d and d multi-radar tracking and data fusion with application to border control. paper presented at the cie international conference on radar, shanghai, china. doi: . /icr. . dietrich, d., gray, j., mcnamara, t., poikola, a., pollock, p., tait, j., & zijlstra, t. ( ). open data handbook. retrieved from opendatahandbook.org/en/ ding, z., kannappan, g., benameur, k., kirubarajan, t., & farooq, m. ( ). wide area integrated maritime surveillance: an updated architecture with data fusion. proceedings of the sixth international conference of information fusion, australia, , - . doi: . /icif. . edlund, j., gronkvist, m., lingvall, a., & sviestins, e. ( ). rule-based situation assessment for sea surveillance. proceedings of the spie - the international society for optical engineering, usa, , - - - . doi: . / . fooladvandi, f., brax, c., gustavsson, p., & fredin, m. ( ). signature-based activity detection based on bayesian networks acquired from expert knowledge. proceedings of the twelfth international conference on information fusion, usa, - . gad, a. s. ( ). a fuzzy logic-based multisensor data fusion for maritime surveillance - real data testing. proceedings of the twenty-sixth national radio science conference, egypt. giompapa, s., farina, a., gini, f., graziano, a., & di stefano, r. ( , april). computer simulation of an integrated multi-sensor system for maritime border control. paper presented at the ieee radar conference, boston, ma, usa, - . doi: . / radar. . giompapa, s., farina, a., gini, f., graziano, a., croci, r., & di stefano, r. ( , may). study of the classification task into an integrated multisensor system for maritime border control. paper presented at the ieee radar conference, rome, italy. doi: . /radar. . guerriero, m., willett, p., coraluppi, s., & carthel, c. ( ). radar/ais data fusion and sar tasking for maritime surveillance. proceedings of the eleventh international conference on information fusion, germany. doi. . /icif. . guyard, a. ., roy, j., & defence r&d canada-valcartier, v. q. ( ). towards case-based reasoning for maritime anomaly detection: a positioning paper. proceedings of the twelfth iasted international conference on intelligent systems and control. usa. hall, d. l., & llinas, j. ( ). an introduction to multisensor data fusion. proceedings of the ieee, ( ), - . doi: . / . hall, d. l., & mcmullen, s. a. h. ( ). mathematical techniques in multisensor data fusion. boston, london: artech house. hall, m. j., hall, s. a., & tate, t. ( ). removing the hci bottleneck: how the human computer interface (hci) affect the performance of data fusion systems. mss national symposium on sensor and data fusion (pp. - ). usa. han, j., kamber, m., & pei, j. ( ). data mining: concepts and techniques. usa: elsevier. hatch, m. d., kaina, j. l., mahler, r. p., & myre, r. s. ( , november). data fusion methodologies to support theater level and deployable surveillance systems. paper presented at the conference record of thirty-second asilomar conference on signals, systems and computers, pacific grove, ca, usa, , - . doi: . /acssc. . henrich, w., kausch, t., & opitz, f. ( ). data fusion for the finnish fast attack craft squadron : concept and architecture. proceedings of the seventh international conference on information fusion, sweden, , - . İnce, a. n., topuz, e., & panayirci, e. ( ). principles of integrated maritime surveillance systems. boston, dordrecht, london: springer. jaro, m. a. ( ). unimatch: a computer system for generalized record linkage under conditions of uncertainty. proceedings of the spring joint computer conference, usa, – . doi: . / . jaro, m. a. ( ). advances in record-linkage methodology as applied to matching the census of tampa, florida. journal of the american statistical association, ( ), - . doi: . / johansson, f., & falkman, g. ( , december). detection of vessel anomalies - a bayesian network approach. paper presented at the third international conference on intelligent sensors, sensor networks and information, melbourne, qld., australia, - . doi: . /issnip. . jouan, a., valin, p., gagnon, l., & bosse, e. ( ). airborne fusion of imaging and non- imaging sensor information for maritime surveillance. proceedings of the conference on quality control by artificial vision, canada, - . lane, r. o., nevell, d. a., hayward, s. d., & beaney, t. w. ( ). maritime anomaly detection and threat assessment. proceedings of the thirteenth international conference on information fusion, uk. laxhammar, r. ( ). anomaly detection for sea surveillance. proceedings of the eleventh international conference on information fusion, germany. lefebvre, e., & helleur, c. ( ). multisource information adaptive fuzzy logic correlator for recognized maritime picture. proceedings of the fourth international conference on information fusion, canada, , thb - - . retrieved from ftp.isif.org/fusion /proceedings/fusion cd/fusion/searchengine/pdf/thb .pdf lefebvre, e., & helleur, c. ( ). automated association of track information from sensor sources with non-sensor information in the context of maritime surveillance. proceedings of the seventh international conference on information fusion, sweden, , - . levenshtein, v. i. ( ). binary codes capable of correcting deletions, insertions, and reversals. soviet physics-doklady, ( ), – . retrieved from www.freearchive.org/o/ e a ffa f ee fbaeed a e df fd b f mano, j. p., georgé, j. p., & gleizes, m. p. ( ). adaptive multi-agent system for multi- sensor maritime surveillance. in y. demazeau, f. dignum, j. m. corchado, & j. b. pérez (eds.), advances in intelligent and soft computing series: vol. . advances in practical applications of agents and multiagent systems (pp. - ). berlin, heidelberg, germany: springer. doi: . / - - - - _ maresca, s., greco, m., gini, f., grasso, r., coraluppi, s., & horstmann, j. ( , june). vessel detection and classification: an integrated maritime surveillance system in the tyrrhenian sea. paper presented at the second international workshop on cognitive information processing, elba, italy, - . doi: . /cip. . molloy, j. c. ( ). the open knowledge foundation: open data means better science. public library of science biology, ( ). doi: . /journal.pbio. nilsson, m., van laere, j., ziemke, t., & edlund, j. ( ). extracting rules from expert operators to support situation awareness in maritime surveillance. proceedings of the eleventh international conference on information fusion, germany. doi: . /icif. . ponsford, a. m., d’souza, i. a., & kirubarajan, t. ( , may). surveillance of the nautical mile eez using hfswr in association with a spaced-based ais interceptor. paper presented at the ieee conference on technologies for homeland security, boston, ma, usa, - . doi: . /ths. . rhodes, b. j., bomberger, n. a., seibert, m., & waxman, a. m. ( , october). maritime situation monitoring and awareness using learning mechanisms. paper presented at the ieee military communications conference, atlantic city, nj, usa, , - . doi: . /milcom. . rhodes, b. j., bomberger, n. a., seibert, m., & waxman, a. m. ( , october). seecoast: automated port scene understanding facilitated by normalcy learning. presented at the ieee military communications conference, washington, dc, usa. doi: . /milcom. . ristic, b., la scala, b., morelande, m., & gordon, n. ( ). statistical analysis of motion patterns in ais data: anomaly detection and motion prediction. proceedings of the eleventh international conference on information fusion, germany. doi: . /icif . . riveiro, m., & falkman, g. ( ). interactive visualization of normal behavioral models and expert rules for maritime anomaly detection. proceedings of the sixth international conference on computer graphics, imaging and visualization, china. - . doi: . /cgiv. . riveiro, m., & falkman, g. ( ). supporting the analytical reasoning process in maritime anomaly detection: evaluation and experimental design. proceedings of the fourteenth international conference information visualisation, uk, - . doi: . /iv. . riveiro, m., falkman, g., & ziemke, t. ( a). visual analytics for the detection of anomalous maritime behavior. proceedings of the twelfth international conference information visualisation, uk, - . doi: . /iv. . riveiro, m., falkman, g., & ziemke, t. ( b). improving maritime anomaly detection and situation awareness through interactive visualization. proceedings of the eleventh international conference on information fusion, germany. riveiro, m., johansson, f., falkman, g., & ziemke, t. ( ). supporting maritime situation awareness using self organizing maps and gaussian mixture models. proceedings of the tenth scandinavian conference on artificial intelligence, – . retrieved from www.his.se/pagefiles/ /scai .pdf roy, j. ( ). anomaly detection in the maritime domain. proceedings of spie - the international society for optical engineering, usa, , w- - w- . doi: . / . roy, j. ( ). rule-based expert system for maritime anomaly detection. proceedings of spie - the international society for optical engineering, usa, , n- - n- . doi: . / . roy, j., & davenport, m. ( , november). exploitation of maritime domain ontologies for anomaly detection and threat analysis. paper presented at the waterside security conference, carrara, italy. doi: . /wssc. . salton, g., & mcgill, m. j. ( ). introduction to modern information retrieval. new york, ny, usa: mcgraw-hill, inc. shadish, w. r., cook, t. d., & campbell, d. t. ( ). experimental and quasi- experimental designs for generalized causal inference. houghton mifflin. van laere, j., & nilsson, m. ( ). evaluation of a workshop to capture knowledge from subject matter experts in maritime surveillance. proceedings of the twelfth international conference on information fusion,usa, - . vespe, m., sciotti, m., & battistello, g. ( ). multi-sensor autonomous tracking for maritime surveillance. paper presented at the international conference on radar, adelaide, sa, australia, - . doi: . /radar. . vespe, m., sciotti, m., burro, f., battistello, g., & sorge, s. ( ). maritime multi-sensor data association based on geographic and navigational knowledge. paper pesented at the ieee radar conference, rome, italy. doi: . /radar. . winkler, w. e. ( ). string comparator metrics and enhanced decision rules in the fellegi- sunter model of record linkage. retrieved from www.eric.ed.gov/pdfs /ed .pdf yin, r. k. ( ). case study research: design and methods. sage publications. zamora, e. m., po llock, j. j., & zamora, a. ( ). the use of trigram analysis for spelling error detection. information processing and management – . . introduction . problem statement . research questions . aims and objectives . contribution . outline . background . terminology . jdl data fusion model . related work . research methodology . open data in the ms domain . identification of maritime anomalies framework design . implementation . . data description . . . ais data . . . ports and pilots data . . data collector module . . database . . . algorithms . . . string matching techniques . . display client . experimental design . validation design . validity threats . verification . experimental results . discussion . conclusion and future work appendix a: open and closed data sources appendix b: ports regions appendix c: algorithms appendix d: user interface appendix e: acronyms references fuzzy linear regression-based detection of earnings management fuzzy linear regression-based detection of earnings management henrik höglund ⇑ hanken school of economics, handelsesplanaden , vasa, finland a r t i c l e i n f o keywords: earnings management discretionary accruals fuzzy linear regression a b s t r a c t a large number of accounting studies have examined the occurrence of earnings management in various contexts. in most of these studies, the earnings management detection model is based on the linear regression model suggested by jones ( ). a considerable problem with the jones model is the require- ment of long time series of financial statement data. an alternative to estimating the linear regression model coefficients with ordinary least squares (ols) is to use fuzzy linear regression (flr) instead. one of the main advantages with flr described in the literature is its ability to handle small data sets. the purpose of this study is to compare the performance of the ols-based jones model with the perfor- mance of the flr-based jones model. the results show that the performance of both types of models decreases when the length of the time series decreases and that there is no significant difference in the estimated discretionary accruals between the models. the results also show that the flr-based jones model outperforms the ols-based jones model in detecting simulated earnings management when the estimation time series is short. overall, the results show that the flr-based jones model is a feasible alternative to the ols-based jones model, especially when the length of the estimation time series is restricted by data availability. � elsevier ltd. all rights reserved. . introduction a large number of accounting studies have examined the occur- rence of earnings management in various contexts. there is, for example, evidence suggesting that firms manage earnings in order to avoid violating debt covenants (defond & jiambalvo, ) or to influence initial public offering (ipo) valuations (teoh, welch, & wong, ). a major challenge in earnings management studies is to find a measure for how much firms have managed their earn- ings. a common assumption is that earnings are managed through accounting accruals. with this assumption the measure of earnings management is the unexpected part of a firm’s total accruals. var- ious models have been suggested for dividing the total accruals into non-discretionary (expected) and discretionary (unexpected) accruals. most of these models are based on the earnings manage- ment detection model suggested by jones ( ). the jones model is a linear regression model where the level of total accruals is as- sumed to be explained by property, plant and equipment and the change in sales. in the original jones model, the regression coeffi- cients are estimated using a firm specific time series comprising data prior to the event year. the expected level of accruals is then calculated using the estimated coefficients with event year data. a considerable problem with the jones model is the require- ment of long time series of financial statement data. typically, in studies where the jones model is used the requirement is set to at least ten years of data prior to the event year (e.g. dechow, sloan, & sweeney, ; thomas & zhang, ). this requirement might lead to several problems, such as survivorship bias and non-station- ary regression coefficients (peasnell, pope, & young, ; young, ). an alternative to estimating the firm specific regression coefficients with ordinary least squares (ols) regression is using fuzzy linear regression (flr), first introduced by tanaka, uejima, and asai ( ). contrary to probability theory-based ols regres- sion, flr is based on possibility theory and fuzzy set theory. with flr the objective is to minimize the fuzziness of the model rather than the sum of squared residuals. one of the main advantages with flr described in the literature is its ability to handle small data sets (shapiro, ). this was corroborated empirically by kim, mosko- witz, and koksalan ( ) who provided evidence that flr outper- forms statistical linear regression in predictive capability with small data sets. thus, it is possible that the jones model regression could be run with shorter time series using flr. if this is the case, an flr-based jones model could reduce several problems originating from the requirement of long time series. the purpose of this study is to compare the performance of the ols-based jones model with the performance of the flr-based jones model. the focus is on the time series version of the jones model and the comparison is made using different lengths of the time series. the remainder of this study is organized as follows. the basic operating principle of the linear regression-based accrual models is covered in section . in section the estimation of flr - /$ - see front matter � elsevier ltd. all rights reserved. http://dx.doi.org/ . /j.eswa. . . ⇑ tel.: + ( ) . e-mail address: henrik.hoglund@hanken.fi expert systems with applications ( ) – contents lists available at sciverse sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.dyndns.org/dialog/?doi= . /j.eswa. . . &domain=pdf http://dx.doi.org/ . /j.eswa. . . mailto:henrik.hoglund@hanken.fi http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa coefficients and prediction with flr models are covered. the re- search design is presented in section and the results from the empirical study are presented in section . section concludes the study. . time series-based discretionary accrual estimation models during the past years a large number of studies have exam- ined the occurrence of earnings management in various contexts. typically, the assumption in these studies has been that deviations from a normal or expected level of accruals constitute earnings management. in the early earnings management detection models the expected level of accruals was defined as the average accruals over an estimation period (healy, ) or as the previous year accruals (deangelo, ). the shortcoming of these models is that they do not consider changing circumstances of the firms. for example, a normal increase in accruals as a result of an increase in sales would turn up as earnings management. jones ( ) re- laxed the assumptions that accruals remain stationary over time by suggesting a regression approach where the level of total accru- als (tacc) is explained by the change in sales (drev) and property, plant and equipment (ppe). the change in sales is assumed to ex- plain the current accruals, such as receivables, inventory and pay- ables, whereas property, plant and equipment should mainly explain the level of depreciation. tacct tat� ¼ a tat� þ a drev t tat� þ a ppet tat� typically, the variables in the jones model are deflated with lagged total assets (ta) in order to reduce heteroscedasticity. the coeffi- cients a . . .a in the jones model are estimated using a sufficiently long firm specific time series. the coefficients are then used to calcu- late the expected level or the non-discretionary part of total accruals using event period data. the difference between the observed event period total accruals and the calculated non-discretionary accruals is considered as unexpected accruals or discretionary accruals equal- ing earnings management. jones ( ) used years of data for estimating the regression coefficients whereas subsequent studies have typically settled for a minimum of years (e.g. dechow et al., ; thomas & zhang, ). a firm specific time series this long might lead to several problems. first, issues with survivorship bias are likely to arise (bartov, gul, & tsui, ; peasnell et al., ; young, ). second, a large number of firms do not have data stretching back ten years, reducing the size of the data set. third, it is unlikely that the regression model coefficients remain stationary over a long period of time (peasnell et al., ). finally, the self-reversing nature of accruals can lead to problems with serial correlation of residuals (peasnell et al., ). to remedy some of these problems and shortcomings, defond and jiambalvo ( ) suggested a cross-sectional approach where the regression coefficients are industry and year specific rather than firm specific. this approach does, however, also have its shortcomings. first, the industry membership is usually defined at a -digit sic level. this might result in that some industries do not have a sufficient number of observations to run the regression, reducing the size of the data set. second, the assumption is that the accrual generating process is similar among the firms in the same industry. there is, however, evidence suggesting that this assump- tion does not necessarily hold (ecker, francis, & olsson, ; höglund, ; kothari, leone, & wasley, ). for example, eck- er et al. ( ) showed that the performance of the jones model improved when running the regression with firms matched on lagged total assets rather than on industry membership. . fuzzy linear regression the concept of fuzzy linear regression (flr) builds on the fuzzy set theory developed by zadeh ( ) and it was first introduced by tanaka, uejima, and asai ( ). the idea behind flr is that the deviations between the observed and estimated values are assumed to originate from imprecise observations or vague relations be- tween the model variables. this is contrary to ols regression where random errors are assumed to be the reason for differences be- tween observed and estimated values. thus, the uncertainty in flr models is fuzziness rather than randomness (yang & lee, ). . . estimating fuzzy linear regression coefficients in ols regression the objective is to minimize the sum of squared residuals. in flr, on the other hand, the objective is to minimize the fuzziness of the model. the general form of the flr model is: by i ¼ ~a i þ ~a i x i þ���þ ~ajixji where x . . .xj denotes the independent variables, à the estimated fuzzy intercept coefficient, à . . .Ãj the estimated fuzzy slope coeffi- cients and ŷ i the estimated fuzzy output. the fuzzy coefficients are generally represented as symmetric triangular fuzzy numbers where aj equals the center value of Ãj and cj the spread (see fig. ). however, other than symmetric triangular membership functions, such as asymmetric triangular and trapezoidal (ishibuchi & nii, ), are also used. before estimating the coefficients of the flr, the fit between the flr model and the data set has to be defined. this is done by set- ting the value of the h term, also called the target degree of belief (e.g. chang & ayyub, ), between and . each observed fuzzy ~y i or crisp yi output must fall within the estimated fuzzy output ŷ i at h (see fig. ). as the value of h increases, the fuzziness of the flr model also increases. the h value does, however, not affect the cen- ter value of the fuzzy coefficients (tanaka & watada, ). the flr coefficients à . . .Ãj are estimated using linear program- ming with the objective of minimizing the total spread or fuzziness of the flr model. the following objective function, where s de- notes the total fuzziness of the regression model, m the number of independent variables and n the number of observations, is to be minimized: s ¼ xm j¼ cj xn j¼ jxjij the flr linear programming problem comprises two sets of con- straints. first, the spread cj of the fuzzy coefficients Ãj has to be zero or positive. second, all observed output variables ~y i must fall within the estimated fuzzy output variables ŷ i at level h. yi is the center αj cj m em b er sh ip . . Ãj cj fig. . fuzzy coefficient with center value and spread. h. höglund / expert systems with applications ( ) – https://isiarticles.com/article/ pii: - ( ) -y pergamon expert systems with applications, vol. . no. . pp. - , copyright elsevier science ltd printed in the usa. all tights reserved - i / $ . + .oa - ( ) - building a fuzzy expert system for electric load forecasting using a hybrid neural network i? k. dash and a. c. liew national university of singapore, , kent ridge crescent, singapore s. rahman department of electrical engineering, virginia polytechnic institute and state university, blacksburg, va g. ramakrishna regional engineering college, rourkela- , india abstract-this paper presents the development of a hybrid neural network to model a fuzzy expert system for time series forecasting of electricc load. the hybrid neural network is trained to develop fuzzy logic rules andjind optimal inputloutput membership values of load and weather parameters. a hybrid learning algorithm consisting of unsupervised and supervised learning phases is used for training the fuzzifred neural network. in the supervised learning phase, both back-propagation and linear kalman jilter algorithms are used for the adjustment of weights and membership functions. extensive tests have been performed on a -year utility data for the generation of peak and average loadprojiles in h. h. and h ahead time frame during summer and winter seasons. from the simulation results, it is observed that the fuzzy expert system using the kalman $lter-based algorithm gives faster convergence and more accurate prediction of a load time series. . introduction it is well known that fuzzy logic provides an inference strategy that enables approximate human reasoning capabilities to knowledge-based systems. also, it pro- vides a mathematical morphology to emulate certain perceptual and linguistic attributes associated with human cognition. although fuzzy theory provides an inference mechanism under cognitive uncertainty, com- putational neural networks offer exciting advantages such as learning, adaptation, fault tolerance, parallelism, and generalization. the computational neural networks, comprising processing elements called neurons, are capable of coping with computational complexity, non- linearity, and uncertainty. to enable a system to deal with cognitive uncertain- ties in a manner more like humans, one may incorporate requests for reprints should be sent to p k. dash, centre for intelligent systems, electrical engineering department of regional engineering college, rourkela , india. the concepts of fuzzy logic into neural network. although fuzzy logic is a natural mechanism for modelling cognitive uncertainty, it may involve an increase in the amount of computation required. this can be readily offset by using fuzzy neural network approaches having the potential for parallel computa- tions with high flexibility. the application of an artificial neural network (ann) and fuzzy logic-based decision support system to time- series forecasting has gained much attention recently. ann-based load forecasts give large errors when the weather profile changes very fast. also extremely slow training or even training failure occurs in many cases due to difficulties in selecting proper structures of the neural network paradigm being used, and due to the errors in associated parameters such as learning rates, activation functions, etc., which are fundamental to any back- propagation neural network. on the other hand, the development of a fuzzy decision system (fuzzy expert system) for load forecasting requires detail analysis of data, and the fuzzy rule base has to be developed p. k. dash et al. heuristically for each season. the rules fixed in this way may not always yield the best forecast. thus, a hybrid neural network model is used that combines the idea of fuzzy logic-based decision system and neural network structure and learning abilities into an integrated frame- work. such a hybrid model provides human understandable meaning to the normal feedforward neural network in which the internal units are always opaque to the users. this structure also avoids the rule matching time of the inference engine in the traditional fuzzy logic system and results in enhanced learning speed and prediction accuracy. the present work is aimed at achieving the said objective of a robust load forecast with much improved accuracy using a fuzzy expert system modelled by the hybrid ann architecture. the two types of fuzzy expert system models, abbreviated fes, and fes, are based on the argument that any fuzzy expert system employing one block of rules may be approximated by a neural network (feedforward, multilayered). the input vector to fes, and fes, consists of differences in weather parameters between the reference and the forecasted day or hour. the output of the fes, and fes, gives the load correction that, when added to the past load, gives the forecasted load. the learning algorithm for fes, and fes, combines unsupervised and supervised learning procedures to build the rule nodes and train the membership functions. the supervised learning proce- dure for fes, uses a gradient back-propagation algorithm for finding the optimum weights and member- ship functions. although for fes,, the supervised learning procedure uses a linear kalman filter-based learning algorithm suggested by scaler and tepede- lenlioglu ( ), which is similar to the least square adaptive learning techniques. the least square adaptive filtering techniques are known to have rapid convergence properties over the back-propagation algorithm. a few examples of peak load forecasting and daily average load forecasting of a typical utility with h and h lead time are shown in this paper. the approach presented in this paper is also applicable to load forecasts with longer lead times. . fuzzy expert system . . fuzzy statements this section portrays inexact information presented by fuzzy statements, and explains both fuzzy conditional statements and inference mechanisms. typically, an engineering model requires the use of exact or mathe- matical statements. these statements correspond to precise information such as x= . , < < , ory= t+ . the value of x = . has a grade membership comparable to % (= l), for all other values ( . , . , . , . ), the grades of membership in the solution is zero. in the case of real world values, however, this grade of membership is not true due to the imprecision of tools, the influence of the observer, etc. additionally, human reasoning is often imprecise, that is, inexact statements such as, “the value of x is not big” or “the temperature is about ”.” therefore, a theory to correctly express the grade of membership is desirable. the fuzzy set theory allows manipulation of both exact and inexact (fuzzy) statements. this is very important for load forecasting because there are so many fuzzy factors that are difficult to characterize by a number. an instance of this could be weather conditions such as temperature, humidity, cloud cover, etc. for example, “the temperature this morning will be about ll’.” temperature is an important factor in load forecasting, but is not easy to characterise by an exact numerical quantity. in addition, the linguistic hedges (such as small, medium, large) can be modified by other linguistic hedges (such as not, very, very very) (zadeh, ) for example, “value of x is not big,” where big is a linguistic hedge and not a modifier. . . fuzzy rule base a fuzzy expert system consists of a collection of fuzzy if-then rules. for example, the rule base r is written as r = [r’, r=, . . ., r”] ( ) with r’:if (x, is f{ and . . and xp is fl) then (y, is g: and . . . y is gb) ( ) where x = (x, , . . ., x,jt and y= (yl, . . ., y,)r are the input and output vectors to the fuzzy system, respectively. fi and gi are labels of fuzzy sets vi and v,, respectively, and ll ( ) where y, e v,. hybrid neural nehvork for fuzzy expert system the final fuzzy set aor, (r,= [r,!, . . . , rf]) in v, determined by the fuzzy inference engine is obtained by combining eqn. ( ) for = , , . . ., m using the triangu- lar co-norm ~~dt,(y,) = pa-r,! (y,) + . . . + pa++ (yj>. ( ) . . . fuzzijication and defuzzification. in many cases it is convenient to express the membership function of a fuzzy subset in terms of a standard nonlinear function. the following gaussian membership function suggested by lin and lee ( ) is used for the fuzzification of input and output linguistic parameters of the fuzzy expert system: pa,(x) = exp -iv i i ( ) where a and h are the center and width of the gaussian membership function, respectively. the defuzzifier is needed to give a crisp output for any practical application like forecasting. the output of the fuzzy inference engine is a fuzzy set aqr in v, therefore, the defuzzifier maps aor into a crisp point y e v. the most commonly used centroid defuzzifier is defined as ( ) . fuzzy expert system model using neural network approach an alternative to the neural network-based load forecast is the expert system approach. a fuzzy expert system approach for load forecast consists of rules similar to the rules r’, . . ., rm given in section . . one of the difficulties with the fuzzy expert system is the rule matching and composition time, apart from the time- consuming process of adapting the rules. the neural network eliminates the rule matching process and stores the knowledge in the link weights. the decision signals can be pumped out immediately after the input data are fed in. figure shows the proposed hybrid neural network to model the fuzzy expert system using the ann architecture. the fuzzy expert system clusters the differential temperatures and humidities of the ith and (i + n)th days into fuzzy term sets. here n is the lead time for the forecast (i.e., n = for h ahead forecast, n = for h ahead forecast, and so on). the output of the expert system is the final load correction (elc). hence, the forecasted load on (i + n)th day (pai + n)) is given by: p,(i + n) = p(i) + ,,(i). ( ) the fuzzy expert system modelled as a hybrid neural network has a total of five layers. nodes at layer are the input linguistic nodes. layer is the output layer and consists of two nodes [one is for the actual load correction (&) and the other is the desired load correction (e&l. nodes at layer and are term nodes that act as membership functions to represent the term sets of the respective linguistic variable. each node at layer represents the preconditions of the rule nodes, and layer links define the consequence of the rules. the functions of each layer is described as follows: a> b) cl layer : the nodes in this layer just transmit the input feature x,, i = , to the next layer. layer : each input feature x,, i = , is expressed in terms of membership values &a,,, l~,~), where i corresponds to the input feature and j corresponds to the number of term sets for the linguistic variable x,. the membership function &, uses the gaussian membership function ( ). layer : the links in this layer are used to perform precondition matching of fuzzy logic rules. hence, the rule nodes perform the product operation (or and operation). prurp = n pj i, ( ) where r,= , ,. . ., n. r, corresponds to the rule node and n is the maximum number of rule nodes. however, if the fuzzy and operation is used d) layer : the nodes in this layer have two operations (i.e., forward and backward transmission). in forward transmission mode, the nodes perform the fuzzy or operation to integrate the fired rules, which have the same consequence: p =c ; ( ) ,=i where p corresponds to the links terminating at the node. in the backward transmission mode, the links function exactly the same as the layer nodes. e) layer : there are two nodes in this layer (i.e., for obtaining the actual and desired output load correc- tion, respectively). the desired output load correction (e,j is fed into the hybrid neural network during learning whereas the actual load correction (a,,) is obtained by using the centroid defuzzification method ( ). . hybrid learning algorithm for the fuzzy expert system (fes) the hybrid learning scheme consists of two phases. in phase i, the initial membership functions of the input and output linguistic variables are fixed and an initial form of the network is constructed. then, during the learning process, some nodes and links of this initial network are deleted or combined to form the final structure of the network. in phase ii learning, the input and output membership functions are optimally adjusted to obtain the desired outputs. phase i represents the unsupervised learning phase. phase ii represents the supervised learning phase. . . phase-i: unsupervised learning phase given the training input data x,(r), i = , , the desired output load correction (e,(r)) and the fuzzy partitions i&, . we want to locate the membership functions (i.e., aij and b,) and find the fuzzy logic rules. the kohonen’s feature maps algorithm (lin & lee, ) is used to find the values for ai, and b,: ai, clo~est(~ + l) = ui, doscar) + tit) [ (‘)i xi(r)-ui, closest u,(r + ) = uij(t) for uii f a,, c,oscst ( ) ( ) ( ) where rl(f) is the monotonically decreasing learning rate and t= it( (i.e., the number of term sets for the linguistic variable xi). this adaptive formulation runs independently for each input linguistic variable xi: the width, b, is determined heuristically at this stage (lin & lee, ) as follows: b,j = ia, jwui, closest r ( ) where r is an overlap parameter. after the parameters of the membership functions have been found, the weights in layer are obtained by using the competitive learning algorithm (rumelhart & zipses, ) as follows: w, = ll; r,-‘u+ ) = [r,-‘w-k,(o$(r) r,- ‘(w/j; where ( ) ( ) and ‘y, is a constant that is inversely proportional to the prediction error, e(r), at that layer. as a result,& remains close to when r, is already large and the weight estimates are sensitive to parameter variations. a lower bound fo and f, is used to prevent the forgetting factor from becoming excessively small, resulting in large estimate fluctuations in spite of small prediction errors. the strategy for employing this forgetting factor is to initially allow it to assume small values during learning such that the initial conditions are quickly forgotten. however, as the average prediction error decreases, it is important to cause to remain near one to avoid the winding up of the weight estimates. this can be simply implemented by setting y, smaller as the mean squared error decreases. the unsupervised learning phase of fes, is same as for fes,. the supervised learning phase of fes, is modified using the linear kalman filter equations instead of the back-propagation method used for fes,. there- fore, the update equations are modified as follows: a) the weight update equation for layer is: p(f) = - r(o +x;(t)r,-‘(t)x,(t) ( ) ( ) p. k. dash et al. where el w, is given by eqn. ( ) and k,(t) is given by eqn. ( ). b) the update equations for a,, and b,, at layer are: b;,(t + ) = b,,(t) + rl,k,(o i v ( . ) where a el a d,, is given by eqn. ( ) and k,(t) is given by eqn. ( ). c) the update equations for a,, and b,j at layer are: a,,@ + ) = a,,(t) + rll k,(t) -ae [ a ( ) u&t + ) = a,(t) + t k,(t) [ c ( ) v where d el i_l a,_ is given by eqn. ( ) and k,(t) is given by eqn. ( ). maximum temperature difference (after supervised learning) maximum temperature difference (after unsupervised learning) . . e : . c . e . - - - - - - (w maximum humidity difference (after unsupervised learning) maximum humidity difference (after supervised learning) - - (%i (%) peak load difference peak load difference (after unsupervised learning) (after supervised learning) .o . e m .> . c :: e . e . .o . e m . @ e . e . - - - - - - (mw (mw figure . learned membership functions for h ahead peak load forecasting, in january (winter) using fes,. hybrid neural network for fuzy expert system r n - ann ----- fes, . . . . . . . . . . fesz ‘!, r i i i l .."\ '; i, :\ .'. . . i number of iterations figure . comparison of the mean absolute percentage errors versus the iteration number for h ahead forecast in january (winter). where, se/ a,] is given by eqn. ( ) and k,(t) is given by eqn. ( ). similarly, b&r + ) = b,,(t)- ~ k,w~e ( ) where df is given by eqn ( ) and k,(t) is given by eqn. ( ). the convergence properties of the supervisory leam- ing scheme for fes, are found to be superior to fes, using the kalman filter equation in the weight and membership update equations, as shown in sections . and . . also, the accuracy of the forecast increases for a similar reason. r,-‘(t+ ) =$(t)r,-‘(t) +$(t)x,(t) and the forgetting factora is modified as f;(t+l)=l-yt?(t+l) ( ) ( ) table peak load forecasting in june (summer) using hour ahead forecast peak peak peak actual forecast forecast forecast load nj) ,k) (few &se,, (few day (mw (mn (mn ww &se,, : ; mape iterations required . - . . - . . . . . . . . - . . . . - . . - . . - . . - . . . . . . - . . - . . - . . - . . - . . . . - . . . . - . . . . . . . . . . . . - . . . . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . . - . - . - . . . - . - . - . - . - . - . . - . . . - . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . - . . . . - . - . . . - . . . - . . . . - . . - . - . - . . - . . . . . - . . p. k. dash et al. table peak load forecasting in january (winter) using hour ahead forecast peak peak peak actual forecast forecast forecast load u'jn) day (m'w (mw io mape iterations required . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . . . - . . . - . . - . . . - . . . . . . - . . . - . . . - . - . . - . . . - . . . . - . . - . . . - . - . . . - . . - . . . . . . . - . . . . . - . . . . . . . . . - . - . . - . . . . - . . . - . . . . . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . - . . - . . . . - . - . - . - . - . . - . - . . . . . . - . . . - . - . . . . . . . t?(t+ l)= kw +x’(qr,(t- )x(t) ( ) where y is a constant inversely proportional to the prediction error e(t). . implementation results in order to evaluate the performance of the fuzzy expert system, load forecasting is performed on a typical utility data. the models ann, fes,, and fes, are tested on a -year utility data for generating peak and average load profiles and some of the results are given in the subsequent subsections. in (bunn and farmer ( ) and brace ( )) it has been shown that ann gives the best prediction and accuracy compared to conventional approaches. so in this case the results of fes, and fes, are compared to that of the ann approach. the training sets are formed separately for each of the seven day types (i.e., tuesdays through thursdays, mondays, fridays, saturdays, sundays, holidays). the selection of training patterns and the selection of variable ranges are given in sections . and . . . . optimum selection of training patterns the utility data studied here are susceptible to large and sudden changes in weather and load, so selection of appropriate training cases plays a vital role in training the network. several techniques for the selection of training patterns have been suggested in peng, hubele, and karady ( ). the following load model is used for peak load forecasting: p,,(i) =f(p,,,(i--n), p,,,(i-n-l), . . ., p,,(i-n-n,), o,,,(i--n), o,,,(i), o,,,(i- l), . . ., omaxg-dr h,,,(i-n), h,,,(i), h&i-l), hybrid neural network for fuzzy expert system table peak load forecasting in january (winter) using hour ahead forecast peak peak peak actual forecast forecast forecast load (nv day (mw (my . . . . . . . . . io . . . . . . . . . . . . . . . . . . . . . . mape iterations required . . . - . - . . - . . . . - . - . . . - . . - . - . . . - . . . - . - . - . . - . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . - . . . . . . . - . . . . . - . - . - . . - . . . - . - . - . . - . - . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . . . . . - . . . . . . - . - . - . . . - . . . . . . . - . - . - . . - . - . . . . . ., h&i- )). ( ) for average load forecast a similar model as eqn. ( ) is chosen, given by: p,,(i) =kp,,(i-n), p,,(i- l), . . ., pji-n,), %di-n), o,,,w~ o,,,g- i), . . ., omax(i-ut omji-nh o,i,(& o,,,(i- , . . ., o,,,u-n,l %,,(i-n), h,,,(i), ff,,a-i), . . ., ~,,,(i-~j). h,,,(i-fi), hmili(i) h,i,(i- i), . . ., h,,,,(i- . ( ) here, n indicates the lead time of the forecast (n * n,, n , n?). p, , and h stand for load, temperature, and humidity, respectively. also different values of n,, n,, and n were tested to find the effect of the past data on the load forecast. with n,, n,, n,>o there was no marked improvement in the results. therefore, we chose, n = n = n = in this study. however, the choice of n,, n , n depends entirely on the utility data concerned. . . scaling of the variable range because the input variables and the estimated ones from the hybrid neural network have wide variations in magnitudes, they will cause convergence problem and the system will behave completely erratic. to circumvent this problem, the variables are scaled between . and . (rahman, drezga, & rajagopalan, ). this is performed so as to maximise accuracy and minimise training time. the following scheme is adopted to scale the variables between . and . . let x,,,,, and x, m,n denote the upper and lower bounds of the observed range of feature x, in all the patterns in the historical data base considering numerical values only. similarly, o,,,,, and ,, m,n denote the upper and lower bounds of the observed range of outputs. then x, is normalised as x: = . + ww-x. ,,,mx.,,,-x,. n,,n ii ( ) where (i, j) corresponds to the jth pattern of the ith training set. p. k. dash et al. table average load forecasting in january (winter) using hour ahead forecast peak peak peak actual forecast forecast forecast load wj) day wv (m'w (%) 'fg (fpe& 'kz &i,, : mape iterations required . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . - . . . . - . . . . - . . . . - . - . . - . . - . - . - . . . . . . - . - . - . . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . - . . - . . . . . - . . . . - . . . - . . - . . . - . - . . . . . . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . . . . - . . . . . - . - . . . - . . - . - . . - . . - . . . . - . . . . similarly, the output , can be expressed ~= ~ +[ ~ ( ~~ ~,~~)/( ~,~~~~ ~,~~~)] ( ) where x and denote the normalised input and output vectors, respectively. the normalised predicted values can be converted back to the actual values using the above expressions. the ideas discussed in sections . . and . are used for obtaining the peak and average load forecasts. . . peak load forecasting for n h ahead peak load forecasting, the following training data are used for the ann: input pattern: p,,,(i), o,,(i), h,,,(i), o’,,,(i + n), h/,,( i + n) output pattern: p ,,,(i + n) for the ann. superscript f denotes the forecasted values for @ and h. the forecasted values for (i + n)th day are used to get a more realistic forecast. for fes, and fes, the training patterns used are: input pattern: a@,,,(i, i + n) and ah_(i, i + n) output pattern: erc(i), the desired load correction. the p,,(i + n) for fes, and fes, is obtained using eqn. ( ). table gives the learned membership functions for h ahed peak load forecasting in winter using fes,, for example, rule is interpreted as: ro: if a@,,, is term and ah_ is term then e^, is term . figure gives the learned membership functions after the first phase (unsupervised learning phase) and the hybrid neural network for fuzzy expert system table average load forecast in june (summer) using hour ahead forecast -- peak peak peak actual forecast forecast forecast load nv day (mw) ww a la i mape iterations required . - . . - . . . i . - . . . . - . . . . - . . - . . - . . . . . . . . . . - . . - . . . . . . . . . i . - . . - . . . . . . . . - . . i .ai . . . - . . . . . . . . . . . . . i . . . . . . . . . . . . . . . . . . i . . . . . . - . . - . . - . - . - . . . - . . - . - . . . . . - . - . . . . . - . . - . . . . . . . . . . . i . . . . . i . . . i . . . . i . . . . . . . i . i . . . - . . . . . - . - . - . i . a - . . - . . - . . . . . - . - . . . . . - . . . . . i .oa second phase (supervised learning phase). figure gives the mean absolute percentage errors (mape’s) versus the number of iterations. the results in figures and are obtained for h ahead load forecasting in january (winter). from figure we find that the fes, gives fastest convergence followed by ffs, and ann. the convergence speed of the fes, was found to be superior because of the linear kalman filter equations used for weight update and the error-dependent forgetting factor was responsible for driving the mape low during the first few hundred iterations until the bias introduced by the initial conditions was eliminated. table gives the peak load forecasting results in terms of mean absolute percentage errors (mapes) for ann, fes, and fes, in the month of june (winter) using h ahead forecast. tables and give the results for the month of january (winter) using h and h ahead forecast, respectively. from tables , and we see that the fes, and fes, give better prediction accuracy compared to ann. also we find that the results for h and h predictions are comparable with that of the h ahead predictions. this is because the load forecasting was performed as a one- step process (i.e. looking h ahead, h ahead, and so on). however, as the lead time increased to h the pes were found to be greater than % even with fes,. as the primary aim of this paper was to make a comparative assessment between ann, fes, and fes,, no attempt was made to improve the forecast errors further. . . daily average load forecast for n h ahead average load forecasting, the following training data are used for ann. input pattern p,,(i), o,,(i), o,,,(i), h,,,(i), &&). @l(i + n), o’,,,(i + n). hf,,,(i + n), hti,(i + n) output pattern: p,,(i + n) for ann. for fes, and fes,, the training patterns used are: r k. dash et al. table average load forecasting in june (summer) using hour ahead forecast peak peak actual forecast forecast load vw day (mw) (mw (k, 'kz cfpe:,, z ; : z; z: mape iterations required . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . - . . - . . . . - . - . - . - . . . . - . - . . . . . . - . . . . - . . . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . - . . - . . . . - . - . - . - . . . . - . - . . - . . . . - . . . . - . . . - . . . peak forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . - . . - . . - . . - . - . . - . . - . - . - . - . . - . . - . - . . . . . - . . . . . . input pattern: a@,,,,(& i + n), aqmin(ir i + n), ah&i, i + n), ahmi,(i, i + n) output pattern: e,,(i), the desired load correction. again forecasted temperature and humidity values are used for the day of the forecast. the p,,(i+n) for fes, and fes, is obtained using eqn. ( ). table gives the average load forecasting results, number of iterations for convergence, pes and the mapes for the ann, fes,, and fes, models in the month of january (winter) using h ahead forecast. tables and give the results for h and h ahead predictions in the month of june (summer). again, from tables , , and we find the improved performance of fes, in terms of faster convergence and improved overall accuracy over the ann and fes,. . discussion the fuzzy expert system presented in this paper is constructed from training examples by machine learning techniques, and the neural network model is trained to develop fuzzy logic rules and find input/output member- ship functions. by combining both unsupervised and supervised learning schemes, the learning speed for time series forecasting problems converges much faster than original back-propagation learning algorithm. further, by using a linear kalman filter for the supervised learning phase, the learning time is shortened with rapid con- vergence due to the adaptive nature of the learning algorithm. the fuzzy expert system using a linear kalman filter produces a more accurate load forecast for lead times varying from h to h in comparison to the one using gradient descent back-propagation algo- rithm. the forecasting results presented in section are based upon only true temperature information because all the data for this utility are historical. however, in reality the temperature information will be the forecasted value and this will add to the forecast error. the developed expert systems are used to forecast the load during weekends but no attempt is made to forecast for days with unusual events. such events will require hybrid neural network for fuzzy expert system extensive data analysis to track the events in that day and to select training cases accordingly from the previous days with similar events. such activities are continuing, and will be reported in a future paper. as a final note, a shorter period is used for training as load patterns change very fast. but training and predic- tion can be done over longer periods with less chaotic data having strong correlation to any particular weather and environmental variable. this flexibility permits the adaptation of the proposed fuzzy expert systems for producing accurate load forecasts for different geo- graphic areas. . conclusions this paper presents the development of a fuzzy expert system using a hybrid neural network approach to predict peak and daily average load profiles in an energy management system. the fuzzy expert system is mod- elled as a hybrid neural network and uses fuzzy membership values of load and weather parameters. a hybrid learning scheme consisting of both self-organized and supervised learning phases is used for training the hybrid neural network. further, the paper presents simulation results using both back-propagation and kalman filter-based algorithms, the latter producing faster convergence during training and more accurate predictions. acknowledgement-the authors gratefully acknowledge the funds from the national science foundation (nsf grant no. int- and int- ). usa for undertaking this research. references box, g. e. p., & jenkins, g. m. ( ). time series onolysisforecosring and control, oakland, ca: holden-day. brace, m. c. ( ). a comparison of the forecasting accuracy of neural networks with other established techniques, firsr inrer- national forum on applications on neural networks to power systems, seattle, wa, july, , pp. - . buckley, j. j., hayashi, y., & czogala, e. ( ) on the equivalence of neural nets and fuzzy expert systems, fuzzy sets systems, - . bunn. d. w.. & farmer, e. d. ( ). comparative models for electric loadforecasting, new york: john wiley and sons. dash, p. k., dash, s., rama krishna, g., & rahman, s. ( ). forecasting of a load time series using a fuzzy expert system and fuzzy neural networks, engineering international systems, l( ), - . dash, p. k., satpathy, j. k., rama krishna, g., & rahman. s. ( ). an improved kalman filter based neural network approach for short-term load forecasting, proceedings of the third international symp. on electricity distribution and energy management, vol. , pp. - . el-sharkawi, m. a., oh, s., marks, r. j., dambourg. m. j., & brace, c. m. ( ). short term electric load forecasting using an adaptively trained layered perceptron, first international forum on applications of neural networks to power systems, seattle wa., july - , , pp. - . hayashi, y., buckley, j. j., & czogala, e. ( ). fuzzy expert systems versus neural networks, proceedings of international joint con- ference on neural networks, baltimore, vol. ii, june - pp. - . haykin, s. s. ( ). adaptive filter theory (pp. - ). englewood cliffs, nj: prentice-hall. ho, k. l., hsu, y. y., &yang, c. c. ( ). short term load forecasting using a multilayer neural network with an adaptive learning algorithm, ieee transactions on power systems, ( ). -i . lee, c. c. ( ). fuzzy logic in control systems: fuzzy logic controller, part . ieee transactions systems, man & cybernetics, ( ), - . lee, k. y., cha, y. t., & park, j. h. ( ). short term load forecasting using an artificial neural network, ieee transactions on power systems, (i), - . lin, c. t., & lee, g. c. s. ( ). neural network-based fuzzy logic control and decision system, ieee transactions on computers, ( ), - . moghram, i., & rahman. s, ( ). analysis and evaluation of five short-term load forecasting techniques, leee transactions on power systems, ( ), - . pal, s. k., & mitra, s. ( ). multilayer perceptrons, fuzzy sets and classifications, ieee transactions on neural networks, ( ), - . park, d. c., el-sharkawi, m. a., & marks, r. j., ( ). adaptively trained neural networks, ieee transactions on neural networks, - . park, j. h., park, y. m.. & lee, k. y. ( ). composite models for adaptive short term load forecasting, leee transactions on power systems, l( ), - . peng, t. m., hubele, n. f., & karady, g. g. ( ). advancement in the application of neural networks for short term load forecasting, ieee transactions on power systems, (l), - . rahman, s., drezga, i., & rajagopalan, j. ( ). knowledge enhanced connectionist models for short-term electric load fore- casting, international conference on ann applications to ponaer systems, yokohama, japan, april. rumelhart, d. e., & zipses, d. ( ). feature discovery by competitive learning. cognitive science, , -l . scalero. r. s., & tepedelenlioglu, n. ( ). a fast algorithm for training feedforward neural networks, ieee transactions on signal processing, ( ). - . singhal, s., & wu, l. ( ). training feed-forward network with the extended kalman algorithm, proceedings of icassp, pp. - . singhal, s., & wu, l. ( ). training feed-forward network with the extended kalman algorithm. proceedings of icassp, pp. - . wang, l.-x. & mendel, j. m. ( ). generating fuzzy rules by learning from examples, ieee transacrions on systems man & cybernetics, ( ), - . zadeh. l. a. ( ). fuzzy sets, information and control. , - . aalborg universitet when to test? troubleshooting with postponed system test ottosen, thorsten jørgen; jensen, finn v. publication date: document version early version, also known as pre-print link to publication from aalborg university citation for published version (apa): ottosen, t. j., & jensen, f. v. ( ). when to test? troubleshooting with postponed system test. department of computer science, aalborg university. general rights copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. ? users may download and print one copy of any publication from the public portal for the purpose of private study or research. ? you may not further distribute the material or use it for any profit-making activity or commercial gain ? you may freely distribute the url identifying the publication in the public portal ? take down policy if you believe that this document breaches copyright please contact us at vbn@aub.aau.dk providing details, and we will remove access to the work immediately and investigate your claim. downloaded from vbn.aau.dk on: april , https://vbn.aau.dk/en/publications/f ecc - fd- -a d - b f f when to test? troubleshooting with postponed system test. thorsten j. ottosen and finn v. jensen technical report - department of computer science aalborg university, denmark october , abstract. troubleshooting is about computing a cost-efficient way to repair a faulty device. in this paper we investigate the most simple action-based trou- bleshooting scenario with a single extension: the overall system test is not required to be performed after each action, but it may be postponed until several actions have been performed. as shown by kadane and simon, the simple troubleshooting scenario is easily solvable in polynomial time. however, we conjecture that the new troubleshooting scenario is np-hard and therefore describe an Θ(n ) time heuristic. the new heuristic guar- antees optimality for a class of models, and for other classes of models we benchmark it against the optimal solution and other simpler greedy heuristics. the benchmark is performed on artificial models as well as a real-world troubleshooting model, and they suggest that the heuristics are close to optimal. introduction imagine that you have a device which has been running well up to now, but suddenly it is malfunctioning. a set of faults f describes the possible causes to the problem. to fix the problem you have a set a of actions, which may fix the problem by repairing one or more faults. an action, α, has a positive cost cα and it has two possible outcomes, namely "α = yes" (the problem was fixed) and "α = no" (the action failed to fix the problem). informally, our task is to determine the best sequence of actions until the problem is fixed or until no more actions remains to be carried out. for an overview of state of the art of decision theoretic troubleshoot- ing we refer the reader to (jensen et al., ), (langseth and jensen, ), (skaanning and vomlel, ), (langseth and jensen, ), (koca and bil- giç, a), (koca and bilgiç, b), and (warnquist et al., ). basically all non-trivial troubleshooting domains are np-hard (vomlelová, ). the scenario investigated in this article is similar to troubleshooting with con- ditional costs, but the proof for np-hardness does not carry through in this simple scenario. nevertheless we conjucture that our scenario is np-hard. in traditional troubleshooting it has been assumed that each action is followed by a test that determines whether the entire system d is working. furthermore, it has generally been assumed that the cost of this test, cd, is so low that it should not even be modelled. we show that this assumption is incorrect in general and develop terminology and algorithms for this new scenario. in this paper we shall examine troubleshooting problems where the fol- lowing model parameters are given: . an overall problem d that we need to solve (e.g. "my printer does not work"). . a set of faults f = {f , . . . , fm} describing the possible causes to the problem (e.g. "toner cartridge empty"). . for each fault f ∈f, a probability p(f) describing how likely it is that f is present when troubleshooting begins. . a set of actions a = {α , . . . ,αn} that can potentially solve d (e.g. "change the toner cartridge"). . for each action α ∈a, a cost cα describing the resources required to perform the action. . for each action α ∈ a, a success probability p(α = yes|f) describ- ing the likelihood of solving d by performing the action when f is present. . a cost of testing if the problem d remains, cd > . furthermore, we have the following simplifying assumptions about the model: assumption . the single fault assumption: initially the problem d is known to exist and it is due to the presence of a single fault from f. assumption . the action result assumption: repeating a failed action will not fix the problem. assumption . the carefulness assumption: by performing an action or testing the system, we never introduce new faults. assumption . the independent actions assumption: each action repairs a dis- tinct set of faults that may be causing the problem d. we uphold these assumptions throughout this paper. remark. assumption implies∑ f∈f p(f) = and by assumption the repair probability of an action can be computed as p(α = yes) = ∑ f∈f p(α = yes|f) · p(f) . however, since actions may be imperfect (p(α = yes|f) < ), assumption does not imply ∑ α∈a p(α = yes) = . furthermore, the above equation can also be violated by the fact that not all faults may be addressed by an action. before we can define our optimization problem formally, we need some notation and terminology. we write εi to denote that the first i actions have failed (ε ⊆ εi), and we have by assumption p ( ε ) = because the device is faulty. we call p(α = yes) for the repair probability of α, and we shall abbreviate it with pα. to make it clear that the cost of an action includes the cost of testing the system cd, we write ca+d for ca + cd and so ca never includes the cost of testing the system. for each v ⊆a we can form a compound action a with the following properties pa = ∑ α∈v pα, ca = ∑ α∈v cα, |a| = |v | . to make the distinction clear, we call actions α ∈ a for atomic actions. we can see that a partition of a into k subsets {v , . . . , vk} induces a new model with independent actions {a , . . . , ak}. we shall also say that {a , . . . , ak} is a partition of a whenever the corresponding sets of atomic actions form a partition of a. from now on we shall not distinguish be- tween the set of actions and the compound action; that is, we use a to mean a set of atomic actions and write a = {α , . . . ,α|a|}. a troubleshooting se- quence is a sequence of compound actions s = 〈a , . . . , a`〉 prescribing the process of repeatedly performing the next action until the problem is fixed or the last action has been performed and such that {a ,. . . ,a`} is a partition of a. when each compound action contains only one action, we say that the sequence is atomic. a subsequence of s, s[k,m] = 〈ak+ , . . . , am〉 with k ≤ m, is called a partial troubleshooting sequence if k ≥ or m < `. if k = , we may simply write s[m]. definition . let s = 〈a , . . . , a`〉 be a troubleshooting sequence. then the expected cost of repair (ecr) of s is the mean of the cost until an action succeeds or all actions have been performed: ecr (s) = ∑̀ i= cai+d · p ( εi− ) . formally, the problem is then to find a troubleshooting sequence with min- imal ecr (·) over all possible sequences. we call such a sequence for opti- mal. an important definition is that of efficiency: definition . the efficiency of an action a is ef(a) = pa ca+d . its importance stems from the following results. theorem (kadane and simon ( )). let s = 〈α , . . . ,αn〉 be an atomic troubleshooting sequences, and let cd = . then s is optimal if and only if ef(αi) ≥ ef(αi+ ) for i ∈{ , . . . ,n− } . theorem (jensen et al. ( )). let s = 〈α , . . . ,αn〉 be an atomic trou- bleshooting sequence. then a necessary condition for s to be optimal, even when cd > , is that ef(αi) ≥ ef(αi+ ) for i ∈{ , . . . ,n− } . corollary . theorem holds for arbitrary troubleshooting sequences. proof. because a troubleshooting sequence partitions a, the compound ac- tions can be seen as virtual actions which are also independent, thus satis- fying assumption . all other assumptions cannot be invalidated. we may also compute the ecr of any atomic troubleshooting sequence: theorem (jensen et al. ( )). let s = 〈α , . . . ,αn〉 be an atomic trou- bleshooting sequence. then the ecr of s may be computed as ecr (s) = n∑ i= cαi+d ·   − i− ∑ j= pαj   , where − ∑i− j= pαj = p ( εi− ) . corollary . theorem holds for arbitrary troubleshooting sequences. proof. same argument as for corollary . thus, due to assumption - , we may completely ignore the f, p(f), and p(α = yes|f) once the repair probabilities have been computed. therefore we only use pα in the rest of this paper. the following example shows that theorem does not extend to arbi- trary troubleshooting sequences when cd > . example . consider a model with four atomic actions α , α , α and α . the other properties of the model are as follows: pα cα cα+d ef(α) α . . α . . α . . α . . so cd = . we have ecr (〈α ,α ,α ,α 〉) = + . · + . · + . · = . ecr (〈{α ,α },α ,α 〉) = + . · + . · = . thus proving that it is more efficient to form the compound action a = {α ,α }. to be able to test heuristic algorithms it is necessary with an algorithm that is guaranteed to find an optimal value. exhaustive searches are al- ways able to do this, albeit the algorithm might take exponential or super- exponential time. algorithm shows an exhaustive search with pruning (line , applying theorem ) and memoization (line - ). algorithm exhaustive search with compound actions : function generateallsequences(a, &s) : let all = ∅ : for i = to |a| do : for all (|a| i ) ways to construct the first compound action a do : if a ∈ s then : compute a from scratch : set s = s∪{a} : end if : let after = generateallsequences(a\ a,s) : for s = 〈a , . . . , a`〉 ∈ after do : if ef(a ) ≤ ef(a) then : let s′ = 〈a, a , . . . , a`〉 : set all = all∪{s′} : end if : end for : end for : end for : return all : end function : function optimalcompoundactionsequence(a) : let s = ∅ : let all = generateallsequences(a,s) : return arg min s∈all ecr (s) : end function let us briefly discuss the complexity of this algorithm. the algorithm re- cursively generates all possible sequences of compound actions. the num- ber of such sequences is n! · n− because for each of the n! permutations of the n atomic actions we can form n− different partitions. thus the run- ning time and memory requirement of the algorithm is o(n! · n). note, however, that this is not a tight upper bound since a troubleshooting se- quence s = 〈a , . . . , a`〉 can be constructed by ∏` i= |ai|! different order- ings. later we show that the complexity of an exhaustive search can be reduced to Θ(n · n!). two simple greedy heuristics in this section we shall develop two simple greedy algorithms. further- more, we also give examples showing that none of the algorithms are guar- anteed to find an optimal troubleshooting sequence. for the first algorithm we consider how the ecr of a troubleshooting sequence can be improved by merging two neighbouring actions into one compound action. we have the following result. theorem . let the two troubleshooting sequences s and s′ be given as s = 〈. . . , ax, ax+ , . . .〉 s′ = 〈. . . , a, . . .〉 with a = {ax, ax+ } and everything else being the same. then s′ has a strictly lower ecr than s if and only if cd > cax+ · pαx − ∑x j= pαj . proof. when the sequence with the new compound action is better, we must have ecr ( s′ ) − ecr (s) ≤ . we observe that most of the terms cancel each other out and we are left with ca+d · p ( εx− ) − [ cax+d · p ( εx− ) + cax+ +d · p(ε x) ] < m ca+d < cax + cd + (cax+ + cd) · p(εx) p(εx− ) m cax+ · ( − p(εx) p(εx− ) ) < cd · p(εx) p(εx− ) m cd > cax+ · ( p ( εx− ) p(εx) − ) m cd > cax+ · ( pαx − ∑x j= pαj ) theorem immediately suggests a procedure where we greedily merge ad- jacent actions until the ecr is no longer improved. we start with an atomic troubleshooting sequence and form compound actions by repeated appli- cation of theorem . justified by theorem , there are two obvious choices of the initial sorting: with respect to descending efficiency or descending p c -value. we have summarized this procedure in algorithm . (the special function st arg(·) returns the first index for which the boolean argument returns true.) as specified, the algorithm runs in Θ(n ) time. however, if the sum in line in compoundaction(·) are precomputed , the total cost of all calls to compoundaction(·) can be made linear. the running time of the algorithm is then dominated by the initial sorting of the actions leading to an Θ(n · lg n) worst case running time. is algorithm guaranteed to find an optimal troubleshooting sequence? the example below show it is not. example . we consider the following model: pα cα cα+d ef(α) pα cα α . . . α . . . α . . . so cd = . using the ef(·) order algorithm first computes < · . − . ≈ . . algorithm computing sequence with compound actions greedily : function compoundaction(〈α , . . . ,αn〉 ,x) : if x = n then : return {αn} : end if : let m = n st arg i=x ( cd ≤ cαi+ · pαi − ∑i j= pαj ) : return {αx, . . . ,αm} : end function : function greedycompoundactionsequence(a) : let s′ = 〈α , . . . ,αn〉 such that ef(αi) ≥ ef(αi+ ) or pαi cαi ≥ pαi+ cαi+ . : let s = 〈〉 : let i = : while i ≤ n do : let a = compoundaction(s′, i) : set s = s + a : set i = i + |a| : end while : return s : end function since this is false, it continues with the second test < · . − . − . = . which means that the algorithm suggest the sequence consisting of a single compound action {α ,α ,α }. if we repeat the calculations using the pc - order, we get the same result. the ecr is easily seen to be , but ecr (〈α ,{α ,α }〉) = + . · = . so the algorithm is not optimal. let us now consider another greedy heuristic. the algorithm is based on the idea that we should repeatedly form the most efficient compound action until all atomic actions have been assigned a compound action. to do this, we need a simple way of computing the most efficient compound action from a set of atomic actions. we first need a lemma. lemma . let a,b,c, and d be strictly positive numbers. then a + b c + d ≥ a c if and only if b d ≥ a c proof. a + b c + d ≥ a c ⇔ a · c + b · c ≥ a · c + a ·d ⇔ b d ≥ a c theorem . let v ⊆ a be given. let a be a compound action consisting of ac- tions in v with ef(a) maximal and with |a| minimal over those actions with max- imal efficiency. then a can be found by the following procedure: define ef(a) = when a = ∅; then repeatedly add an action α ∈ v to a with the highest pα cα value until ef(a) does not increase. proof. let a consist of actions in v such that ef(a) is maximized and |a| is minimal. then ef(a) equals∑ α∈a pα cd + ∑ α∈a cα = ∑ α∈a\{β} pα + pβ cd + ∑ α∈a\{β} cα + cβ = sp + pβ sc + cβ = p c where β is chosen arbitrarily. let furthermore α ∈ v \ a. we shall prove pβ cβ ≥ p c ≥ pα cα which implies the theorem. we first prove pβ cβ ≥ p c (i). because ef(a) is maximal and a has minimal size, lemma gives sp + pβ sc + cβ > sp sc ⇔ pβ cβ > sp sc (ii). then assume equation (i) is false; then by lemma sp + pβ sc + cβ ≥ pβ cβ ⇔ sp sc ≥ pβ cβ which is a contradiction with equation (ii). we then consider the second inequality. since ef(a) is maximized, lemma implies p c ≥ p + pα c + cα ⇔ p c ≥ pα cα which completes the proof. remark. in (langseth and jensen, ) a similar heuristic procedure is reached albeit for a slightly different problem. algorithm greedy algorithm with max-efficient actions : function greedycompoundactionsequence (a) : let s′ = 〈α , . . . ,αn〉 such that pαi cαi ≥ pαi+ cαi+ : let s = 〈〉 : repeatedly add the most efficient compound action to s : return s : end function algorithm shows a greedy algorithm based on theorem . like al- gorithm we can precompute the sums used in the efficiency and get a running time of Θ(n · lg n). the following example shows that algorithm is not guaranteed to find an optimal troubleshooting sequence. example . consider the following model pα cα cα+d pα cα α . . α . . α . . where the ordering considered is 〈α ,α ,α 〉. algorithm computes the following values to determine the first compound action: ef({α }) = . = . ef({α ,α }) = . + . = . ef({α ,α ,α }) = = . thus the suggested sequence is one single compound action with ecr . however, ecr (〈{α ,α },α 〉) = + . · = . which shows the algorithm is not optimal. having established the heuristic nature of both greedy algorithms, we would like to know if they at least find a local optimum. we use the following def- inition to define a local optimum. definition . let s = 〈α , . . . ,αn〉 be a sequence of atomic actions, and let sc = 〈b , . . . , bm〉 be a compound action sequence. then sc is said to be consistent with s if ∀ k < l αi ∈ bk and αj ∈ bl =⇒ i < j. in other words, sc can be seen as an ordered partition of s. if an atomic action sequence s is given, we call the set of all compound action sequences consistent with s for c(s). a local optimum is then a com- pound action sequence sc ∈ c(s) with minimal ecr. we then say the (or- dered) partition induced by sc is optimal. from example and it follows that neither algorithm nor algorithm guarantees an optimal partition. an algorithm for optimal partitioning we have now seen how the greedy algorithms fail to provide optimal se- quences even under the restriction that the solution must be consistent with a single seed-sequence s of atomic actions. however, since there are n− ways to partition the actions of s, a brute force approach is quite impracti- cal even for models of moderate size. by exploiting theorem below we can construct an o(n ) dynamic programming algorithm. lemma . let k ≥ and let sc = 〈a , . . . , a`〉 be a partial troubleshooting sequence that is performed after the first k actions have been performed. then the contribution of sc to the total ecr is given by ecrk(s c) = ∑̀ i= p ( εk+i− ) · cai+d where p ( εk+i− ) = − k∑ j= pαj − i− ∑ j= paj . proof. follows directly from corollary . the lemma ensures that the following definition is sound. definition . let s = 〈α , . . . ,αn〉 be an atomic troubleshooting sequence. for any partial sequence s[k,m] with k ≤ m the ecr of an optimal com- pound action sequence consistent with s[k,m] is defined by ecr∗(s[k,m]) = { min sc∈c(s[k,m]) ecrk(s c) if k < m if k = m . theorem . let s = 〈α , . . . ,αn〉 be an atomic troubleshooting sequence. then for k < m, ecr∗(s[k,m]) can be calculate as ecr∗(s[k,m]) = m min i=k+ ( ecrk(〈{αk+ , . . . ,αi}〉) + ecr∗(s[i,m]) ) proof. we use induction in m−k. basis step: m−k = . since there is only one compound sequence consistent with 〈αm〉, we get ecr∗(s[k,m]) = min(ecrk(〈{αm}〉) + ) = ecrk(〈{αm}〉) so the theorem holds in this case. induction step: assume the theorem holds for m−k ≤ x and let m−k = x + . then let s∗ = arg min sc∈c(s[k,m]) ecrk(s c) for which we must prove ecr∗(s∗) = m min i=k+ ( ecrk(〈{αk+ , . . . ,αi}〉) + ecr∗(s[i,m]) ) . by lemma it is correct to split the ecr into two terms corresponding to an ordered partition of s∗. furthermore, by the induction hypothesis we can compute ecr∗(s[i,m]) for any i since m−i ≤ x. then consider that s∗ must start with a compound action consisting of y ∈ { , , . . . ,x + } actions. these choices of the first action correspond exactly to the expressions that we minimize over, and the result follows. theorem shows that the problem of finding an optimal partition of a seed sequence s requires optimal partitions of smaller partial sequences (the optimal-substructure property) and that these smaller partial sequences are needed by the computation of several bigger sequences (the overlap- ping subproblems property). these two properties enables a dynamic pro- gramming algorithm, that is, an algorithm that only computes solutions to subproblems once by storing the results in a table. algorithm uses this approach to find the optimal partition of a seed-sequence s. the algorithm runs in Θ(n ) time and uses Θ(n ) space. the correctness of the algorithm follows from the above theorem. algorithm now also means that we can make a more efficient exhaustive search for the best troubleshooting se- quence. the algorithm runs in Θ(n · n!) time and is shown in algorithm . unfortunately, algorithm is not guaranteed to find an optimal trou- bleshooting sequence, as the following example shows. example . consider the model from example , but take cd = : pα cα cα+d ef(α) pα cα α . . . α . . . α . . . algorithm optimal partition into compound actions : procedure fillcell(row,col, s, cd, &ecrmatrix, &prevmatrix) require: row > col : let n = row − col : let a = 〈αcol, . . . ,αcol+n〉 from s : let ecr = p ( εcol− ) · ca : set ecrmatrix[row,col] = ecr : set prevmatrix[row,col] = col : for prev = col + to row − do : let ecr′ = ecrmatrix[prev,col] + ecrmatrix[row,prev] : if ecr′ < ecr then : set ecrmatrix[row,col] = ecr′ : set prevmatrix[row,col] = prevmatrix[row,prev] : end if : end for : end procedure : function optimalpartition(s = 〈α , . . . ,αn〉, cd) : let n = |s| : let ecrmatrix = { } : let prevmatrix = { } : for row = to n do : for col = row − down to do : fillcell(row,col, s, cd,ecrmatrix,prevmatrix) : end for : end for : let s∗ = 〈〉 : let prev = n : repeat : let i = prevmatrix[prev, ] : let a = 〈αi+ , . . . ,αprev〉 : set s∗ = 〈a, s∗〉 : until prev < : assert(ecr (s∗) = ecrmatrix[n, ] ) : return s∗ : end function algorithm compound action sequence with optimal partition : function compoundactionsequence(a) : let s = 〈α , . . . ,αn〉 such that ef(αi) ≥ ef(αi+ ) or pαi cαi ≥ pαi+ cαi+ : let s = optimalpartition(s) : return s : end function algorithm optimized exhaustive search with compound actions : function optimalcompoundactionsequence (a) : let bestecr = ∞ : let bestseq = 〈〉 : for all n! permutations of a s do : let s′ = optimalpartitioning(s) : if ecr (s′) < bestecr then : set bestecr = ecr (s′) : set bestseq = s′ : end if : end for : return s′ : end function for the ef(·) ordering, the algorithm constructs the following tables—here visualized as one table with a pair of values (ecr∗(s[k,m]), i) where i− is the split that minimized the (partial) ecr:   ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( . , ) ( . , ) ( . , ) ( , )   looking at the bottom left entry, we get that the last action is {α } and then looking at the entry above it, that the first action is {α ,α } and 〈{α ,α },α 〉 has ecr . . even though the initial p c ordering is different, the algorithm returns the same troubleshooting sequence. however, ecr (〈{α ,α },α 〉) = + . · = . which shows that the algorithm does not lead to an optimal sequence. a heuristic for the general problem we have seen that by changing the ordering of the atomic actions, we could improve the ecr compared to keeping the ordering fixed. we shall there- fore consider the orthogonal task of optimizing the ordering given that the partitioning is fixed. definition . let sc = 〈a , . . . , a`〉 and sd = 〈b , . . . , b`〉 be two compound action sequences of the same size. then sc and sd are partition equivalent if |ai| = |bi| ∀i ∈{ , , . . . ,`}. lemma . let sc = 〈. . . , ax, . . . , ay, . . .〉 and sd = 〈. . . , bx, . . . , by, . . .〉 be two partition equivalent compound action sequences. furthermore, let the only difference between sc and sd be that two actions α ∈ ax and β ∈ ay have been swapped s.t. β ∈ bx and α ∈ by. then ecr (sc) − ecr(sd) = (cα − cβ) · [ p(β) − p(α) + y− ∑ i=x p(ai) ] + [p(β) − p(α)] · y∑ i=x+ cai+d . proof. let sc and sd be given as above. we then have ecr(sc) = ∑̀ i= cai+d · ( − i− ∑ j= p(aj)) = ∑̀ i= cai ·ai ecr(sd) = ∑̀ i= cbi+d · ( − i− ∑ j= p(bj)) = ∑̀ i= cbi · bi . the difference between these two numbers can be expressed as ecr(sc) − ecr(sd) = ∆x + ∆i + ∆y (i) where • ∆x is the difference caused by changes to cax, • ∆i is the difference caused by ai changing into bi for all compound actions between ax and ay, and • ∆y is the difference causes by changes to cay and ay. for i ∈{ , . . . ,x− ,y + , . . . ,`} the terms cancel out. we have ∆x = ax · (cax+d − cbx+d) = ax · (cα − cβ) . furthermore, since bi = ai + p(α) − p(β)∀ i ∈{x + , . . . ,y}: ∆i = y− ∑ i=x+ (ai − bi) · cai+d = [p(β) − p(α)] · y− ∑ i=x+ cai+d finally we have ∆y = ay · cay+d − by · cby+d = ay · cay+d − by · (cay+d − cβ + cα) = (ay − by) · cay+d − by · (cα − cβ) = [p(β) − p(α)] · cay+d − by · (cα − cβ) . now observe that by = ay − [p(β) − p(α)] = ( ax − y− ∑ i=x p(ai) ) − [p(β) − p(α)] which implies that ∆y = [p(β) − p(α)] · cay+d − [ ax − p(β) + p(α) − y− ∑ i=x p(ai) ] · (cα − cβ) . inserting these three results into equation (i) yields ecr(sc) − ecr(sd) = (cα − cβ) · [ p(β) − p(α) + y− ∑ i=x p(ai) ] + [p(β) − p(α)] · y∑ i=x+ cai+d as required. theorem . let sc, α and β be given as in lemma . then the ecr of sc can be improved by swapping α and β if and only if (cα − cβ) · [ p(β) − p(α) + y− ∑ i=x p(ai) ] > [p(α) − p(β)] · y∑ i=x+ cai+d proof. we should swap the two actions if ecr(sc) − ecr(sd) > m (cα − cβ) · [ p(β) − p(α) + y− ∑ i=x p(ai) ] + [p(β) − p(α)] · y∑ i=x+ cai+d > m (cα − cβ) · [ p(β) − p(α) + y− ∑ i=x p(ai) ] > [p(α) − p(β)] · y∑ i=x+ cai+d . it turns out that in certain cases we can avoid the full analysis of theorem . proposition . let sc, α and β be given as in lemma . if p(β) > p(α) and cα ≥ cβ, or p(β) = p(α) and cα > cβ then ecr(sc) can be improved by swapping α and β. proof. if p(β) > p(α) and cα ≥ cβ, then p(β)−p(α) < , but cα−cβ ≥ . in theorem the two other factors are strictly positive, so the left hand side becomes non-negative while the right hand side becomes negative making the inequality true. if p(β) = p(α) and cα > cβ, then the left hand side is strictly positive while the right hand side is zero which also makes the inequality true. based on the above results, we can easily search for a better partition equivalent troubleshooting sequence. the procedure is given in algorithm . arguably, after the algorithm has terminated, it could be that more ac- tions could be swapped to improve the ecr. however, we have been un- successful in proving the number of possible swaps. therefore we prefer an algorithm that terminates deterministically. furthermore, we tried to run the procedure until no more swaps could be made on our test models, and we found that it did not improve the ecr compared to algorithm . algorithm : procedure optimizepartition(& sc=〈a , . . . , an〉) : for x = to n do : for α ∈ ax do : for y = x + to n do : for β ∈ ay do : if (cα − cβ) · [ p(β) − p(α) + ∑y− i=x p(ai) ] > : [p(α) − p(β)] · ∑y i=x+ cai+d : then : swap(α,β) : end if : end for : end for : end for : end for : end procedure it should be clear that the algorithm runs in Θ(n ) where n is the number of atomic actions. the reason it cannot run in Θ(n ) time is that checking theorem takes linear time on average, albeit when proposition applies it only takes constant time. because any swapping invalidates the involved sums, it is not possible to precompute these as with the greedy approaches we saw earlier. we can also formulate the improved heuristic described in algorithm . algorithm compound action sequence with optimized partition : function compoundactionsequence(a) : let s = 〈α , . . . ,αn〉 such that ef(αi) ≥ ef(αi+ ) or pαi cαi ≥ pαi+ cαi+ : let s = optimalpartition(s) : optimizepartition(s) : return s : end function we have by means of a computer constructed examples that show that the algorithm is not guaranteed to find optimal troubleshooting sequences. however, for certain models we can prove that optimality is ensured. con- sider a model where the actions can be ordered into the sequence s = 〈α , . . . ,αn〉 such that pαi ≥ pαi+ and cαi ≤ cαi+ for i ∈{ , , . . . , n − }; we call such models for non-reordable models. theorem . let s = 〈α , . . . ,αn〉 be an atomic troubleshooting sequence in a non-reordable model such that ef(αi) ≥ ef(αi+ ). then there exists an optimal troubleshooting sequence sc consistent with s. proof. let sd = 〈b , . . . , b`〉 and sc = 〈a , . . . , a`〉 be partition equivalent compound troubleshooting sequences where sd is optimal and sc is consis- tent with s. now assume sd is not consistent with s. since sc = sd, we can find the index of the first pair of compound actions ax and bx where the two sequences differ. then at least one action β ∈ ax, but β ∈ bx and at least one action α ∈ bx, but α ∈ ax. furthermore, α must be found in ay with y > x. since sc is consistent with s and since the model is non- reordable pβ ≥ pα and cβ ≤ cα (i). we now consider two cases. case : at least one of the inequalities in (i) is strict. then if we apply theorem on sd we find that we can improve the ecr of sd by swapping α and β. however, this is a contradiction to sd being optimal. hence in this case any optimal sequence must be consistent with s. case : we have equality in equation (i). in that case we can safely swap α and β in sd without altering the ecr. if we now have sd = sc, then we are done. if not, we can repeat the procedure until sd = sc. since each swap puts at least one atomic action into its rightful compound action, the procedure terminates in a finite number of steps, thus proving that there exists an optimal sequence consistent with s. corollary . algorithm and find an optimal troubleshooting sequence in non- reorderable models. proof. both algorithms calls optimalpartition(·) after having sorted the actions with respect to descending efficiency. results in this section we shall investigate the performance of the heuristic algo- rithms. in particular, we investigate how the following two model param- eters affect the precision: . the closeness of the initial efficiency of actions. . the closeness of the cost of the actions. due to the difficulty of finding an optimal solution, we could not inves- tigate models with more than actions. in the models below the repair probabilities have not been normalized to make the specification accurate (but they are of course normalized at runtime) and the efficiency stated is the efficiency rounded to three decimals when cd = . model : close efficiencies + close costs α α α α α α α α pα . . . . . . . . cα . . . . . . . . ef(α) . . . . . . . . model : close efficiencies + non-close costs α α α α α α α α pα . . . . . . . . cα ef(α) . . . . . . . . model : non-close efficiencies + close costs α α α α α α α α pα . . . . . . . . cα . . . . . . . . ef(α) . . . . . . . . model : non-close efficiencies + non-close costs α α α α α α α α pα . . . . . . . . cα ef(α) . . . . . . . . for each of the four initial models we run all the algorithms for increas- ing values of cd. the following observation shows when it does not make sense to increase cd anymore. proposition . once cd is made so high that the optimal sequence consists of one compound action, then increasing cd can never change the optimal sequence. the increment in cd was determined as one permille of the largest cost of any action in the model. starting from cd = we kept sampling until the optimal sequence consists of only one compound action. the tables below summarize the result for the four models where each column describes the results of a given algorithm. the first four rows summarize the percent- wise deviation from the optimal ecr, and the last row shows the percent of cases where the algorithm was optimal. the number of increments to cd is given in the parenthesis after the model name. the first algorithm column is the simple ef(αi) ≥ ef(αi+ ) sorting. for algorithm , , and we give two numbers. the first corresponds to an initial ef(·) ordering and the second corresponds to an initial p c ordering. model : close efficiencies + close costs ( ) ef(αi) ord. alg. alg. alg. alg. min / / / max . . . / . . / . . / . mean . . . / . . / . . / . median . . . / . . / . / % opt. . . . / . . / . . / . model : close efficiencies + non-close costs ( ) ef(αi) ord. alg. alg. alg. alg. min / / / max . . . / . . / . . / . mean . . . / . . / . . / . median . . . / . . / . . / % opt. . . . / . . / . . / . model : non-close efficiencies + close costs ( ) ef(αi) ord. alg. alg. alg. alg. min / / / max . . . / . / / mean . . . / . / / median . / / / % opt. . . . / . / / model : non-close efficiencies + non-close costs ( ) ef(αi) ord. alg. alg. alg. alg. min / / / max . . . / . / / mean . . . / . / / median . / / / % opt. . . . / . / / we can summarize the results as follows: . algorithm is by far the best heuristic. algorithm is the worst of the new heuristics. algorithm performs surprisingly well, almost as good as algorithm and would thus be considered the preferred choice for a system under real-time constraints. . models with non-close efficiencies are much easier to solve than mod- els with close efficiencies. models with non-close costs seem to induce larger deviations from the optimal ecr than models with close costs. we consider none of these findings surprising. we have also tested the algorithms on a real-world model with actions. the results are summarized in the table below where the statistics are com- puted as the percent-wise deviation from the overall best algorithm (algo- rithm with p c ordering). the − indicates that a small percentage of the cases where better than algorithm . also notice that the apparent conflict between "median" and "% best" in column two is due to rounding of the median. model : actions ( ) ef(αi) ord. alg. alg. alg. alg. min / − / − max . . . / . . / . . mean . . . / . . / . median . . / . . / . % best . . . / . / . . the conclusion here is the same with one noticeable difference: algorithm is now the third best algorithm, only beaten by algorithm and with p c ordering and far better than any other heuristic. the fact that algorithm and are very close to each other also suggests that we are quite close to the optimal value. conclusion and further research we have extended a classical polynomial troubleshooting scenario with postponed system test. albeit this is a somewhat simple extension, we conjecture that the new scenario is np-hard. we have identified a class of non-reordable models where the best of the described heuristics are guar- anteed optimal. for reorable models, we found that the suggested heuris- tics provide quite close approximations to an optimal solution—even for heuristics with an o(n·lg n) complexity. apart from proving or disproving np-hardness, future research includes finding guarantees on the perfor- mance of the heuristics as well as guarantees about local optimality when optimizing a given partition. acknowledgements we would like to thank sven skyum for the discussions that lead to the- orem . we are also grateful to dezide aps for providing the real-world model used in the experiments. finally, we would like to thank our col- legues of the machine intelligence group at aalborg university for provid- ing such a fruitful research environment. references f. v. jensen, u. kjærulff, b. kristiansen, c. skaanning, j. vomlel, and m. vomlelová. the sacso methodology for troubleshooting complex sys- tems. artificial intelligence for engineering design, analysis and manufac- turing, : – , . j. kadane and h. simon. optimal strategies for a class of constrained se- quential problems. the annals of statistics, : – , . e. koca and t. bilgiç. troubleshooting with questions and conditional costs. proceedings of the th turkish symposium on artificial intelligence and arti- ficial neural networks, pages – , a. e. koca and t. bilgiç. a troubleshooting approach with dependent actions. in r. l. de mántaras and l. saitta, editors, ecai : th european con- ference on artificial intelligence, pages – . ios press, b. isbn - - - . h. langseth and f. v. jensen. decision theoretic troubleshooting of coherent systems . reliability engineering & system safety, pages – , . h. langseth and f. v. jensen. heuristics for two extensions of basic trou- bleshooting. in proceedings of the seventh scandinavian conference on arti- ficial intelligence, pages – . ios press, . isbn - - - . c. skaanning and j. vomlel. troubleshooting with simultaneous models. symbolic and quantitative approaches to reasoning with uncertainty, th, . m. vomlelová. complexity of decision-theoretic troubleshooting. int. j. intell. syst., ( ): – , . h. warnquist, m. nyberg, and p. säby. troubleshooting when action costs are dependent with application to a truck engine. in proceeding of the conference on tenth scandinavian conference on artificial intelligence, pages – , amsterdam, the netherlands, the netherlands, . ios press. isbn - - - - . ubipapago: context-aware path planning expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a ubipapago: context-aware path planning chiung-ying wang a, ren-hung hwang b,⇑, chuan-kang ting b a department of information management, transworld university, taiwan, roc b department of computer science and information engineering, national chung-cheng university, taiwan, roc a r t i c l e i n f o keywords: ubiquitous network human-centered computing context-awareness path planning genetic algorithm (ga) spatial conceptual map - /$ - see front matter � elsevier ltd. a doi: . /j.eswa. . . ⇑ corresponding author. address: , university taiwan, roc. tel.: + x ; fax: + e-mail addresses: wjy@cs.ccu.edu.tw (c.-y. wang) h. hwang), ckting@cs.ccu.edu.tw (c.-k. ting). a b s t r a c t the increased prevalence of digital devices with communication capability heralds the era of ubiquitous computing, as predicted by mark weiser. ubiquitous computing aims to provide users with intelligent human-centric context-aware services at anytime anywhere. optimal path planning in a ubiquitous net- work considers the needs of users and the surrounding context. this approach is very different from that applied by existing research on car navigation and mobile robots. this study proposes a context-aware path planning mechanism based on spatial conceptual map (scm) and genetic algorithm (ga), referred to as ubipapago. the scm model is adopted to represent the real map of the surrounding environment. the optimal path is planned using a ga, which is a robust metaheuristic algorithm. ubipapago attempts to automatically find the best path that satisfies the requirements of an individual user. a prototype of ubipapago is implemented to demonstrate its feasibility and scalability. experimental results validate the effectiveness and the efficiency of ubipapago in finding the optimal path. � elsevier ltd. all rights reserved. . introduction consider the path distance, accessibility of wifi access point (ap) the demand for intelligent navigation applications in path find- ing, such as papago (papago) and tomtom (tomtom interna- tional), has risen significantly in recent years, helped by the rapid development of wireless and communication technologies. these systems provide mobile users, such as car drivers, with intelligent path planning, routing and navigation services, and direct users to their destination using electronic maps and global position system (gps). however, these systems are computer-centric, and thus are unable to achieve the vision of mark weiser (weiser, ) of a ubiquitous environment with human-centric applications. con- text-awareness is a key to human-centric applications. a context- aware application extracts, interprets and uses context, and automatically adapts its functionality to the current context of an individual user. consequently, mobile users can adopt context to plan the ‘‘best-fitting” path automatically. to help address the problem and develop the solution, a sce- nario is described for finding john’s optimal path that satisfies his requirements. john needs to drive to the taipei arena as quickly as possible to watch a show, because he is late. the automobile assistant connects to a streaming server to watch the live show re- motely before john arrives at taipei arena. in this case, john needs to find the shortest path to the taipei arena, while keeping connec- tivity with the streaming server. therefore, the optimal path must ll rights reserved. rd., min-hsiung chia-yi , . , rhhwang@cs.ccu.edu.tw (r.- or g base station (bs), accessibility of the streaming service and quality of service (qos) of the required service. the problem con- sidered herein is formally defined as how to plan an optimal (shortest) path for an individual user based on his requirements and the context of his surrounding environment, such as user pref- erence, location, network connectivity, bandwidth, service avail- ability and quality of service (qos). the objective is to find an optimal (shortest) path given constraints, such as higher band- width, available service and qos guarantee. these constraints en- sure that the shortest path is also the one that best fits the user requirements. the context is very important for inferring the optimal path. the path planning framework relies on various contexts acquired from various sensing sources. the collected contexts are stored in an open and distributed context management server (lu, wang, & hwang, ), called ubiquitous gate (u-gate), which is responsi- ble for context acquisition, context representation, context retrie- val, context protection and context inference. the path planning framework adopts an ontology as the underlying context model. the ontology model is implemented using rdf (resource descrip- tion framework) ([rdf primer] and [rdf schema]) and owl (web ontology language) (owl). the context model includes user con- texts, such as personal information and active applications, as well as environmental contexts, such as map information and network connectivity. the proposed solution, called ubipapago, is based on spatial conceptual map (scm) (kettani & moulin, ) and genetic algo- rithm (ga) (holland, ). a modified scm is proposed to model the real world map and related environment context, and a path http://dx.doi.org/ . /j.eswa. . . mailto:wjy@cs.ccu.edu.tw mailto:rhhwang@cs.ccu.edu.tw mailto:ckting@cs.ccu.edu.tw http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa c.-y. wang et al. / expert systems with applications ( ) – planning map based on this model is then created. in addition, a ga is proposed to find the optimal path. chromosomes are encoded based on the method proposed in inagaki, haseyama, and kitajima ( ). the fitness function is carefully designed to fit the user requirements. prototyping and simulations are performed to dem- onstrate the feasibility and the efficiency of the proposed ubipapago. the remainder of this paper is structured as follows. section describes related works on path planning and context-aware path prediction. section introduces the scm model. section de- scribes the design of ubipapago based on ga. section provides implementation details of the prototyping system. experimental results are discussed in section . conclusions are finally drawn in section , along with our future research. . related work path planning problems have been studied extensively in the past, but mostly focusing on mobile robots (bodhale, afzulpurkar, & thanh, ; castilho & trujilo, ; lei, wang, & wu, ; li, zhang, yin, zhang, & liu, ; liu, lu, yang, & chen, ; tu & yang, ), transportation systems (li, liu, et al., ) and game systems (arikan, chenney, & forsyth, ; wan, chen, & earnshaw, ). these proposed path planning algorithms emphasize features, such as collision avoidance and ease of track- ing. although these solutions address information about roads, obstacles and maps, they do not adapt to user situations or context. context-aware path finding (moving path) has recently been studied. studies which focused on context-aware path prediction are most related to our work. these approaches adopt context, such as history data, user’s preference, location and map data, to reason the future moving path in advance. path prediction ap- proaches are able to yield high prediction accuracy and efficient network resource management and pre-configuration, thus pro- viding high quality of service for users. neural-network based pre- diction algorithm (poon & chan, ) predicts path from both global and user-level information. profile-based next-cell predic- tion (pbnp) algorithm (bharghavan & jayanth, ), shadow clus- ter concept (levine, akyildiz, & naghshineh, ), mobile motion prediction (mmp) algorithm (liu & maguire, ), per-user profile replication scheme (jannink, lam, shivakumar, widom, & cox, ), hierarchic position-prediction (hpp) algorithm (liu, bahl, & chlamtac, ), regular path recognition method (erbas, steuer, kyamakya, eggesieker, & jobmann, ) and sectorized mobility prediction algorithm (chellappa doss, jennings, & shenoy, ) predict the further location from a user’s moving history. in addi- tion, hpp and mobility tracking algorithms (zaidi & mark, ) consider instantaneous rssi measurements of surrounding cells. pbnp also considers location classification. path prediction (anag- nostopoulos, anagnostopoulos, hadjiefthymiades, kalousis, & kyriakakos, ) considers both spatial information, such as mov- ing profile and moving history, and temporal information, such as morning or afternoon. moving history is one of the important con- texts. however, if the moving history is not available, then those approaches are not applicable. hence, the mobility prediction architecture (samaan & karmouch, ) performs path prediction based on user context, such as user preferences, goals and spatial conceptual maps (scm). additionally, most approaches use high- level contexts in the prediction process. however, sigg, haseloff, and david ( ) use low-level context that contain additional information not available with high-level context to improve pre- diction accuracy. the prediction results of these approaches are useful for many context-aware applications. for instance, if the moving path is obtained, then the handoff process can be prepared such that seamless handoff can be better achieved. however, the high accuracy of preferable or habitual moving path may not be the optimal (best-fitting) path in terms of the user’s objective and requirement. consequently, planning an optimal path (short- est path) based on various contexts that best fit the user require- ments is an essential issue. genetic algorithm (ga) (holland, ) is a search method for solving large-scale optimization problems. ga is inspired by bio- logical evolution and has proved to be a robust and powerful adap- tive search technique. many researches use ga to solve path finding (ji, chen, & kitti, ), path planning (kanoh, ; castillo & trujillo, ; kanoh & hara, ) and route selection (chakraborty, maeda, & chakraborty, ) problems. numerical experiments indicate that ga can result in effective searching, shortest path, improved qos, low system resource consumption and quick decision time. therefore, the proposed ubipapago mech- anism adopts a ga-based approach to obtain the optimal path. . design of ubipapago this section introduces the design goal of ubipapago. section . presents the context model of ubipapago. section . describes the architecture of ubipapago. finally, the privacy policy of ubipapago is introduced in section . . . . design goal this study develops a human-centric, context-aware and intel- ligent path planning mechanism to satisfy individual user require- ments in a ubiquitous network. accordingly, the following objectives are proposed: ( ) the mechanism should be scalable and possible to implement in a distributed manner. ( ) based on various contexts acquired from the environment, the solution should be able to automatically plan an optimal path that best fits the user requirements, such as network connectivity, bandwidth and available services. ubipapago relies on various types of context stored in an open and distribute framework (lu et al., ). this framework allows a digital device to join the ubiquitous environment and search for context-aware services based on standard protocols, such as uni- versal plug and play (upnp) (universal plug and play), hypertext transfer protocol (http) (fielding et al., ) and dns. a service or application, such as ubipapago, can be easily deployed into the system through standard protocols and distributed context managers. . . context model table presents the context model of ubipapago. in a ubiqui- tous network, billions of sensors located in our physical environ- ment accumulate a huge amount of context. these contexts are fed into context model for context-aware path planning applica- tions. this study limits the description of context needed to sup- port path planning mechanism. the context model includes static and dynamic types describing the user, terminal and map. the user profile contains details of the user. the user preference allows a user to specify the required service and cost constraints. meanwhile, the proposed system also considers user privacy. the interface privacy setting enables users to specify their privacy pol- icies, since user privacy is always one of the most important issues in developing context-aware applications. users can hide sensitive information, such as schedule and location, or filter access to per- sonal context, using their privacy settings. the terminal profile includes id, capability, software interface status and types of running application. the terminal capability describes the process speed, memory, screen size, resolution and table context model of ubipapago. static dynamic user – user identity – user preference � required service � cost constrain � habit – privacy setting – schedule – location – moving direction – user preferences terminal – terminal id – capability � process speed � memory � screen size � resolution � supported interface types – application type � voip, video conference � iptv, mobile tv � email, ftp � im, web browsing – interface status – running application type map – block id – block distance – accessible service � conversational � real-time streaming � non-real-time streaming � interactive � background – qos parameter of accessible service � rssi � required bandwidth � delay � jitter � packet loss ratio – accessible context server – accessible ap/bs – qos parameter of accessible ap/ bs � rssi � required bandwidth � delay � jitter � packet loss ratio server wlan service providing server context management server middleware (web service) c on text a cqu isition g gps rfid zigbee hand-held devices client application : ubipapago sensor lan wimax network fig. . architecture of ubipapago. c.-y. wang et al. / expert systems with applications ( ) – supported interface types. the terminal application types describe software applications installed in the terminal. application types based on quality of service (qos) parameters, such as response time, delay, jitters and bandwidth, can be categorized into four types, namely (i) conversational service, such as voip and video conferencing; (ii) real-time (rt) service, such as internet protocol television (iptv) and mobile tv; (iii) non-real-time (nrt) services, such as email and ftp; (iv) interactive services, such as instant mes- sage (im) and web browsing. the map and environment profile includes block id, block dis- tance, accessibility of ap/bs, qos parameter of accessible ap/bs, accessible service, qos parameter of accessibility service, and accessible context server (u-gate). these contexts are represented using an scm model, which is described in detail in section . the qos parameters of a service are pre-defined based on service pro- file and stored in the context server. the qos of ap/bs is dynami- cally updated to the context management server. . . architecture of ubipapago fig. shows the ubipapago architecture. the ubipapago servers are composed of a service providing server and a context manage- ment server. the service providing server provides the intelligent human-centric services. the context management server mainly acts as a semantic database server to manage context accumulated from the sensors in the environment. the context management server manages and refers to context by constructing an ontology tree based on owl and rdf. the hand-held device includes a ubi- papago client application to find the best-fitting path. the commu- nication model (lu et al., ) among the device and the server is mainly based on the standard protocols, including universal plug and play (upnp), simple service discovery protocol (ssdp) (simple service discovery protocol), simple object access protocol (soap) (simple object access protocol), general event notification proto- col (gena) (general event notification architecture) and hyper- text transfer protocol (http). . . privacy policy ubipapago relies on various contexts which are highly related to personal privacy. protecting user privacy is always one of the most important issues in developing context-aware applications. unfortunately, due to large amount of context information and de- vices in dynamic and heterogeneous environments, creating a total solution for managing security and privacy policies is known to be a very challenging open issue in ubiquitous environments. in cur- rent development, ubipapago tries to preserve privacy by using a privacy setting to hide sensitive information, such as schedule, location, moving direction and user preference in context manage- ment server, and filtering access to personal context based on user’s preference and privacy setting. future versions of ubipapago will deploy advanced mechanisms or policies for preserving pri- vacy on the context management server. . ubipapago based on scm model this section represents the trajectory map and environment context using a scm model. the scm is an abstraction of real map representation containing of a set of landmark objects (oj), a set of medium objects (wy) and the influence areas of those ob- jects. the landmark objects define those areas, such as buildings or user’s target destination. the medium objects (also called ‘‘ways”) define areas, such as streets, roads, trajectories and virtual connections between objects (kettani & moulin, ). a way ob- ject is further partitioned into continuous blocks, called way ele- mentary areas (weas). to encode additional context information, a novel approach for partitioning a way object is proposed as follows. a wea may be a landmark object (oj) or a crossing point of two or more ways (ket- tani & moulin, ). the proposed solution requires additional contexts, including received signal strength indication (rssi) of fig. . a portion of taipei city of taiwan map. c.-y. wang et al. / expert systems with applications ( ) – ap/bs and available bandwidth. thus, besides partitioning a way object based on crossings, the proposed solution also considers characteristics of ap/bs, such as rssi. specifically, a way object is partitioned into several continuous blocks (weas) such that the rssi of an ap/bs within each wea is roughly at the same level. fig. (a) illustrates a portion of the map of taipei city in taiwan, and the distribution of ap/bs (wifly). to clearly show the distribu- tion and the context of ap/bs, the blue circle in fig. (a) shows an enlarged portion of the map. in this study, a series of simulations were performed using the qualnet network simulator (qualnet) to implement the context of the ap/bs, as shown in fig. (b). fig. shows its scm representation. for interpretation of color in figs. , , – , the reader is referred to the web version of this article. the matrix of orientation and adjacency and characteristic (moac) (samaan & karmouch, ) was adopted to represent the weas of the scm. the moac contains characteristic informa- tion of each wea and landmark object, and the relative informa- tion and displacement direction between two weas. fig. shows a portion of the moac of fig. . the diagonal cell (x, x) represents the characteristics, i.e. context of wea ax. the cell (x, y), where x – y denotes the characteristics of relative direction between ax and ay, which can be represented in eight-way directions, namely, east (e), west (w), south (s), north (n), northeast (ne), southeast (se), southwest (sw) and northwest (nw). for instance, a is on the south (s) side of a , and a is on the northwest side (nw) of a . according to moac, the list of blocks can be inferred along a moving path, along with the direction of movement and all re- quired context information. fig. . scm representation of taipei city of taiwan map of fig. . fig. . a portion of moac of fig. . c.-y. wang et al. / expert systems with applications ( ) – fig. . path planning map. c.-y. wang et al. / expert systems with applications ( ) – the characteristic function co( ) describes several characteris- tics of an object. specifically, co(oj) and co(ai) describe the context information of a landmark object and a wea, respectively. more- over, since state information of aps is also considered in this study, each ap is associated with a characteristic function co(api). for in- stance, if o is a rapid metro station, then the characteristic func- tion denotes the characteristics in terms of name, location. thus an example of co(o ) could be {c = rapid transit station, c = a }. similarly, if wea a is covered by two aps and provides a service, then the characteristic function of co(a ) is denoted by {c = {ap , ap }, c = {streaming server}}. the characteristics of ap are described in terms of available rssi (c ), bandwidth (c ), delay (c ), jitter (c ), and packet loss ratio (c ). an example of co(ap ) is {c = � dbm, c = mbps, c = ms, c = ms, c = %}. . ubipapago based on ga this section describes our context-aware path planning mecha- nism based on ga. the ga procedure is shown in fig. . before applying ga, the scm model obtained in section is first converted to path planning map, as shown in fig. , where node ai denotes a crossing in fig. . restated, to reduce the length of chromosome and to increase the scalability of ubipapago, the path planning map considers only crossing nodes. additionally, the path planning map needs to consider the source or destination node, if these are not at a crossing. however, the fitness function of a chromosome is still calculated according to all weas and their characteristic func- tions, as described later. the path planning map is an undirected graph g = (v, e), where v denotes a set of crossing, source and destination nodes, and e de- notes a set of edges (weas). the aim is to identify the shortest path for g. each edge e e e includes blocks(ai). corresponding to each edge, a weight xij denotes the fitness value calculated from the fit- ness function from node ai to node aj. moreover, if the user’s re- quired service is assumed to be located at the edge, then this edge is called a constraint edge e e e. the found path ~s must con- tain this edge, so that the required service can be accessed. for in- stance, if the streaming services are located at block a and a , respectively, then the links ( , ) and ( , ) are constraint edges that must be included in the path. note that end-to-end initialization population evaluate fitness termination criterion met? solution start end selection/ reproduction crossover mutation no yes fig. . the procedure of ga. qos attributes are evaluated, e.g., the delay represents the end- to-end delay of a path. the ubipapago defines the optimal path as the shortest path subject to seven constraints, including received signal strength indication (rssi), available bandwidth (bw), response time (rt), delivery latency (dl), jitter (j) and packet loss ratio (pl), available service (as). the optimization problem is formulated as follows: minimize dð~sÞ¼ x k ~s distðekÞ ð Þ subject to linkðekÞ ~s : xðekÞ p hx; x ¼frssi; bwg; x x ð Þx k s yðekÞ hy; y ¼frt; dl; j; plg; y y ð Þ as~s exists required service ð Þ where dð~sÞ denotes the total distance of moving path, dist(ek) de- notes the distance of edge ek, hx and hy are thresholds for rssi, rt, etc. the constraints ( )–( ) ensure that the computed result indeed satisfies the user’s requirements. . . chromosome representation and population initialization chromosome representation is a key to the design of ga for solving a problem. this study adopts the encoding method pro- posed in wan et al. ( ), where all chromosomes have the same length, thus facilitating crossover operation. fig. gives an exam- ple of chromosome representation, where the upper sequence is the node id and the lower sequence is the chromosome. here a chromosome is represented by a sequence of integer numbers; a gene denotes the node id through which it passes. for an edge (vi, vj) from node vi to node vj, node vj is put in the ith id of the chro- mosome. if the id is not along the path, then its corresponding gene is filled with a neighboring node randomly chosen from the set of its neighboring nodes. for instance, the chromosome in fig. de- notes a path from source node id (block a ) to destination node id (block a ), passing through a , a , a and a . the initial population influences the convergence speed and the quality of the solution. random and heuristic initializations are the most commonly used ways of generating candidate solutions. however, random initialization possibly causes a large computing overhead and loop path, while heuristic initialization (arikan et al., ) has difficulty in finding a global optimal. heuristic initializa- tion explores a small part of solution space, making it faster than random initialization. this study adopts the probability network approach (poon & chan, ) that combines the features of ran- dom and heuristic initialization to generate initial candidate paths in the gene pool. starting with the source node, the next node is chosen from its directly connected neighbors. to speed up the con- vergence speed, each neighboring node is assigned a different selection probability. a neighboring node located in the direction id gene chromosome ) () ( ( ) ( ) ( ) fig. . example of encoding path from a to a . c.-y. wang et al. / expert systems with applications ( ) – toward the destination is assigned a higher probability. the se- lected next node then becomes the gene of the source node’s id. this procedure repeats until the path reaches the destination node. to avoid looping, a node is forbidden to visit twice in the same path during the next node selection procedure. considering the example shown in fig. , a has four neighboring nodes, a , a , a and a (cf. fig. ), among which a and a have higher probabilities to be chosen as the next node of a , since they are located in the direction toward a and have not appeared in the path before. for nodes that are not in the path under consideration, their next nodes are randomly selected. for instance, a , a and a can be chosen as the next node of a . . . fitness function the fitness function measures the distance of the path that satis- fies all the quality constraints. the concept of penalty function is adopted to handle the constraints (yeniay, ). in this study, the quality of a chromosome depends on properties of each link(ek) of the path, including received signal strength indication (rssi), avail- able bandwidth (b), response time (rt), delivery latency (dl), jitter (j), packet loss (pl) and available service (as). the fitness function is defined by f ¼ dð~sÞ � ð þ apÞ ð Þ with p ¼ x k ~s pðekÞþ hasð~sÞ ð Þ where pðekÞ¼ x x frssi;bwg fxðekÞþ x y frt;dl;j;plg gyðekÞ ð Þ fxðekÞ¼ ; if xðekÞ hx ; otherwise � x frssi; bwg ð Þ gyðekÞ¼ ; if p k s yðekÞ p hy ; otherwise ( y frt; dl; j; plg ð Þ id parent parent mask offspring chro fig. . an example of c hasð~sÞ¼ ; if required service is not available along path ; otherwise � ð Þ . . selection the selection operation chooses a chromosome from the cur- rent population for reproduction. the ubipapago adopts tourna- ment selection (goldberg & deb, ) to select parents for producing offspring. in tournament selection, two individuals are chosen at random from the population, and the one with the higher fitness value is chosen as a parent for the crossover operation. the tournament process repeats until two parents are selected. . . crossover the crossover operation recombines two chromosomes (called parents) to produce a new chromosome (called offspring). this study proposes a modified uniform crossover by additionally con- sidering genes (nodes) that have not appeared (been visited) be- fore. the proposed crossover operation is carefully designed to avoid the path looping problem. the crossover starts with the gene at the id of the source node. if the two parental genes at this id are identical, then this gene is simply copied to the offspring. other- wise, the gene that has not appeared before is chosen. a tie is bro- ken by random selection. after selecting the gene of the current id, the procedure moves to the id of the next node (i.e., the gene we just selected), and is repeated until the destination node is reached as the gene of the current id. fig. shows an example of the crossover operation. the path in parent is a ? a ? a ? a ? a ? a , and that of par- ent is a ? a ? a ? a ? a ? a . for the node id , the operation randomly selects among a and a . if a is selected, then node id is considered. the operation is then repeated until the destination node is reached. the new offspring a ? a ? a ? a ? a ? a ? a ? a ? a ? a is generated after the crossover operation. for the id’s in which the path does not pass through, e.g. , the genes at them are randomly selected from either parent, for example, a in node id . mosome rossover operation. id before after fig. . an example of mutation operator. c.-y. wang et al. / expert systems with applications ( ) – . . mutation with a given probability, mutation introduces variation into the chromosome by adjusting partial chromosome to avoid local opti- ma. in this study, a random node (mutation node) that is neither a source nor a destination node along a path is chosen for mutation (wan et al., ). the new gene value is randomly chosen from directly connected nodes of the mutation node. fig. shows an example of mutation operation. the old path is a ? a ? a ? a ? a ? a . in the mutation operation, the node id a is randomly chosen as the mutation gene. the new path following the mutation operation is a ? a ? a ? a ? a ? a ? a ? a . after crossover and mutation operations are completed, the off- springs are evaluated by the fitness function for another round of reproduction until a termination condition is satisfied. . prototyping and implementation this section demonstrates the feasibility of the ubipapago by prototyping a real system in u-gate. because u-gate is an open, effi- cient and distributed context management architecture and com- munication model based on standard protocols, ubipapago can be implemented in a distributed manner, enabling any service provider to deploy its service into the ubiquitous environment easily. for easy deployment, ubipapago was implemented in a small scope environ- ment. therefore, the demonstration scenario is in our campus (chung-cheng university), as displayed in fig. , and made up with user using eeepc with wifi and rfid. in the scenario, the user uses ubipapago service in eeepc to connect ubipapago service provider (u-gate) in local area, as shown in fig. (a) and (b), to plan a path from the college of engineering (o ) to the auditorium (o ) in the ccu campus. u-gate collects the contexts of user’s current location, requirement, id, schedule and scm map to infer the path planning result, as shown in fig. (c). readers are referred to lu et al. ( ) for the detailed implementation of u-gate. fig. . the implementation . experimental results and discussion this section evaluates the proposed ubipapago through com- puter simulation. the proposed ubipapago was implemented in a java (jdk ) program on a pc with a core quad . ghz cpu and g mb memory. fig. shows the road map of some part of taipei city that was used in the experiments. in the experiment, the road map with nodes was first converted into a graph with crossing nodes. the parameters are set as follows. the experiment tries five population sizes (n): , , , , and . termination cri- terion is generation. the crossover probability (pc) was set to , and the mutation probability (pm) was set to . . the tournament size was set to . the alpha (a) value of the penalty function weight was empirically set as a = . the collected results were averaged over runs. notably, ubipapago uses probability network based initializa- tion and a modified crossover. to verify the effect of these two ele- ments, we compare the performance of ubipapago with that of ga, i.e. using random initialization and uniform crossover. the perfor- mance metric, mean best fitness (mbf), is defined by mbf ¼ sum of the best fitness value of each run number of runs ð Þ . . optimal path evaluation fig. (a)–(c) shows the best obtained path, the optimal path, and the shortest path for the indicated source and destination. the result of the proposed ubipapago in fig. (a) coincides with the optimal path in fig. (b) and the shortest path in fig. (c). table summarizes the results of qos constraint and parameters of the obtained path in fig. (a)–(c). the qos parameters of the resulting path means the minimum rssi and bandwidth and max- imum delay, jitter and packet loss rate along the path. the best ob- tained path of ubipapago guarantees shorter path and qos to scenario of ubipapago. fig. . the demonstration of ubipapago. c.-y. wang et al. / expert systems with applications ( ) – satisfy user’s requirements, while the shortest path has shortest distance but does not satisfy user’s requirements. . . convergence behavior fig. compares the convergence behavior of the ubipapago and ga. the result is helpful for choosing the appropriate popula- tion size and termination condition. moreover, it shows that a small population size reduces evaluation cost but results in prema- ture convergence. on the other hand, a large population size finds better solutions and yet requires longer computation time. we can also observe that ubipapago converges faster than ga. according to the result in fig. , the best convergence curve of ubipapago was achieved with population size of in generations. fig. . the portion of taipei city. c.-y. wang et al. / expert systems with applications ( ) – therefore, unless otherwise stated, the following experiments of ubipapago were performed with population size and generations. the ga differs from ubipapago by adopting a random approach in the initialization and crossover operations, such that some chromosomes may contain loops. hence, the ubipapago converges faster than the ga. the results indicate that the ubipapago improves the convergence speed by preventing looping in the initialization and crossover operations. fig. . path comparison (paths are table the qos constraint and result of found path, optimal path and shortest path. parameters distance fitness value min rssi min b range n/a n/a � to � – threshold n/a n/a p� p best obtained path of ubipapago � optimal path � shortest path � a a the qos parameter is not satisfied user’s requirement. . . optimization results comparison table compares the performance of the ubipapago and ga in terms of mbf, average computation time, probability of obtaining the best path. note that the ga uses random initialization and uni- form crossover, while ubipapago uses probability network ap- proach initialization and modified uniform crossover. the mbf values indicate that ubipapago generally outperforms the ga in terms of fitness. the effectiveness of the ubipapago was measured under vari- ous numbers of generations. the ubipapago was found to have a shorter computation time and significantly higher success rate than the ga, because ubipapago is a probability network approach in which the next node is located forward of the destination with a higher probability. additionally, ubipapago avoids looping. the probability of obtaining the optimal solution is higher than % with a population size of . this approach is also efficient, be- cause the population size is small and the evolution process is short. the result is very important and applicable for context- aware applications. short computation time and high success prob- ability enable ubipapago to achieve satisfactory solution quality and real time service. table also shows that ubipapago has lower standard deviations than ga does. . . effect of alpha value finally, table compares the fitness, distance, and penalty val- ues at different alpha values ranging from . to . . note that pop- ulation size with generations are adopted herein. the table shows that the increaseing alpha value causes decreasing penalty value, increasing average fitness, and increasing distance value. this is because a high alpha value intensifies penalty so that the evolutionary search tends to avoid penalty and to find a solution without penalty but with ordinary distance. conversely, a low al- pha encourages finding a short route that may violate some constraints. denoted by blue dotted lines). andwidth max delay max jitter max packet loss rate available service – – . – . . . = . . a a . a a fig. . convergence curve of ubipapago and ga. table the qos constrain and result of found path. generation number fitness average computation time (s) probability of obtaining the optimal path best mbf std ubipapago n = . . . . n = . . . . n = . . . . n = . . . . n = . . . . n = . . . . ga n = . . . . n = . . . . n = . . . . n = . . . . n = . . . . n = . . . . table fitness, distance and penalty for alpha values . , . , . , and . . alpha (a) parameters average fitness (f) average distance (dð~sÞ) average penalty (p) . . . . . . . . . . . . . . . c.-y. wang et al. / expert systems with applications ( ) – . conclusions and future work this study proposes an intelligent navigation application, ubi- papago, which focuses on providing human-centric context-aware path planning mechanism in a ubiquitous network. specifically, ubipapago considers requirements and contexts of users and envi- ronment when planning the best-fitting path. ubipapago adopts the scm model to present map context and uses ga to find the optimal path. the map simplifies conversion to graph, thus reduc- ing the computational complexity and increasing scalability. experimental results examine the effect of initialization and pro- posed crossover. moreover, the results validate that the proposed ubipapago can efficiently achieve the shortest path that guaran- tees all qos parameters. some issues of the proposed ubipapago remain for further re- search. this study assumes that the destination is inferred from the user’s personal context, such as schedule. if the destination of the path is uncertain or unknown, predicting the destination from the related context becomes an a priori problem. additionally, the proposed system only considers how to plan an optimal path for a user but does not consider issues, such as vertical handoff between ap and bs, and resource reservation along the path in advance to guarantee quality of service of the path. finally, security and pri- vacy issues also need further investigation. advanced framework or policies for preserving privacy will be deployed in the future. acknowledgment the research was supported by the nsc - -e- - - my and nsc - -e- - -my , national science council, roc. references anagnostopoulos, t., anagnostopoulos, c., hadjiefthymiades, s., kalousis, a., & kyriakakos, m. ( ). path prediction through data mining. in ieee international conference on pervasive services (icps), istanbul, turkey (pp. – ). arikan, o., chenney, s., & forsyth, d. a. ( ). efficient multi-agent path planning. in eurographic workshop on computer animation and simulation, manchester, uk (pp. – ). bharghavan, v., & jayanth, m. ( ). profile-based next-cell prediction in indoor wireless lan. in ieee singapore international conference. bodhale, d., afzulpurkar, n., & thanh, n. t. ( ). path planning for a mobile robot in a dynamic environment. in ieee international conference on robotics and biomimetics (pp. – ). castilho, o., & trujilo, l. ( ). multiple objective optimization genetic algorithms for path planning in autonomous mobile robots. international journal of computers, systems and signals, ( ), – . castillo, o., & trujillo, l. ( ). patricia melin, multiple objective genetic algorithms for path-planning optimization in autonomous mobile robots. soft computing – a fusion of foundations, methodologies and applications, ( ), – . chakraborty, b., maeda, t., & chakraborty, g. ( ). multiobjective route selection for car navigation system using genetic algorithm. in ieee mid-summer workshop on soft computing in industrial application (smxia/ ) (pp. – ). chellappa doss, r., jennings, a., & shenoy, n. ( ). mobility prediction for seamless mobility in wireless networks. in th international network conference (inc) (pp. – ). erbas, f., steuer, j., kyamakya, k., eggesieker, d., & jobmann, k. ( ). a regular path recognition method and prediction of user movements in wireless networks. in ieee vehicular technology conference, atlantic city (pp. – ). fielding, r., gettys, j., mogul, j., frystyk, h., masinter, l., leach, p., et al. ( ). hypertext transfer protocol – http/ . , ietf rfc , june . general event notification architecture (gena). . goldberg, d. e., & deb, k. ( ). a comparative analysis of selection schemes used in genetic algorithms. foundation of genetic algorithms (pp. – ). san mateo calif: morgan kaufman. holland, j. ( ). adaptation in natural and artificial systems. the university of michigan press. inagaki, j., haseyama, m., & kitajima, h. ( ). a genetic algorithm for determining multiple routes and its applications. in ieee international symposium (iscas’ ) (pp. – ). jannink, j., lam, d., shivakumar, n., widom, j., & cox, d. ( ). efficient and flexible location management techniques for wireless communication system. acm/ baltzer wireless networks, ( ), – . http://www.tools.ietf.org/html/draft-cohen-gena-p-base- http://www.tools.ietf.org/html/draft-cohen-gena-p-base- c.-y. wang et al. / expert systems with applications ( ) – ji, z. w., chen, a., & kitti, s. ( ). finding multi-objective paths in stochastic networks: a simulation-based genetic algorithm approach. in congress on evolution computation (cec ) (pp. – ). los alamitos: ieee press. kanoh, h. ( ). dynamic route planning for car navigation systems using virus genetic algorithms. international journal of knowledge-based and intelligent engineering systems, ( ), – . kanoh, h., & hara, k. ( ). hybrid genetic algorithm for dynamic multi-objective route planning with predicted traffic in a real-world road network. in genetic and evolutionary computation conference (gecco ) (pp. – ). kettani, d., & moulin, b. ( ). a spatial model based on the notions of spatial conceptual map and of object’s influence areas. in conference spatial information theory (cosit) (pp. – ). lei, l., wang, h., & wu, q. ( ). improved genetic algorithms based path planning of mobile robot under dynamic unknown environment. in ieee international conference on mechatronics and automation (pp. – ). levine, d., akyildiz, i., & naghshineh, m. ( ). a resource estimation and call admission algorithm for wireless multimedia networks using the shadow cluster concept. ieee/acm transactions on networking, ( ), – . li, q., liu, g., zhang, w., zhao, c., yin, y., & wang, z. ( ). a specific genetic algorithm for optimum path planning in intelligent transportation system. in th international conference on its telecommunications (pp. – ). li, q., zhang, w., yin, y., zhang, w., & liu, g. ( ). an improved genetic algorithm of optimum path planning for mobile robots. in the sixth international conference on intelligent systems design and applications (pp. – ). liu, t., bahl, p., & chlamtac, i. ( ). mobility modeling, location tracking, and trajectory prediction in wireless atm networks. ieee jac, ( ), – . liu, z., lu, q., yang, p., & chen, l. ( ). path planning for tractor-trailer mobile robot system based on equivalent size. in th world congress on intelligent control and automation (wcica) (pp. – ). liu, g., & maguire, g. ( ). a class of mobile motion prediction algorithms for wireless mobile computing and communications. acm/baltzer monet, ( ), – . lu, j. h., wang, c. y., & hwang, r. h. ( ). an open framework for distributed context management in ubiquitous environment. in international symposium on ubicom frontiers – innovative research, systems and technologies (ufirst- ), brisbane, australia (pp. – ). owl. w c document. . papago. . poon, w. t., & chan, e. ( ). traffic management in wireless atm network using a hierarchical neural-network based predication algorithm. in th international conference on computers and their applications, new orleans, united states (pp. – ). qualnet, scalable network technologies inc. . rdf primer. w c document. . rdf schema. w c document. . samaan, n., & karmouch, a. ( ). a mobility prediction architecture based on contextual knowledge and spatial conceptual maps. ieee transactions on mobile computing, ( ), – . sigg, s., haseloff, s., & david, k. ( ). a novel approach to context prediction in ubicomp environments. in th annual ieee international symposium on personal, indoor and mobile radio communications (pimrc’ ) (pp. – ). simple object access protocol (soap) version . . . simple service discovery protocol (ssdp). . tomtom international. . tu, j., & yang, s. ( ). genetic algorithm based path planning for a mobile robot. in ieee international conference on robotics and automation, taipei (pp. – ). universal plug and play (upnp). . wan, t. r., chen, h., & earnshaw, r. ( ). real-time path planning for navigation in unknown environment. in theory and practice of computer graphics (pp. – ). weiser, m. ( ). the computer for the twenty-first century. scientific american, – . wifly. . yeniay, Ö. ( ). penalty function methods for constrained optimization with genetic algorithms. journal of mathematical and computation application, ( ), – . zaidi, z. r., & mark, b. l. ( ). real-time mobility tracking algorithms for cellular networks based on kalman filtering. ieee transactions on mobile computing, ( ), – . http://www.w .org/ /owl/ http://www.papago.com.tw/ http://www.scalable-networks.com http://www.w .org/tr/rdf-primer/ http://www.w .org/tr/rdf-schema/ http://www.w .org/tr/soap/ http://www.pwg.org/pub/pwg/ipp/new_ssdp/draft-cai-ssdp-v - .txt http://www.pwg.org/pub/pwg/ipp/new_ssdp/draft-cai-ssdp-v - .txt http://www.tomtom.com/ http://www.upnp.org/ http://www.wifly.com.tw/wifly /tw/wiflynet/servicerange/hotpoint ubipapago: context-aware path planning introduction related work design of ubipapago design goal context model architecture of ubipapago privacy policy ubipapago based on scm model ubipapago based on ga chromosome representation and population initialization fitness function selection crossover mutation prototyping and implementation experimental results and discussion optimal path evaluation convergence behavior optimization results comparison effect of alpha value conclusions and future work acknowledgment references text categorization methods for automatic estimation of verbal intelligence f. fernández-mart́ıneza, k. zablotskayab, w. minkerb agrupo de tecnoloǵıa del habla, universidad politécnica de madrid, madrid, spain. binstitute of communications engineering, university of ulm, germany abstract in this paper we investigate whether conventional text categorisation methods may su!ce to infer di"erent verbal intelligence levels. this research goal relies on the hypothesis that the vocabulary that speakers make use of reflects their verbal intelligence levels. automatic verbal intelligence estima- tion of users in a spoken language dialogue system may be useful when defining an optimal dialogue strategy by improving its adaptation capabili- ties. the work is based on a corpus containing descriptions (i.e. monologues) of a short film by test persons yielding di"erent educational backgrounds and the verbal intelligence scores of the speakers. first, a one-way analysis of vari- ance was performed to compare the monologues with the film transcription and to demonstrate that there are di"erences in the vocabulary used by the test persons yielding di"erent verbal intelligence levels. then, for the classi- fication task, the monologues were represented as feature vectors using the classical tf-idf weighting scheme. the naive bayes, k-nearest neighbours and rocchio classifiers were tested. in this paper we describe and compare these classification approaches, define the optimal classification parameters and discuss the classification results obtained. keywords: spoken dialogue systems, naive bayes classification, rocchio approach, k-nearest neighbours email addresses: ffm@die.upm.es (f. fernández-mart́ınez), kseniya.zablotskaya@uni-ulm.de (k. zablotskaya), wolfgang.minker@uni-ulm.de (w. minker) preprint submitted to expert systems with applications december , *manuscript click here to view linked references figure : spoken language dialogue system . introduction next-generation spoken language dialogue systems (slds), developed to provide users with required information and/or to help them to accomplish certain goals, are expected to be able to deal with di!cult tasks and react to a wide range of situations and problems. they should help users to feel free and comfortable when interacting with them. moreover, they should also be user-friendly and easy to use. including aspects of adaptation to users into slds may help to increase the systems’ communicative competences and influence on their acceptability (figure ). next-generation slds may change the level of dialogue depending on users’ experience. for example, a spoken dialogue system aimed at providing guidance and support for the installation of some software may try to estimate whether the user is an ex- pert or a novice in this field. based on this information suitable words and explanations may be generated. these explanations may be very detailed and without specific vocabulary for a non-experienced user; in contrast, for an expert, the system may provide only a sequence of important steps or inform about more di!cult operations. from the beginning of the dialogue, slds may analyse the user’s speech, behaviour and requests and also the ex- isting di!culties. when deciding on the best response to a user, the dialogue manager may change words and sentence structures based on the informa- tion about cognitive processes. its responses may become more helpful and the user-friendliness of the system may be improved. for this purpose it is necessary to identify di"erences in language use of people yielding di"er- ent educational background and abilities to analyse situations and to solve problems. the ability to use language for accomplishing certain goals is called ver- bal intelligence (vi) [ , ]. in other words, verbal intelligence is “the ability to analyse information and to solve problems using language-based reason- ing” [ ]. automatic verbal intelligence estimation may help dialogue sys- tems to choose the level of communication and be more simple, useful and e"ective. figure explains the adaptation process of spoken dialogue systems based on verbal intelligence estimation in more detail. when talking to the system, all j spoken utterances of a user are analysed for the verbal intelligence determination. this means that the intelligence level is re-estimated at each turn based on features extracted from the new spoken utterances and from all the phrases which were pronounced at the previous turns. in figure the slds has three di"erent dialogue scenarios corresponding to users yielding a higher, an average and a lower verbal intelligence. at the beginning of the dialogue, the systems uses scenarios corresponding to users yielding an average verbal intelligence. at the following turns, the system might switch to alternative dialogue scenarios. figure : adaptation to the user. the automatic estimation of users’ verbal intelligence may help slds to more e"ectively control the flow of the dialogues, engage users in the interaction and be more attentive to human needs and preferences. for training machine learning algorithms, we need to know a maximum number of language features that reflect di"erences in language use of people yielding di"erent verbal intelligence. in this work we investigate to which extent the vocabulary of test persons reflect their levels of verbal intelligence when they all describe the same event and explain their thoughts and feelings about it. the investigation is based on a corpus containing descriptions of a short film along with the corresponding intelligence scores of the speakers [ ] . the paper is structured as follows. in section we describe the cor- pus which was used for the experimental research. section describes our primary e"orts at defining film related features which could be useful for distinguishing test persons yielding di"erent verbal intelligence. section describes typical tf-idf approaches and explains the details of the feature selection process for the monologues. in section we describe and compare the naive bayes, k-nearest neighbours and rocchio classifiers. classification results are presented and discussed in section . finally, section presents conclusions and future work. . corpus description for the data acquisition in [ ], a short film was shown to german native speakers. it described an experiment on how long people could stay without sleep. the test persons were asked to imagine that they met an old friend and wanted to tell him about this film. our goal was to record everyday speech when talking to relatives and friends. this corpus, described in [ ], consists of descriptions ( , hours of audio data) of a short film (i.e. monologues). the test persons were also asked to participate in the verbal part of the hamburg wechsler intelligence test for adults (hawie) [ ]. according to wechsler, intelligence is “the global capacity of a person to act purposefully, to think rationally, and to deal e"ectively with his environment” [ ]. the verbal part consists of the following subtests: • information: this subtest measures general knowledge and includes questions about history, geography, literatures, etc. for example, what is the capital of italy? • comprehension: test persons are asked to solve di"erent practical prob- lems and explain some social situations. for example, what would you do if you lost your way in a forest? • digit span: test persons are asked to repeat increasingly longer strings of numbers forward and then backward; the subtest measures short- term memory. • arithmetic: test persons are asked to solve some arithmetic problems given in a story-telling way; the subtest measures their concentration and computational ability. for example, how many rolls you can buy if you have cents and one roll costs cents? • similarities: test persons are asked to find a similarity between a pair of words. for example, please find a similarity between “wood” and “alcohol”? • vocabulary: test persons are asked to explain increasingly more di!- cult words using their vocabulary. for example, what does “to creep” mean? raw scores of each test person on the verbal test are based on the correct answers (figure ). the raw scores are then converted into “scaled scores” using special tables [ ]. the scaled scores vary between and and may be used to compare the performance of the participants. the sum of the scaled scores and the age of a test person are used to estimate the corresponding verbal intelligence score.  figure : verbal part of the hamburg wechsler intelligence test for adults overall, test persons yielding di"erent educational levels were tested, therefore monologues about the same topic were collected. all the mono- logues and the film were transcribed according to the transcription standards by mergenthaler [ ]. . modelling verbal intelligence by using film derived features to analyse the vocabulary of people yielding di"erent verbal intelligence when describing the same event, at first we decided to compare the mono- logues with the film transcription. figure shows excerpts from the film and from one of the monologues . excerpt from the film max and funda have been without sleep for fifty eight hours. they have laid down on the sofa. is it a mistake? actually they would like to move. but now they cannot any more. the blood pressure is down, the energy reserves are over. they both are freezing despite the fire-place and the jacket. the question is who closes the eyes first. it is max. funda wins. she stays awake a few minutes longer. excerpt from a corresponding monologue after fifty eight hours, they were really tired. and, they had frozen. despite they had very warm clothes. and then the man fell asleep and then the woman. figure : excerpts from the film and one of the recorded monologues. for the comparison, the following features were extracted: • number of reused words - number of words which a test person “reused” from the film. for our example in figure the reused words are: fifty, eight, hours, they, and, they, despite, they, and, the, and, the. • number of unique reused words. it includes the number of reused words without repetitions. in figure , the unique reused words are: fifty, eight, hours, they, and, despite, the. • number of all reused lemmas. this feature is similar to the number of all reused words, but referred to lemmas instead. as the conversation language is german, the example was directly translated into english. • number of unique reused lemmas. this feature is similar to the number of unique reused words, but considering unique lemmas instead. • cosine similarity between the film and a kth monologue using lemmas. for this feature extraction, we created a matrix consisting of all unique lemmas from the film, including the frequency of these lemmas within the film and within a kth monologue. table shows this matrix for the texts from our example (figure ). table : matrix for lemma frequency lemmas from film frequency (film) frequency (monologue) max and funda have be without sleep for fifty eight hour the frequencies were normalized by the total amount of words in the corresponding text; the cosine similarity between the two normalized vectors (lemma frequencies within the film and lemma frequencies within a kth monologue) was calculated as: similarity = !n i= aibi!n i= ai !n i= bi , where n is the number of unique lemmas in the film, ai - frequency of ith lemma in the film, bi - frequency of ith lemma in the monologue . • number of reused n-grams. for this feature we have calculated the number of n-grams (n = , ) that were used in the film and then cosine similarity will be further used in the rocchio classification approach that will be explained in detail in section . . reused by a test-person in the corresponding monologue. in our exam- ple, the number of reused -grams equals to (reused -grams are fifty eight and eight hour), the number of reused -grams equals to (fifty eight hour), etc. • cosine similarity using n-grams. the cosine similarity was calculated from a feature vector composed by the counts of di"erent n-grams for each monologue. • we also determined the number of lemmas that were used by the candi- dates but were not used in the film. for each monologue the following features have been calculated: own lemmas = n" i= frequency(lemmai) ! count(lemmai) and own lemmas = n" i= frequency(lemmai), where n is the number of unique lemmas that were used by a test person but were not used in the film; count(wordi) shows how many times lemmai was used in the monologue; frequency(lemmai) shows the frequency of lemmai according to a frequency dictionary of the german language [ ]. this dictionary consists of german words with frequency from to : corresponds to more frequent words, corresponds to less frequent words. if a word from the monologues was not found in the dictionary, its frequency was set to . . . feature analysis the k-means algorithm, which is frequently used for data clustering in machine learning, was applied on the scaled scores of the test persons (fig- ure ). for the feature analysis two experiments were performed. in the first experiment the observations were partitioned into two clusters: cluster p consisted of test persons yielding a lower verbal intelligence, p contained candidates yielding a higher verbal intelligence. in the second experiment the test persons were partitioned into three clusters: p - lower verbal intel- ligence, p - average verbal intelligence, p - higher verbal intelligence.       figure : the k-means algorithm the averaged values of all the features from the two clusters were com- pared using a one-way analysis of variance (anova). in experiment i with two clusters, features with small p-values were: • number of reused -grams (averaged value for the first class avlow = . , averaged value for the second class avhigh = . , p = . , f = , ); • cosine similarity using lemmas (avlow = . , avhigh = . , p = . , f = , ); • cosine similarity using repeated n-grams (avlow = . , avhigh = . , p = . , f = , ). in experiment ii with three clusters, a feature with a small p-value was: • cosine similarity using repeated n-grams (avlow = . , avaver = . , avhigh = . , p = . , f = , ). as we can see, participants with a higher verbal intelligence used more words from the film and the similarity between their descriptions and the film was higher than the similarity of participants with an average and a lower verbal intelligence. this may be explained in the following way. test persons yielding a higher verbal intelligence (class high ) may have a better ability to listen to and recall spoken information from the film. mem- ory is indeed one of the verbal sub-tests of hawie so that a high memory score relates to a high verbal intelligence score of a test person. therefore, people with good memory (i.e. higher verbal intelligence) were easier able to remember many details of the film and to use words which they heard when watching the program. they may also better understand the relationships between language concepts, make more sophisticated language analogies or comparisons and perform a more complex language-based analysis. hence, we may conclude that the vocabulary of test persons yielding di"erent verbal intelligence was di"erent when they talked about the same event, even despite they were asked to talk about this film just after they had watched it. . text categorization solutions film derived features presented in the previous section showed to be good predictors of verbal intelligence. particularly, some of them suggested that test persons belonging to di"erent verbal intelligence classes may be distin- guished by word or lemma patterns, even regardless of the order of these words and lemmas in the monologues. this result led us to the main hypothesis that we investigate in this work: is it possible to solve the problem of inferring the corresponding level of verbal intelligence by simply applying conventional text categorization (tc) techniques? to validate this hypothesis, typical tf-idf features (introduced in the next section) have been extracted from the transcripts of the monologues (henceforth we do not make any use of the film transcription). three of the most popular tc methods have been applied for the auto- matic classification of monologues into three groups: test persons yielding a lower, an average and a higher verbal intelligence. . . tf-idf based approaches tf-idf (term frequency - inverse document frequency) based approaches are often used in information retrieval and text mining. as an example of a typical text mining task we may refer to the text categorization. the goal of tc is the classification of documents into a fixed number of predefined categories. the applicability of tc techniques has significantly grown in recent years. organizing news by subject topics (e.g. to disambiguate information and to provide readers with greater search experiences) or papers by research domains (e.g. for large databases of information that need indexing for re- trieval) are just some of the most popular examples. moreover, security (e.g. analysis of plain text sources such as internet news), biomedical (e.g. indexing of patient reports in health care organizations according to disease categories) or software (e.g. for tracking and monitoring terrorist activities) domains also have benefit from these techniques. new domains, like marketing (e.g. analytical customer relationship ma- nagement) or sentiment analysis (e.g. analysis of movie reviews), start using text mining solutions. in this work we have applied these techniques to a new domain: the estimation of speakers’ verbal intelligence. for tc, every document has to be transformed into a representation which could be suitable for learning algorithms and classification tasks. as reviewed in [ ], most tc algorithms are based on the vector space model (vsm). tc state-of-the-art systems widely apply the vsm approach [ , , ]. information retrieval (ir) research suggests that words work well as rep- resentation units. in vsm, each document in a corpus is represented by a list of words (i.e. bag of words). each word is considered as a feature; the value of the feature is a weight transformation of the number of times the word occurs in the document (i.e. word’s frequency). thus, a document is represented as a feature vector and its relevance to a query submitted by a user is measured through appropriate matching functions. these match- ing functions are typically based on statistical measures, like tf-idf, that basically weight the importance of each word. the importance of a word increases proportionally to its frequency within a document but is o"set by its frequency within a corpus. variations of this tf-idf weighting scheme are often used by search engines as a central tool in scoring and ranking a document’s relevance given a user query. . . . mathematical details tf-idf is a common feature transforming or weighting function. the term count, ni,j, denotes the frequency of a given term ti in a given docu- ment dj. this count is usually normalized to prevent a bias towards longer documents. thus, the term frequency tfi,j measures the importance of a term ti within a document dj and is defined as follows: tfi,j = ni,j! k nk,j ( ) where the denominator is the number of words in a document dj, that is, the size of the document |dj|. the inverse document frequency idfi is a measure of the general impor- tance of a term: idfi = log |d| {j : ti " dj} ( ) where |d| is the total number of documents in the corpus, {j : ti " dj} is the number of documents where the term ti appears (i.e. documents for ni,j #= ). the feature weighting function is then computed by using the following formula: tfidfi,j = tfi,j · idfi ( ) these weights show the importance of the words in each document. as can be seen, more frequent terms in a document are more representative and, if the number of documents in which this term occurs increases, this term becomes less discriminative. at this point, we may view each document as a vector that contains terms and their corresponding weights. for those terms from the vocabulary that do not occur in a document this weight equals to zero. in the following sections we will show the advantage of such a document representation. . . feature selection typical tc approaches make use of di"erent feature selection techniques to further reduce the dimensionality of the data space by removing irrelevant features that have no contribution to category discrimination. di"erent feature selection techniques through information theory were well studied in [ ]. as a result of this study, information gain (ig) and v -test (chi) were reported to be the top performing methods out of five methods under test in terms of feature removal aggressiveness and classifica- tion accuracy improvement. however, the document frequency thresholding approach, the simplest method with the lowest cost in computation, was reported to perform similarly. the document frequency (df) is the number of documents in which a term occurs. as described in [ ], it is possible to compute the document frequency for each unique term in the training corpus and to remove from the feature space those terms whose document frequency is less than a certain predefined threshold. by doing so we are adopting a basic assumption: rare terms are either non-informative for the category prediction (i.e. intelligence estimation in our case) or not influential in global performance. in either case, removal of rare terms contributes to the reduction of dimensionality of the feature space and improves the classification accuracy (i.e. if rare terms happen to be noise terms). if we try to summarize both pros and cons of using the document fre- quency thresholding approach, we may say that positive aspects are: • it is the simplest technique for vocabulary reduction (easily scalable to a very large corpora). • computational complexity is approximately linear with the number of documents. while on the other hand, negative aspects are: • the technique is usually considered as an ad-hoc approach to improve the e!ciency instead of a principled criterion for a predictive features selection. • the technique is typically considered, from an ir point of view, as a non-appropriate approach for aggressive term removal (low-df terms are assumed to be relatively informative and therefore should not be removed aggressively). in this work a slightly modified version of this df thresholding approach was applied to the data: tf-idf measures instead of df measures were used. as another remarkable di"erence, we did not remove the lowest tf-idf terms but just selected the highest tf-idf terms. in particular, instead of defining a threshold for tf-idf measures, we defined a fixed number of terms to be selected (i.e. n). therefore, we first sorted all the terms according to their tf-idf measures. then, we selected the top n most representative or indicative terms according to their tf-idf weights. the remaining terms were removed as stop or common words that did not add any meaningful content. by observing the evolution of the classification accuracy with an increasing n value, we determined the minimum size of the vocabulary (i.e. dimensionality) required to achieve the optimum performance. . . . class-based vs corpus-based as stated above, in our framework each word is considered as a feature and each document is represented as a feature vector. in [ ] two alternative ways for implementing the selection of these keywords or features are presented. in the first one, the so-called corpus-based keyword selection, a common keyword or feature set that reflects the most important words for all classes (i.e. highest tf-idf terms) in all documents is selected. in the alternative approach, named as class-based keyword selection, the keyword selection process is performed separately for each class. in this way, the most important and specific words for each class are determined. . . . word lemmatisation word lemmatisation is often applied in the area of ir, where the goal is to enhance the system performance and to reduce the number of unique words [ ]. particularly, word lemmatisation is part of the data pre-processing required to convert a natural language document to the feature space. for- mally, it is the process for reducing inflected (or sometimes derived) words to their lemmas. for example, as a result of lemmatisation, di"erent words like “play”, “plays”, “playing” and “played” are related to the same feature identification (i.e. lemma) “play”. word lemmatisation was applied to our monologues to assess its impact on performance (i.e. classification accuracy). like removing stop words, lemmatisation also contributed to the reduction of the size of the lexicon, thus saving on computational resources. . vector space classification as stated above, the vector space model represents each document as a vector with one real-valued component (i.e. tf-idf weight) for each term. therefore, we need text classification methods that can operate on real- valued vectors. in this section we introduce those ones that have been tested so far. a number of classifiers has been used to classify text documents, including regression models, bayesian probabilistic approaches, nearest neighbours approaches, rocchio algorithm, decision trees, inductive rule learning, neural networks, on-line learning, support vector machines (svms), and combining classifiers [ , , , ]. in this work we used three well-known vector space classification methods: naive bayes (nb), rocchio and nearest neighbour classification (knn). nb is often used as a baseline in text classification research as it com- bines e!ciency (training and classification can be accomplished with one pass over the data) and good accuracy (particularly if there are many equally im- portant features that jointly contribute to the classification decision). the rocchio algorithm is a very simple and e!cient text categorization method for applications such as web searching, on-line query, etc. because of its sim- plicity in both training and testing [ ]. knn requires no explicit training and can use the unprocessed training set directly in classification. however, it is less e!cient than the other classification methods (i.e. with knn all the work is done at run-time so that it can have poor run-time performance if the training set is large). rocchio and naive bayes are linear classifiers whereas knn is an example of a non-linear one. generally speaking, if a problem is non-linear and its class boundaries cannot be approximated well with linear hyperplanes, non- linear classifiers are often more accurate than linear classifiers (particularly, if the training set is large, then knn can handle complex classes better than rocchio and nb). on the other hand, if a problem is linear, then it is better to use a simpler linear classifier. however, this needs to be taken with a little bit of salt since the previous assertion is always conditioned by the well-known bias-variance trade-o" (i.e. with limited training data, a more constrained model tends to perform better). these approaches are described in more detail in the following sections. among the enumerated alternatives, svms are widely used mainly be- cause they have much current theoretical and empirical appeal and perform at the state-of-the-art level. according to [ ], svms, example based meth- ods, regression methods and boosting based combining classifiers deliver top- notch performance. lewis et al. ( ) also found that svms perform better on reuters-rcv corpus than knn and rocchio. nonetheless, recent revisions of the selected algorithms have proposed en- hanced versions of these methods that achieve relatively close performance to the top-notch tc classifier: svms. for instance, miao and kamel ( ) have re-examined the applicable assumptions and parameter optimization method of the traditional rocchio algorithm and proposed an enhanced ver- sion of this method that clearly outperforms the former one by using a pair- wise optimized strategy. salles et al. ( ) also presents a methodology to determine the impact that may have temporal e"ects on tc and to minimize it. by extending the three algorithms (namely knn, rocchio and nb) to in- corporate a temporal weighting function (twf), experiments showed that these temporally-aware classifiers achieved significant gains, outperforming (or at least matching) state-of-the-art algorithms. in any case, and as discussed in [ ], despite believes of many researchers that svm is better than knn in terms of e"ectiveness, knn is better than rocchio and rocchio is better than nb, the ranking of classifiers ultimately depends on the classes, the document collection and the experimental setup. . . naive bayes classification the first supervised learning method we introduce is the multinomial naive bayes or multinomial nb model, a probabilistic learning method [ ]. according to this method, the probability of a document d being in a class c can be computed as: p(c, d) $ p(c) · # !k!nd p(tk|c) ( ) where p(tk|c) is the conditional probability of a term tk occurring in a doc- ument of a class c. it may also be interpreted as a measure of how much evidence tk contributes that c is the correct class. p(c) is the prior probabil- ity of a document occurring in a class c. terms %t , t , · · · , tnd& are part of the vocabulary that is used for the classification; nd is the number of terms. in nb classification, the best class for a document d is determined as: cmap = arg max c"c $p(c|d) = arg max c"c $p(c) · # !k!nd $p(tk|c) ( ) where $p refers to the parameters to be estimated from the training data by applying the maximum likelihood estimation (mle). the interpretation of this equation is rather simple. each conditional parameter p(tk|c) is a weight that indicates the quality of an indicator tk for a class c. similarly, the prior p(c) is a weight that indicates the relative frequency of c. more frequent classes are more likely to be determined as the correct class. to reduce the number of parameters, we adopted the naive bayes con- ditional independence assumption where attribute values are independent of each other given the class so that for our multinomial model: p(d|c) = p (%t , t , · · · , tnd&|c) = # !k!nd p (xk = tk|c) ( ) where xk is a random variable for a position k in the document and the values of xk are terms from the vocabulary. p (xk = tk|c) is the probability that in a document of a class c a term t will occur in a position k. to further reduce the complexity of our multinomial model (assuming a di"erent probability distribution for each position k in the document still re- sults in too many parameters), we made a second independence assumption: conditional probabilities for a term are the same regardless of its position in a document: p (xk = t|c) = p (xk = t|c) ( ) where x is a single distribution of terms which is exactly the same for any po- sition k , k , · · · , ki. equation applies for all terms t and classes c. this po- sitional independence assumption is equivalent to adopting the bag of words model, which we introduced in section . . this bag-of-words model dis- cards all information that is communicated by the order of words in natural language sentences. . . . a variant of the multinomial model a critical step in solving a text classification problem is to choose the document representation. an alternative formalization of the multinomial model represents each document d as a m-dimensional vector of counts: tf'idft ,d, tf'idft ,d, · · · , tf'idftm ,d where tf'idfti,d is the tf-idf measure for a term ti in a document d. p(d|c) is then computed as follows: p(d|c) = p % %tf-idf t ,d, tf-idf t ,d, · · · , tf-idf tm ,d&|c & ( ) all the model parameters (i.e. class priors and feature probability distri- butions) may be estimated from the training set by using mle. for every class’ prior we calculated an estimate for the class probability from the train- ing set (i.e. (prior for a given class) = (number of samples in the class) / (total number of samples)). to estimate the parameters for our feature dis- tribution, we adopted the typical assumption that the continuous values as- sociated with each class are distributed according to a gaussian distribution. particularly, assuming that the training data contains continuous attributes, i.e. tf-idf measures for each term and document, we first segmented the data by the class and then computed the mean and variance of every term specific tf-idf measure in each class. . . . about the independence assumptions typically, tc tasks rather look at the words themselves and not at their corresponding positions in the documents (i.e. bag of words). this relies on the hypothesis that each topic or class to be distinguished is fairly represented by only some specific words from our bag. the nb models often perform well for tc tasks despite the conditional independence and the positional independence assumptions. in fact, both assumptions are very important to avoid problems in estimation owing to data sparseness. by adopting both independence assumptions, we are committed to a spe- cific way of processing the evidence. particularly, in the nb classification we look at each term separately so that we do not make a di"erence between word a followed by word b and word b followed by word a (although there is a di"erence between them). however, the conditional independence assump- tion does not really hold for text data as terms are conditionally dependent on each other. additionally, the position of a term in a document by itself may carry more information about the class than expected. . . the rocchio approach the rocchio classification [ ] divides the vector space into di"erent re- gions centred on prototypes. these prototypes or centroids, one for each class, define the class boundaries (i.e. hyperplanes). for a given training dataset, the centroid of a class c can be computed as the vector average or centre of mass of its members (i.e. all documents in the class) [ , ]. '(µ (c) = |dc| " d"dc '( v (d) ( ) where dc is the set of documents from class c: dc = ' d : %d, c& " d ( ; '( v (d) = v (d) · · ·vm(d) is a vector that contains tf-idf weights for each term of a document d. as many vector space classifiers (e.g. computing the nearest neighbours in knn classification), the rocchio approach relies on distance-based deci- sions (from a tc point of view, the relatedness of two documents can be typically expressed in terms of similarity or distance). particularly, the roc- chio classification rule is to classify a point in accordance with the region it falls into. to do this, basically we determine the centroid '(µ (c) that the point is closest to and then assign it to c. in our experiments, we used the cosine similarity measure as the un- derlying distance. cosine similarity is the cosine of the angle between two vectors and determines whether they are pointing in roughly the same direc- tion. since the components of our vectors (i.e. tf-idf weights) could not be negative, the angle between two tf-idf vectors could not be greater than #. the vector representation '( v (d ) and '( v (d ) of the cosine similarity be- tween two documents d and d is: sim (d , d ) = '( v (d ) · '( v (d ) | '( v (d )|| '( v (d )| ( ) where | '( v (d )| and | '( v (d )| it the euclidean length of the vectors. by using this measure, we are also applying a normalization process which makes each vector of the same length [ ]. if we have a look at the magni- tude of the vector di"erence between two vectors corresponding to documents with very similar content, it may happen that this di"erence is significantly simple because one is much longer than the other. cosine similarity measure compensates this e"ect of document length so that the similarity between document vectors is reduced to only measuring the cosine of the angle be- tween them. we can rewrite equation as follows: sim (d , d ) = '(v (d ) · '(v (d ) ( ) where '(v (d ) = '( v (d )/| '( v (d )| and '(v (d ) = '( v (d )/| '( v (d )|. the assign- ment criterion for a document d and its vector representation '( v (d) can be defined as: crocchio = arg max c"c sim ) '(µ (c), '( v (d) * ( ) in our implementation of the rocchio approach [ , ] only positive train- ing samples are considered for obtaining the prototype for each class (i.e. training samples that belong to the corresponding class). however, recent variations of rocchio [ , , ] consider the e"ects of negative samples (i.e. training documents that belong to all other classes) when computing the prototypes for the defined classes. di"erent parameters may be used to con- trol the relative importance of positive and negative samples. these rocchio classifiers reward not only the closeness of a test document to the centroid of the positive training instances, but also its distance from the centroid of the negative training instances. . . k-nearest neighbours in pattern recognition, the k-nearest neighbour algorithm (knn) is a method for classifying objects based on the closest training examples in the feature space. in tc, knn takes an arbitrary input document and ranks the k nearest neighbours among the training documents through the use of a similarity score (i.e. cosine similarity distance). it then assigns to the input the category or the class of the most similar document or documents. a constant k, defined by a user, denotes the number of neighbours included in the evaluation. the knn algorithm is a valid non-parametric method. despite being amongst the simplest of all machine learning algorithms, it is one of the best methods when the text is described by using vsm [ ]. however, traditional knn has two main drawbacks: the intensive computational e"ort, especially when the size of the training set grows (training examples are vectors in a highly multidimensional feature space), and its sensitiveness to the local structure of the data [ ]. new nearest neighbour algorithms have been recently proposed mainly with the purpose of reducing the number of distance evaluations actually performed, thus trying to make knn computationally tractable even for large data sets. for instance, [ ] presents a fast knn algorithm that reduces the cost of similarity computing in order to raise the classifying speed and applicability of knn. in naive bayes and rocchio classification we have to estimate correspond- ing parameters: priors and conditional probabilities and centroids. in knn we do not need to estimate any parameters but simply memorize all exam- ples in the training set and then compare a test document to them. for this reason, knn is also called memory-based learning or instance-based learning. the knn algorithm is known because of its strong consistency results. as the amount of data approaches infinity, the algorithm is guaranteed to yield an error rate no worse than twice the bayes error rate (the minimum achiev- able error rate given the distribution of the data) [ ]. knn is guaranteed to approach the bayes error rate for a certain value of k (where k increases as a function of the number of data points). the k-nearest neighbour methods may be improved by using proximity graphs [ ]. . . . choosing the class for an unclassified document to make a decision on a number of unclassified documents, we measure their similarity with all the documents that have already been classified. the unclassified documents are then ranked according to their similarity scores. appropriate classes for the documents may be assigned in the following ways: • if we choose k = , the class is predicted to be the class of the closest training sample. this is called the nearest neighbour algorithm. • if we choose k > , then all the documents which ranks are smaller than or equal to k will be included in the ranked list. we can then use di"erent means to find a class for our document, like: – we may assign the document to the most common class amongst its k nearest neighbours (if we are dealing with a binary, i.e. two- class, classification problem, it is helpful to choose k to be an odd number as this avoids tied votes). – we may estimate the probability of membership in a class c as the proportion of the k nearest neighbours in c. this is commonly referred as the probabilistic version of the knn classification al- gorithm. – for the individual classes, we may sum the distances to all the doc- uments in which the class occurs, and then choose the class corre- sponding to the highest accumulated distance (remember that we are using cosine distance). – etc. if we decide to use either the basic “majority voting”, the probabilistic method or the sum of distances based classification, those classes with more frequent examples will tend to dominate the prediction of a new vector. this is actually a drawback as they tend to come up in the k nearest neighbours when the neighbours are computed due to their large number (it is important to remind that the available data is certainly imbalanced). to overcome this problem, we compensate this possible imbalance by introducing a slightly modified classification method for our knn based approach. typically, the implementation of these versions of the algorithm starts by computing the distances from the test sample to all stored vectors of the training data set. next, all these training samples are sorted according to these distances thus ranking the nearest k training samples regardless of their corresponding class. if we look at those already labelled classes instead (neighbours are taken from a set of documents for which the correct classification is known), we could then identify specific top k neighbours for each class c (i.e. k nearest neighbours labelled as c, thus resulting in an overall list composed of k ) c elements). finally, by computing distance to each class as the average distance between the test sample and those top k class specific neighbours, we will then manage to compensate any possible imbalance in the distribution of the training data among the defined classes. the best class for knn classification can then be derived from: cknn = arg max c"c score(c, d) = arg max c"c k · " d!"sk(c,d) sim('(v (d$), '(v (d)) ( ) where sk(c, d) is the c class specific set of d’s k nearest neighbours. as could be derived from equation , it may also be useful to weight the contributions of the neighbours so that nearer ones contribute more to the average than more distant ones [ ]. this classification method is weighted by taking into account not only the distance from the test sample to the c set of k nearest neighbours, but also the class compactness particularly for high values of k. . . . parameter selection the parameter k in knn is typically defined by using some previous experience or specific knowledge about the classification domain. normally, nn is found to be not very robust. the accuracy of the knn algorithm can be severely degraded by the presence of noisy or irrelevant features (also if the feature scales are not consistent with their importance). nn means that the classification decision for each test document only relies on the class of a single training document, whose label could eventually be incorrect or atypical. knn for k > are more robust as larger values of k tend to reduce the e"ect of noise on the classification (although also make boundaries between classes less distinct). as an alternative, a good value of k can be assigned heuristically via cross validation technique or empirically via bootstrap method [ ]. in our experiments, instead of applying any of these parameter selection methods, we tried di"erent k values, thus finally selecting the optimal k as the value which was used when obtaining the best performance. . experimental results and discussion . . experimental set-up our main goal is to identify the algorithm that best computes class bound- aries and reaches the highest classification accuracy. in our experiments for comparing the performance of the di"erent approaches, a leave-one-out cross validation (loo-cv) method was used. the idea of this method is to use n- observations for training (where n is the number of data points) and only data point for testing. this procedure is repeated n times and each observation is used once as the testing data. . . baseline approach: class-based vs corpus-based feature selection as introduced in section . . , our experiments covered the comparison of the class-based and corpus-based keyword selection approaches. the corpus-based approach implies the selection of a common feature set for all classes with the top n most representative or indicative terms. the class-based approach instead implies the selection of the most important words for each particular class. in this case, to preserve the balance between classes, n/m words for each specific class were selected, where m is the number of classes. for our classification task m equals to , where the first class contained test persons yielding a lower verbal intelligence, the second class contained participants yielding an average verbal intelligence and the third class contained participants yielding a higher verbal intelligence. then, we composed our feature vector by concatenating all the class-specific features, thus resulting into a vector comparable to the n-dimensional vector corresponding to the corpus-based approach. however, when using the class-based approach, a particular word may be included in various class-specific subsets (i.e. a word that is important to not only one single class but to several classes). to avoid using duplicate features, we only used the intersection between all the class-specific subsets. therefore, the dimension of the resulting feature vector in these cases had to be necessarily lower than n. for simplicity, we will better refer to the number of features per class (i.e. f = n/ ) rather than to the final dimensions of the vectors. consequently, if we report, for instance, about words or features per class, this means that we are using a -dimensional corpus- based vector. in this case for the class-based approach, is the maximum number of dimensions. to definitely determine its value, it is necessary to check the possible intersection. of course, the higher the value of f , the more significant the intersection between class-specific word subsets, and also the bigger the di"erence with respect to the corpus-based vector dimensions. analysing the corpus, di"erent words were extracted from all the monologue transcripts. table shows how the intersection evolved according to f . considering the size of the vocabulary, the observed di"erence is significant. table : dimension di!erences between class-based and corpus-based approaches. # of features per class ( classes) corpus-based class-based di!erence rel. di!. (%) , , , , , , , , figure presents the accuracy results obtained using either the corpus- based or the class-based feature selection methods. the results were obtained using the nb approach for di"erent dimensions of the feature vector. confi- dence intervals of % are also shown in the figure. as it can be derived from the figure, the class-based approach clearly out- performed the corpus-based one regardless of any di"erence about the used dimension. although the observed di"erences were not statistically signifi- cant in any case, it is interesting to pinpoint the result for the -dimensional value. at this point the class-based approach reached the top performance while the di"erence with the corpus-based alternative also turned to be the biggest one thus becoming almost significant. from a di"erent point of view, we may also try to analyse the min- imum dimensionality required by the class-based approach to outperform the corpus-based one. the corpus-based approach obtained a maximum ac- curacy of , %. as can be observed in figure , this performance was reached with dimensionality equal to or higher than . also derived from this figure, we may check that the class-based approach obtained a better performance of , % (though not statistically significant) using “only” features per class. the class-based feature selection, by definition, focuses on finding the most crucial or indicative class keywords. on the other hand, figure : baseline approach: class-based vs corpus-based feature selection. the corpus-based one simply tends to find general keywords concerning all classes. this clearly tips the balance in favour of the class-based approach particularly when we use a reduced set of features. this is important as there may be a significant gain in classification time when a small number of features is used. by confirming these di"erences with additional statistical evidence (i.e. more data), we may also conclude that the class-based feature selection im- proved the performance of the corpus-based one for the nb approach not only in terms of accuracy but also in terms of time. similar results were already confirmed in [ ]. when using the corpus-based approach, most features (i.e. words) tend to be selected from the prevailing classes so that rare classes are not well rep- resented. in contrast, when using the class-based approach all the classes are represented equally well as for their representation class specific features are used. thus, the class-based approach achieved consistently higher accuracies than the corpus-based approach. similar di"erences between the class-based and corpus-based methods have been consistently observed throughout all of our experiments. there- fore, in the next sections we will only focus on the class-based versions. . . comparison between approaches: rocchio “wins” in this section we compare the results that were obtained using di"erent approaches. before proceeding with this comparison, we need first to assign the optimal configuration (i.e. k value) for the knn approach. figure presents classification results corresponding to several k values. as expected, nn was found to be not very robust. optimal performance may be reached by using k = in combination with dimensionality of . however, if we keep increasing the value of k, which is typically more robust as it helps to reduce the e"ect of noise on the classification, then the results apparently start to be a"ected by sparse data bias. figure : knn results for di!erent k values. as a result of the initial k-means clustering, only samples were defined to be part of the least populated class. therefore, starting with nn we checked out up to k = values leaving one sample out for testing (the loo approach was applied). for clarity, figure presents classification results only with some values of k. the observed di"erences were found to be statistically significant for the top performance dimensionality (i.e. f = ) when comparing the best configuration (i.e. k = ) with all the others for k > . no statistically significant di"erences where observed between the best k and any k * configurations. figure allows to compare the results of the nb approach, the rocchio approach and the knn approach with k = . figure : comparison between approaches: rocchio wins. a first important result that we can derive from figure is that both roc- chio and knn are clearly outperforming the nb approach, although the top performance is defined for di"erent dimensionalities in each case. the knn performance had a maximum accuracy of , % for -dimensionality, while rocchio just required features per class to improve it up to , %. both results denoted a statistically significant di"erence when compared to the nb top performance, , % also for -dimensionality. however, we did not observe any significant di"erence between rocchio and nn (natu- rally, rocchio was also significantly outperforming any k > approach). as it typically occurs in tc tasks, most of the learning takes place with a small yet crucial portion of features (i.e. keywords) for a class. this is evident in the steeper learning curves that reach the top performance at relatively low dimensionality. therefore, we may conclude that the class- based feature selection approach is shown to be successful in quickly finding the most crucial or indicative class keywords. another visible result in figure , common to all the tested approaches, is the performance decrease as the value of f increases (particularly beyond a -dimensional value). as we already introduced in section . and proved in table , the higher the value of f , the more significant the intersection between class-specific word subsets. if we expand this interpretation, the more significant the intersection, the less discriminative the class-specific subsets, the more likely to include words that are not really indicative of any of the classes, and so the performance decreases. . . using words vs lemmas as we introduced in section . . , we also tried a word lemmatisation strategy (i.e. to group together those words that are in di"erent forms but with the same lemma). this strategy was implemented as part of the data pre-processing stage during the classification task. figure shows the results with and without word lemmatisation for our top performing approach: the rocchio one. the main advantage of word lemmatisation is to reduce the dimension- ality of the data space. in a tc task, it is basically applied under the assumption that all the documents belonging to the same category or topic may include these lemmas appearing in di"erent forms, and of course, it makes sense to use them as they refer to words with similar meanings. tc tasks typically rely on this. however, to be successful and thus really en- hance system performance, there is another important hypothesis that also needs to be confirmed: each topic or class to be distinguished should be fairly represented by only some class-specific lemmas. while the former one happens to be true for most of the cases, the latter one, though also successfully applied in typical tc tasks, may reasonably not be true in our case. the main reason for this would be that, from this point of view, all the documents (i.e. monologues) could be regarded as belonging to the same category according to their topic or content: all the documents are about the film which the participants watched. consequently, we could figure : classification results using words vs lemmas. expect an important number of lemmas to be shared among the participants as they all were talking about the same topic. this is an important di"erence with conventional tc tasks where, nor- mally, the topics or classes are well separated according to their conceptu- alization. in contrast, in our domain we may expect the participants to be identifiable among others not by the concepts or ideas themselves but by the way they express these ideas. therefore, in this particular case, we may expect lemmas not to have much contribution to category discrimination but the di"erent endings and forms instead. hence, missing this discriminative information because of lemmatisation (simplifying words with di"erent forms into their more common roots) could have some undesirable consequences in classification and clustering. moreover, the fact that all the participants were german native speakers could be particularly critical for this problem. in this regard it is important to remark that german is a very agglutinative language [ ]. compound words or words that consist of more than one lemma (i.e. compounding or word- compounding occurs when a person attaches two or more words together to make one word), can be found very often in the german language. “donau- dampfschi"fahrtsgesellschaftskapitänsmütze” (i.e. danube steamboat ship- ping company captain’s hat) is a good example of how long these compound words could be (they can be practically unlimited in length, particularly in case of biochemistry). the meaning of a compound word di"ers from the meanings of words which it consists of. lemmatisation of compound words would simply re- duce them to their more common lemmas thus loosing this discriminative information. to what extent this argument could be either true or false is something that can be derived from figure . in fact, the word-based approaches sys- tematically outperform the lemma-based ones. confidence boundaries for both cases are also shown in this figure. as we can observe, di"erences be- come statistically significant mostly around the same dimensionality range that was previously pinpointed when referring to the top performance for both knn and nb (particularly at a -dimensional value). however, the di"erences are not statistically significant at f = , the point at which roc- chio reaches its maximum accuracy for both word-based and lemma-based approaches. . . tempted to use more classes although the three-classes scheme can be found entirely suitable from a practical implementation point of view (i.e. participants yielding a lower, an average and a higher verbal intelligence), we were also interested in analysing the performance of the suggested approaches for a higher number of classes. this would enable a better granularity for the verbal intelligence classifica- tion. figure presents benchmarking results for classes instead of (as it was shown in figure ). from a practical point of view, these classes may correspond to the following levels of verbal intelligence: poor, average, high and very high respectively. in this regard it seems to be important to remark that working with a higher number of classes, like or more, was practically infeasible because of sparse data problems (i.e. k-means resulted into unpopulated classes). as for classes, the rocchio approach showed the highest accuracy again (i.e. , % at f = ). the optimal dimensionality remained to be the same as for classes (i.e. f = ). regarding the comparison between rocchio and knn, the observed di"erences also remained to be not statistically significant. figure : tempted to use more classes: classes. additionally, both rocchio and knn clearly outperformed the nb algo- rithm, once again by a significant margin. for these two, another e"ect starts to become evident: the top performance region, previously observed for di- mensionality values up to , now turns to be narrower, locating its limit approximately around a value of . as we simply increase the number of classes, it seems to be evident that the number of terms or features that are really indicative of each particular class becomes smaller, thus a"ecting the performance. from a general point of view, the resulting performance can still be deemed to be satisfactory as the error rate is only roughly % higher than with three classes. if we look at the confusion matrix, presented in table , we may check that by adopting the four classes into three by grouping high and very high classes, predictability would be more similar reducing the gap roughly to the half (i.e. , % for three classes and , % for four classes adopted into three). finally, it may be interesting, particularly from a practical point of view, to have a look at the upper-right and lower-left corners of the matrix: table : confusion matrix corresponding to rocchio’s top performance using classes (f = ). prediction outcome a b c d actual value a = poor b = average c = high d = very high errors. this means there was not any critical errors like regarding a lower verbal intelligence individual as a higher verbal intelligence one and vice- versa. . conclusions and future work this work showed that verbal intelligence may be recognized by com- puters through language cues. the achieved classification accuracy can be deemed as satisfying for a number of classes that is reasonably high enough to enable its integration into sldss. to our knowledge, this is the first report of experiments attempting to automatically predict verbal intelligence. some of the most popular tc algorithms were applied to this task: nb, rocchio and knn. nb models are typically expected to perform well for tc tasks despite the conditional independence and the positional independence assumptions. however, the performance of nb approach was significantly worse than with the other approaches: knn and rocchio. this suggests that this probabilistic classifier was more sensitive to the low number of examples available, mainly resulting into inaccurate probability estimates, than the vector space ones (computing distances to some relevant members or to a prototype of each defined class seems to be more robust against sparse data). on the other hand, and connecting with those independence assumptions, it is well known that conditional independence does not really hold for text data (even worse considering that our features are highly correlated). fur- thermore, we firmly believe that, for this specific task, the position of a term in a document by itself could carry more information about the class than expected, mainly because of the above mentioned peculiarities of our classi- fication task (i.e. it is not only about the words that participants used to denote their intelligence, but also the way they combined them). therefore, our data is somehow violating these independence assumptions, thus finally explaining why the nb approach performed so poorly. in this regard, it would be very interesting to test a lm based tc approach to better validate this argument. using the class-based feature selection approaches has proven to be an essential factor, not only to achieve a better inference performance but also to reduce its computational cost. despite typically successful when applied to tc tasks, word lemmatisa- tion was not really helpful for our task. the word-based approaches system- atically outperformed the lemma-based approaches, thus pointing out some peculiarities of the classification task. particularly, these results were found to be mainly explained by two di"erent factors: the same topic for collecting monologues and the use of the german language. unlike typical tc tasks, our verbal intelligence prediction task is influ- enced by the necessary fact that the di"erent categories or classes to be identified are not well separated from a conceptualization point of view. of course, it might be easier to distinguish people talking about di"erent topics from their everyday life although the results for such a comparison might not be objective. by letting the participants (i.e. people with di"erent interests and hobbies) to discuss their own topics, we would be then recognizing the topics themselves rather than people with di"erent cognitive processes. on the other hand, the use of german, a very agglutinative language, resulted to be a drawback with regards to word lemmatisation. by lemma- tisation of compound words (compounding is a pretty common phenomena in german) we are basically loosing the extra meaning that arises from the combination of the interrelated words. this meaning has proven to be really helpful to correctly discriminate between di"erent levels of verbal intelligence. in future work, it would also be interesting to examine how well the suggested approaches perform when integrated into existing slds. in this regard, it is important to remark that any application involving speech recog- nition will always introduce noise in the features that we used. this needs to be considered as it will surely reduce the presented accuracies. testing these approaches with conventional slds would allow us to assess whether the accuracies we achieve are high enough or not for our intended application (i.e. dialogue system adaptation). on the other hand, this also suggests the importance of finding some other features that could be more robust when being used in a conventional system. prosodic features could be a good alternative; so it would be interesting to start working on an multimodal inference framework that could jointly exploit the potential of, among others, this kind of features. as we have already mentioned, the linguistic cues that we have used in this work could pose a problem, for instance, if we want to apply these solutions with the same users but across di"erent domains. in this regard, prosodic features would be found to be advantageous as they would also allow us to explore the possibility of finding topic independent solutions. acknowledgement this work is partly supported by the daad (german academic ex- change service). parts of the research described in this article are supported by the tran- sregional collaborative research centre sfb/trr ”companion-technology for cognitive technical systems” funded by the german research founda- tion (dfg). for this work, fernando was granted a fellowship by the caja madrid foundation. references [ ] baeza-yates, r. a., ribeiro-neto, b., . modern information re- trieval. addison-wesley longman publishing co., inc., boston, ma, usa. [ ] bi, y., bell, d., wang, h., guo, g., guan, j., march . combining multiple classifiers using dempster’s rule for text categorization. appl. artif. intell. , – . url http://dl.acm.org/citation.cfm?id= . [ ] cianciolo, a. t., sternberg, t. j., . intelligence: a brief history. blackwell publishing. [ ] dumais, s., platt, j., heckerman, d., sahami, m., . inductive learn- ing algorithms and representations for text categorization. in: proceed- ings of the seventh international conference on information and knowl- edge management. cikm ’ . acm, new york, ny, usa, pp. – . url http://doi.acm.org/ . / . [ ] goethals, g., sorenson, g., burns, j., . encyclopedia of leadership. no. v. in encyclopedia of leadership. sage publications. url http://books.google.es/books?id=kjlspnszs uc [ ] hall, p., park, b. u., samworth, r. j., . choice of neighbor order in nearest-neighbor classification. annals of statistics , . url doi: . / -aos [ ] hui, g. g., wang, h., bell, d., bi, y., greer, k., . using knn model- based approach for automatic text. in: in proc. of odbase’ , the nd international conference on ontologies, database and applications of semantics, lncs. pp. – . [ ] ittner, d. j., lewis, d. d., ahn, d. d., . text categorization of low quality images. in: in proceedings of sdair- , th annual symposium on document analysis and information retrieval. pp. – . [ ] jianliang, y., yongcheng, w., . application of iterative-knn based on knn and automatic retrieval in automatic categorization. journal of the china society for scientific and technical information , – . [ ] joachims, t., . text categorization with support vector machines: learning with many relevant features. in: european conference on ma- chine learning (ecml). springer, berlin, pp. – . [ ] kupietz, m., belica, c., keibe, h., witt, a., . the german ref- erence corpus dereko: a primordial sample for linguistic research in: calzolari, nicoletta et al. (eds.). in: proceedings of the th conference on international language resources and evaluation (lrec ). pp. – . [ ] lewis, d. d., yang, y., rose, t. g., li, f., december . rcv : a new benchmark collection for text categorization research. j. mach. learn. res. , – . url http://dl.acm.org/citation.cfm?id= . [ ] logsdon, a., . learning disabilities. url http://www.learningdisabilities.about.com/ [ ] manning, c. d., raghavan, p., schtze, h., . introduction to infor- mation retrieval. cambridge university press, new york, ny, usa. [ ] mergenthaler, e., . emotion-abstraction patterns in verbatim pro- tocols: a new way of describing psychotherapeutic processes. journal of consulting and clinical psychology ( ). [ ] miao, y.-q., kamel, m., january . pairwise optimized rocchio al- gorithm for text categorization. pattern recogn. lett. , – . url http://dx.doi.org/ . /j.patrec. . . [ ] moschitti, a., . a study on optimal parameter tuning for rocchio text classifier. in: proceedings of the th european conference on ir research. ecir’ . springer-verlag, berlin, heidelberg, pp. – . url http://dl.acm.org/citation.cfm?id= . [ ] olsen, s., . ein internationales handbuch zur flexion und wortbil- dung. in: booij, g., lehmann, c., mugdan, j. (eds.), morphologie. berlin / new york: de gruyter, pp. – . [ ] özgür, a., özgür, l., güngör, t., . text categorization with class- based and corpus-based keyword selection. in: iscis. pp. – . [ ] rocchio, j., . relevance feedback in information retrieval. prentice-hall inc., ch. , pp. – . [ ] salton, g., . automatic text processing: the transformation, analy- sis, and retrieval of information by computer. addison-wesley longman publishing co., inc., boston, ma, usa. [ ] sebastiani, f., march . machine learning in automated text catego- rization. acm comput. surv. , – . url http://doi.acm.org/ . / . [ ] solka, j., jul . text data mining: theory and meth- ods. statistics surveys , vol. , - comments: pub- lished in at http://dx.doi.org/ . / -ss the statistics surveys (http://www.i-journals.org/ss/) by the institute of mathematical statis- tics (http://www.imstat.org). [ ] tan, s., may . neighbor-weighted k-nearest neighbor for unbalanced text corpus. expert syst. appl. , – . url http://dx.doi.org/ . /j.eswa. . . [ ] toussaint, g. t., . geometric proximity graphs for improving near- est neighbor methods in instance-based learning and data mining. int. j. comput. geometry appl. ( ), – . [ ] vinciarelli, a., october . application of information retrieval tech- niques to single writer documents. pattern recogn. lett. , – . url http://dx.doi.org/ . /j.patrec. . . [ ] wang, y., wang, z.-o., aug. . a fast knn algorithm for text cat- egorization. in: machine learning and cybernetics, international conference on. vol. . pp. – . [ ] wechsler, d., . the measurement of adult intelligence. baltimore (md): williams & witkins. [ ] wechsler, d., . handanweisung zum hamburg-wechsler- intelligenztest fuer erwachsene (hawie). separatdr., bern; stuttgart; wien, huber. [ ] yang, y., liu, x., . a re-examination of text categorization meth- ods. in: proceedings of the nd annual international acm sigir con- ference on research and development in information retrieval. sigir ’ . acm, new york, ny, usa, pp. – . url http://doi.acm.org/ . / . [ ] yang, y., pedersen, j. o., . a comparative study on feature se- lection in text categorization. in: fisher, d. h. (ed.), proceedings of icml- , th international conference on machine learning. morgan kaufmann publishers, san francisco, us, nashville, us, pp. – . url citeseer.nj.nec.com/yang comparative.html [ ] zablotskaya, k., walter, s., minker, w., may . speech data cor- pus for verbal intelligence estimation. in: proceedings of lrec’ . pp. – . response modeling with support vector regression dongil kim, hyoung-joo lee and sungzoon cho ∗ department of industrial engineering, seoul national university, san - , shillim-dong, kwanak-gu, seoul, - , korea abstract response modeling, which predicts whether each customer will respond or how much each customer will spend based on the database of customers, becomes a key factor of direct marketing. in previous researches, several classification approaches, include support vector machines (svm) and neural networks (nn), have been applied for response modeling. however, there are two drawbacks of conventional approaches: ( ) response models only predict classification scores rather than predicting total amount of money spent, ( ) too large training data. for the first drawback, we applied support vector regression (svr) for response modeling to predict total amount of money spent of each respondent. for the second drawback, we employed a pattern selection method designed for svr. this paper provides experimental results of a direct marketing dataset in terms of model fit, training time complexity and profitability. key words: response modeling, direct marketing, support vector machines; regression; pattern selection ∗ corresponding author. tel: + - - - , fax: + - - - email addresses: dikim @snu.ac.kr (dongil kim), imhjlee@gmail.com preprint submitted to elsevier science may introduction a response model, given a mailing campaign, predicts whether each customer will respond or how much each customer will spend based on the database of customers’ demographic information and/or purchase history. marketers will send mails or catalogs to customers who are predicted to respond or to spend large amounts of money. a well-targeted mail increases the profit, while a mistargeted or unwanted mail not only increases the marketing cost but also may worsen the customer’s relationship to the firm (gönül et al., ; potharst et al., ). various methods have been used for response modeling such as statistical techniques (bentz and merunka, ; haughton and oulabi, ; ling and li, ; suh et al., ), machine learning techniques, (wang et al., ; chiu, ; cheung et al., ; shin and cho, ; viaene el al., ; yu and cho, ) and neural networks (nn) (potharst et al., ; bentz and merunka, ; zahavi and levin, ). so far, a response model have been usually formulated as a binary classification problem because of its straightforwardness. the customers are divided into two classes, respondents and non-respondents. a classifier is constructed to predict whether a given customer will respond or not. from a modeling point of view, however, as pointed out in (kdd cup, ) for the kdd-cup- task, there is an inverse correlation between the likelihood to buy and the dollar amount to spend (wang et al., ). this is because the more dollar amount is involved, the more cautious the customer is in making a purchase decision. in this sense, the likelihood to buy that is estimated by a classification model may not lead to more profit. therefore, in addition to the classification approach, a regression model needs to be applied for response modeling to (hyoung-joo lee), zoon@snu.ac.kr (sungzoon cho). predict total amount of money of each customer. support vector machine (svm) is known as the most spot-lighted algorithm with great generalization performances by employing structural risk mini- mization (srm) principle (vapnik, ). support vector regression (svr), a regression version of svm was developed to estimate regression functions (druker et al., ). like svm, svr is capable of solving non-linear prob- lems using kernel functions and successful in various domains (druker et al., ; müller et al., ). however, there was a difficulty to train svr on real-world dataset. as the number of training patterns increases, svr train- ing takes much longer with a time complexity of o(n ) where n denotes the number of training patterns. so far, many algorithms such as chunking, smo, svmlight and sor have been proposed to reduce the training time. however, their training time complexity is still strongly related to the number of train- ing patterns (platt, ). we take another direction called pattern selection which is focusing on reducing the number of training patterns. npps (shin and cho, ) is a proved method of pattern selection for svm, but it’s only for svm classifiers. rather we use a pattern selection method based on ε-tube which is especially designed for svr (kim and cho, ). in this paper, we applied svr for response modeling to predict total amount of money spent of each respondent. as mentioned before, a single response model based on classification method can score the likely to respond of each customer. however, direct marketers would like to know not only respondents but also profitable customers who will spend more money than others. hence, after predicting respondents by a classification response model, a regression model is needed to predict total amount of money spent of each respondent. as the classification model was not our concerned, we supposed there was a primitive ideal classifier which could find all respondents without false posi- fig. . the ideal procedure of the paper. especially, the paper is focused on the two dark boxes. tive (fp) errors. we made a subset consists of only but all respondents from original dataset. svr model was applied to the subset of respondents. the ideal procedure of this paper is presented in fig . we used the dmef dataset from the direct marketing educational foundation (dmef) which is collected from a catalog mailing task. the remaining of this paper is organized as follows. in section , we introduce the concepts of svr and provide the main idea of the pattern selection method with a simple toy example. in section , we present details of dmef datasets and parameters for experiment. in section , experiment results are following. in section , we summarize the result and conclude the paper with a remark on limitations and future research directions. pattern selection for support vector regression . support vector regression for a brief review of svr, consider a regression function f(x) to be estimated with training patterns {(xi, yi)}, f (x) = w · x + b with w, x ∈ rn , b ∈ r ( ) where {(x , y ), · · · , (xn, yn)} ∈ rn × r ( ) svr is moved around to include training patterns inside ε-insensitive tube (ε-tube). by the srm principle, the generalization accuracy is optimized by the flatness of the regression function. since the flatness is guaranteed on a small w, svr is moved to minimize the norm, ‖w‖ . an optimization problem could be formulated with constraints where c, ε, and ξ, ξ∗ are trade-off cost between empirical error and the flatness, size of ε-tube and slack variables, respectively, for the following soft margin problem. minimize ‖w‖ + c n∑ i= (ξi + ξ ∗ i ), ( ) subject to yi − w · xi − b ≤ ε + ξi, w · xi + b − yi ≤ ε + ξ∗i , ξi, ξ ∗ i ≥ , i = , · · · , n. hence, svr is trained by minimizing ‖w‖ with including training patterns inside the ε-tube. with adding lagrangian multipliers α and α∗, the qp prob- lem can be optimized as dual problem. the estimated regression function from svr is following where ns is the number of support vectors: f (x) = ns∑ i= (α − α∗)k(xi, x) + b ( ) . pattern selection for support vector regression the training time complexity of svr is o(n ). if the number of training patterns increases, the training time increases more radically, i.e. in a cubic proportion. marketing databases usually consists of over one millions of cus- tomers and hundreds of input variables. hence, it takes too long time to train svr directly to marketing dataset. we applied a pattern selection method from our previous research (kim and cho, ). fig. . (a) the regression function after training original dataset, and (b) the re- gression function after training only patterns inside estimated ε–tube. svr trains patterns based on ε-loss function foundation. svr makes ε-tube on the training patterns. the patterns in ε-tube are not counted as error, and patterns out of ε-tube, i.e. support vectors (svs), are used for training. in addition, svr estimates the regression function as the center-line of ε-tube. hence, if ε-tube can be estimated before training, we can find the regression function with only those patterns inside ε-tube (see fig. ) . however, remov- ing all patterns outside ε-tube could lead to reduction of ε-tube itself, thus it is desirable to keep some of “outside” patterns for training. hence, we defined a “fitness” probability for each pattern based on its location with respect to ε-tube and then selected patterns stochastically. we made k bootstrap samples of size l (l < n) from original training pattern set (d). we trained an svr with each bootstrap sample and obtained k svr regression functions. each regression function was used to see if a training pattern is located inside ε-tube. each training pattern in d is located inside a minimum of zero ε-tubes to a maximum of k ε-tubes. let mj denote the number of times that pattern j is found in-side an ε-tube. we use mj as the likelihood that pattern j is actually located inside the real ε-tube. each mj is converted to a probability, pj as in eq. ( ). since we want to select patterns fig. . (a) original dataset and an svr trained on it, (b) a bootstrap sample and an svr trained on it, (c) original dataset and ε-tube of (b)’s svr, and (d) selected patterns and an svr trained on them inside ε-tube, pattern j is selected with a probability of pj, pj = mj∑n i= mj . ( ) the procedure of the pattern selection method is presented in fig. with a simple toy example. the algorithm is presented in fig. . . initialize the number of bootstrap samples, k initialize the number of patterns in each bootstrap sample, l initialize the number of patterns to be selected, s . make k bootstrap samples, di, (i = , · · · , k), from the origi- nal dataset d by random sampling without replacement . train svr fi with di, (i = , · · · , k) . count the number of times mj that pattern j is found inside ε-tube of fi . convert mj to pj according to eq. ( ) . select s patterns stochastically from d without replacement based on pj . train final svr with s selected patterns fig. . ε-tube based pattern selection algorithm dataset and experimental settings . dataset: dmef dataset we utilized dmef dataset which contains , customers and input variables. the response rate is . % with , respondents and , non- respondents. we selected a subset based on the weighted dollars used from previous researches (yu and cho, ; ha et al., ). we would like to design a regression model as a scoring model after picking up respondents with a primary classification model. since a classification model was not our table input variables name formulation description original variables purseas number of seasons with a purchase falord ltd fall orders ordtyr number of orders this year puryear number of years with a purchase sprord ltd spring orders derived variables derived variables recency order days since / tran i( ≤ recency ≤ ) tran i( ≤ recency ≤ ) tran i( ≤ recency ≤ ) tran /recency comb ∑ m= prodgrpm number of product groups purchased from this year tran √ comb tran log( + ordtyr × falord) interaction between the number of orders tran √ ordhist × sprord interaction between ltd orders and ltd spring orders tran /( +lorditm) inverse of latest-season items interest in this research, we assumed that there was an ideal response model built with a classification algorithm that could pick all respondent without false acceptances. hence, we selected a new dataset consists of customers only whose target dollars are positive, i.e. only respondents. the final selected dataset consists of , customers. for performance evaluation, the dataset was partitioned into training and test sets. a half of customers were randomly assigned to the training set while the other half to the test set. performance of a model shows a large varia- tion with regard to a specific data split (malthouse, ). so ten different training/test splits were generated. all results are averaged of each exclusive training/tes sets. we are not interested in feature selection/extraction. malt- house extracted input variables for this dataset and ha et al (ha et al., ) used out of them, removing two variables whose variations are neg- ligible. in this paper, these variables were used as input variables as listed in table . we formulated this dataset to a regression problem by setting the total amount of dollars as target variable. . experimental settings we made three different response models based on svr in the experiments. all experimental results of svr with pattern selection (svr-ps) were compared with results of svr with all data (svr- ) and svr with random sampling (svr-random). we set hyper-parameters of svr. the hyper-parameters of svr were deter- mined by cross–validation for svr- with c × ε = { . , , , , , }× { . , . , . , . , . , . , }. rbf kernel was used as a kernel function and the kernel parameter σ was fixed to . for all datasets. the parameters of the pattern selection method were set as follows. in this case, the number of boot- strap samples k was set to . the other parameter that controls the number of patterns in a bootstrap sample l was set to % of the number of patterns in dataset, n. the number of selected patterns was set to % to % of n. during the experiments, we used all data as normalized vectors. root mean squared error (rmse) is used to estimate the model fits. we calculated the averaged profits per one catalog mail to estimate the profitability of response models. also we recorded training times. all stochastic results were the aver- aged values of -time repeats. experimental results . model fit and time complexity fig. shows the experimental results. the solid lines present the result of svr with all data (svr- ). the lines with circles presents the results of svr with pattern selection (svr-ps) while the lines with boxes presents the results of svr with random sampling (svr-random). fig. (a) is rmses which have real values (dollars). it shows svr-ps is slightly less accurate than svr- , but is significantly more accurate than svr-random. as more patterns were selected, the training dataset became similar to the original dataset and the gaps between all svrs became narrower. with % pattern selection, the gap of rmses between svr-ps and svr-random is larger than dollars. when the percentage of selected patterns is % of all, the gaps of rmses between svr-ps and svr- are less than dollars. fig. (b) shows averaged training time of all svrs. it takes seconds to train all data with svr. however, with % of pattern selection, svr-ps needed only % of training time of svr- . although the percentage of selected patterns was up to %, which resulted only dollars’ gap, svr-ps needed only % of training time of svr- . svr-ps is more efficient than svr- . percentage of patterns selected (%) r m s e svr− svr−ps svr−random (a) percentage of patterns selected (%) t im e (s ec ) svr− svr−ps svr−random (b) fig. . (a) rmse resulted from experiments and (b) the training times . . . . . . . . . mailing depth (%) a ve ra ge p ro fi t p er m ai l ($ ) svr− svr−ps svr−random . . . . . . . . . mailing depth (%) a ve ra ge p ro fi t p er m ai l ($ ) svr− svr−ps svr−random (a) (b) . . . . . . . . . mailing depth (%) a ve ra ge p ro fi t p er m ai l ($ ) svr− svr−ps svr−random . . . . . . . . . mailing depth (%) a ve ra ge p ro fi t p er m ai l ($ ) svr− svr−ps svr−random (c) (d) fig. . average profit per one mail. (a) results of % pattern selection and random sampling, from (b) to (d) are %, %, % respectively . profitability fig. compares three response models of svr in terms of profit. for each setting, the response model predicted the dollar amount to each test customer. then, we sorted the test customers by descending order in terms of the dollar amount predicted by model. a mailing depth of fig. is an expected profit made in a promotion to each percentage of top-decile. as it shows, if we send catalog mails to all respondents, the averaged profit of each mail is roughly dollars. however, if we want to select % of all respondents, we can make more profit based on svr model than only use classification model. as mailing depths go larger, averaged profits per mail decrease. svr-ps resulted between svr- and svr-random. as we saw the training time complexity analysis in fig. (b), svr-ps guaranteed efficient models with acceptable profitability and fast training speed. conclusions and discussion we applied svr for response modeling. we assumed that there was a pre- vious response model that could find all respondents, perfectly. with svr, we estimated total amount of dollar spent for each customer to find more profitable respondents. as results, we could find high profit customers rather than just response rate. also, to reduce the training time complexity, we used the pattern selection method. the pattern selection method made svr be an efficient model and, at the same time, avoid worsening the accuracy. svr- resulted the best and svr-random resulted the worst in terms of accuracy and profit, while svr-random resulted the best in terms of training speed. svr-ps located a reasonable area in terms of efficiency which guarantees ac- curate like svr- and fast like svr-random. svr-ps needs only ∼ % of training time of svr- , with acceptable accuracy loss. there are some limitations of this research. first, it was an early research of applying svr for response modeling. we formulated the problem as a regres- sion form, however, we might have missed some key factors of data setting for regression formulation. with further researches, we can find more effec- tive regression formulation to be applied response modeling. second, the re- sults of svr-ps looked efficient, however, we couldn’t decide those were good enough to apply real-world problems. the pattern selection method should be improved to be more accurate. finally, various experiments including profit analysis should be followed in further study. we assumed there was a per- fect classification model, however, there will be a comparison with the real classification model in further study. references bentz, y., merunka d., . neural networks and the multinomial logit for brand choice modeling: a hybrid approach, journal of forecasting , - . cheung, k.-w., kwok, j.t., law, m.h., tsui, k.-c., .mining customer product ratings for personalized marketing, decision support systems , - . chiu, c., .a case-based customer classification approach for direct mar- keting, expert systems with applications ( ), - . drucker, h., burges, c. j. c., kaufman, l., smola, a., vapnik, v., .sup- port vector regression machines, in: mozer, m. c., jordan, m. i., petsche, t.(eds.): advances in neural in-formation processing system , mit press, cambridge, ma, - . gönül, f.f., kim, b.d., shi, m., . mailing smarter to catalog customer, journal of interactive marketing ( ), - . ha, k., cho, s., maclachlan, d., .response models based on bagging neural networks, journal of interactive marketing ( ), - . haughton, d., oulabi, s., .direct marketing modeling with cart and chaid, journal of direct marketing ( ), - . kdd , the kdd-cup- result, . http://www.kdnuggets.com/ meetings/kdd /kdd-cup- .html. kim, d., cho, s., .ε-tube based pattern selection for support vector machines lecture notes in artificial intelligence , - . ling, c.x., li, c., .data mining for direct marketing: problems and so- lutions, proceedings of acm sigkdd international conference on knowl- edge discovery and data mining (kdd- ), new york, pp. - . malthouse, e.c., .performance-based variable selection for scoring mod- els. journal of interactive marketing ( ), - . müller k.-r., smola a., rätsch g., schölkopf b., kohlmorgen j., vapnik v., .predicting time series with support vector machines. in: gerstnerw., germond a., hasler m., and nicoud j.-d. (eds.), artificial neural networks icann , berlin. lecture notes in computer science , - . platt, j. c., .fast training of support vector machines using sequen- tial minimal optimization, advanced in kernel methods; support vector machines, mit press, cambridge, ma, pp. - . potharst, r., kaymak, u., pijls w., . neural networks for target selection in direct marketing, erasmus research institute of management (erim), rsm erasmus university, research paper ers- - -lis, available at http://ideas.repec.org/s/dgr/eureri.html. shin, h., cho, s., .fast pattern selection algorithm for support vector classifiers: time complexity analysis, lecture notes in computer science , - . shin, h., cho, s., .response modeling with support vector machines, expert systems with applications ( ), - . suh, e.h., noh, k.c., suh, c.k., .customer listsegmentation using the combined response model, expert systems with applications ( ), - . vapnik, v., . the natural of statistical learning theory, springer, new york. viaene, s., baesens, b., gestel, t., suykens, j.a.k., .van den poel, d., vanthienen, j., de moor, b., dedene, g., knowledge discovery in a direct marketing case using least squares support vector machines, international journal of intelligent systems , - . wang, k., zhou, s., yang, q., yeung, j.m.s., . mining customer value: from association rules to direct marketing, data mining and knowledge discovery , - . yu, e., cho, s., .constructing response model using ensemble based on feature subset selection, expert systems with applications, in press, available online at http://www.sciencedirect.com/science/journal/ . zahavi, j., levin, n, .applying neural computing to target marketing. journal of direct marketing ( ), - . draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . a scada oriented middleware for rfid technology ismael abad cardiel ∗ , ruben heradio† , carlos cerrada somolinos‡ , and jose cerrada somolinos§ ets de ingenieria informatica, universidad nacional de educacion a distancia, madrid, spain abstract radio frequency identification (rfid) has emerged as the new technology paradigm for acquisition and information management. rfid can be used to improve significantly the efficiency of business processes by providing the capability of automatic identification and data capture. this technology introduces new challenges on data and process information management in current systems. rfid data are time- dependent and dynamically changing. in addition, data carry implicit semantics. the homogeneous data processing of such implicit semantics allows us to propose rfid middleware as a who–when–where data problem. this paper presents depcas, a new middleware for rfid information based on the scada architecture for control systems. an application of depcas is the resolution of heterogeneous situations, which solves the what or context–aware to apply the auto identification data received from rfid systems in business applications. introduction due to the ubiquity of radio frequency identification (rfid) deployments and low cost tagging [ ], a growing amount of data are being generated from multiple sources that need to be processed and managed. so, new questions are arising, such as how to encompass current and future identification technology at the same time and how this information can be combined with existing applications [ , , ]. the large amount of information generated in the rfid startup projects requires to know how and when the information is going to be summarized [ , ]. many of the non–finished rfid pilot projects have been focused in the rfid data instead of being fo- cused on the business related events [ ]. raw rfid data have little value until they are is transformed into a form suitable for application-level interaction [ ]. moreover, data have implicit meanings and associated relationships with other rfid data (e.g., data about containment, position, aggregation...). so, rfid appli- cations must make appropriate inferences. for instance, observations of rfid tags associated to several items and a rfid tag associated to a container in a certain time interval can imply that the items are packed in the container. in addition, the wide spread use of rfid technology involves a huge migration problem. there are thousands of applications using identification through codes, bar codes, plates, badges, d codes, operator identifications, etc [ ]. that in a mid-time will change to rfid auto identification [ ]. from the application ∗iabad@issi.uned.es †rheradio@issi.uned.es ‡ccerrada@issi.uned.es §jcerrada@issi.uned.es draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . that can be found in any little market, to the large and more complex information system of a big and global company, the change from identification to auto identification will arrive [ , ]. the state of the art in rfid integration in existing and real systems provide a research prospect to propose a general middleware architecture that processes and consolidates rfid information solving the former issues and helping efficiently to migrate from identification to auto identification systems [ ]. this paper proposes a new middleware architecture for rfid called depcas (data epc acquisition system), which integrates existing proposals for rfid middleware and it is based on the supervisory control and data acquisition (scada) architecture for control systems [ ]. key concepts supported by depcas are: . converting rfid information to business information: rfid applications receive basic and simple data source inputs as time–stamp, tag code or reader dock (reader plus antenna) [ ]. this raw information should be translated to business relevant information using abstract process that can be applied in heterogeneous rfid installations. . distributing rfid information to business processes: the rfid processed information need to be de- fined and accessible for external applications in a general way. the format and the exchange mech- anisms must be transparent and general for any kind of applications. the exchange flow between existing running programs and rfid acquisition system is bidirectional. it is necessary to solve how to send the necessary business information data to the rfid acquisition system and the opposite flow from rfid middleware to external entities. . filtering information: the computer systems that manage a big quantity of input data must solve a filter policy. filter process could be as simple as eliminating all unknown information or as complex as operations that manages every time–stamp and any network structure to ensure the data quality. in the rfid acquisition systems the raw tag readings are filtered according to a previous algorithm that helps the back end application to receive only relevant information. . triggering business process from rfid process and vice versa: it is a common situation to solve that rfid acquisition system can receive any kind of input that triggers external activities. and in the opposite way, external input can modify rfid acquisition process. in any direction, rfid acquisition must solve the event trigger process. . managing rfid devices: this requirement deals with an abstraction of physical communication and protocol exchange between reader devices and the acquisition platform. it is necessary to manage also the topological reader network, i.e., how readers are distributed or centralized. . consolidating rfid information: the general rfid process generates relevant information that has to be consolidated. this permanent information is the result of specific process according to the homogeneous process. the homogeneous process creates detailed information about how, where, and why rfid information is being used in a productive process. the remainder of this paper is structured as follows. section summarizes related work. section describes depcas. section outlines our current implementation of depcas. section presents a scenario where depcas is used. finally, section presents the conclusions of our work. related work the need of rfid middleware is gradually becoming a basic question for non–trivial rfid installations. this is particularly true in heterogeneous environments, including multiple readers, applications instances, supply chain management, complex process and sophisticated business semantics [ , , , ]. epc stands for electronic product code. draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . rfid middleware is indispensable for these main reasons: . the need to filter rfid data in order to avoid redundant or erroneous information that is not needed to business applications, while at the same time optimizing resources. . the need to process and interface rfid data from origin (tag reads) in heterogeneous deployments (multi-tagging and multi-readers systems) without resorting to business integration logic. . the flexibility to integrate rfid systems to support auto–identification in different applications and process models. there are several surveys in the literature that propose system taxonomies and compilations about rfid middleware [ , ]. the rfid middleware market is dominated by three tendencies: . specialized companies which offer rfid middleware. these companies can be classified into product oriented and specific rfid middleware producer. if they are reader/antenna, or general hardware producer, then they include specific rfid middleware to support the hardware. if we speak about a general middleware producer it is quite common they have included an specific plugin to acquire rfid data. there are new companies that have born with the specific commercial target to create, distribute and install rfid middleware. we can mention in this group: (a) winrfid from ucla-winmec research lab [ ] is a research result to solve rfid features like scalability and administration, event and data intelligent process and dispatching. (b) oat foundation suite is a rfid middleware platform that includes two-way process integration between enterprise applications and rfid acquisition systems. (c) detego/you-r open is the rfid middleware from rf-it solutions based on three main func- tions: business integration, data management and device management. (d) truevue rfid platform from vuetechnology is a complete software infrastructure to manager rfid deployment. it contains several modules to support different points of view in item level management. . the second group is the one of giants’ software vendors such as ibm, microsoft, oracle/sun mi- crosystems, or sap. every of these huge companies have dedicated in the last decade important resources to include rfid acquisition inside their own products. between the rfid middleware we can mention: (a) rfid java middleware from oracle/sun microsystems . with oracle and sun microsystems company fusion, the rfid middleware from both companies was renamed using the oracle reference. now the product name is oracle sensor edge server based on oracle fusion mid- dleware. (b) sap solutions for auto-id and item serialization is a set of components to include rfid tech- nologies in sap management tool. the solution offers three major components: sap auto-id infrastructure, sap object event repository and sap event manager. http://www.oatsystems.com see “plug and trace. excellent rfid solutions” available at http://www.enso-detego.com http://tycoretailsolutions.com see “oracle application server wireless developer’s guide” available at http://docs.oracle.com see “sap research report / . from creativity to value” available at http://research-report- .research-events.com/ http://www.oatsystems.com http://www.enso-detego.com http://tycoretailsolutions.com http://docs.oracle.com http://research-report- .research-events.com/ draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . (c) ibm websphere rfid is the ibm middleware to implement rfid acquisition. it consists of edge- controllers and the websphere rfid premises servers which includes websphere application server and db workgroup server. (d) rfid anywhere from sybase ianywhere is a software infrastructure that simplifies the develop- ment, deployment, configuration and management task for rfid networks. . the last group of rfid middleware producers is the open source initiatives. these proposals are research results that take origin in university groups or consortiums. there are several open source initiatives to solve an rfid middleware: (a) cuhk rfid middleware from mobitec technologies centre of the chinese university of hong kong. (b) aspire rfid middleware is an eu funded project to deliver an advanced sensors and lightweight programmable middleware for innovative rfid enterprise applications as an open source mid- dleware. (c) rifidi edge server is a software architecture proposed by the rfid research center in the university of arkansas. (d) fosstrak [ ] (called accada) is a middleware implementation sponsored by auto-id labs oriented to produce a free and open source software for track and trace. scada and rfid middleware systems share a set of commonalities about structure, functions and characteristics. from an acquisition point of view, the physical low level is a set of devices acquiring data: we have called sensors and rtu’s in scada world and we call antenna/edge/reader in rfid world. the communication networks to integrate the data capture system with the master processing system. this network communication could be defined with a variety of technologies and topologies. and the last element: the server data processor. the functions in scada and rfid middleware systems are similar. main question of both of them is to acquire data: scada data acquisition could be expressed in terms of discrete inputs (equipment is on or off, tripwire alarms, power failures, etc.) and analog values (level in tanks, voltage levels, temperature, or any other factor, etc.). rfid middleware data acquisition system is a three-tuple: the rfid code or the who, the time–stamp or the when, and the reading point or the where. if we analyze the characteristics in the two kind of acquisition systems, there are a set of common points between scada software systems and rfid acquisition system: . hiding hardware: scada software system hides the specific remote terminal units (rtus) hard- ware. it is a common situation that scadas manage a set of rtus from different companies or with different structure. this same situation could be found in rfid middleware with a deployment of an rfid infrastructure. at our days we can find a set of various rfid readers’ trades like alien, intermec, feig, siemens, symbol, sick, etc. . hiding communication between devices and master software: data transport from acquiring devices to master software can be made using diverse communication infrastructures and different protocols. there are all kind of scada communication infrastructures, from the simplest serial wired structure to the most complex satellite structure. additionally, any advance in telecommunication field will be used if there was any opportunity. the same situation may be repeated in rfid systems. the present see “ibm websphere premises server” available at http://www.ibm.com see “getting the most from mobile rfid with middleware” available at http://www.sybase.com see “cuhk rfid middleware - system design document” available at http://mobitec.ie.cuhk.edu.hk http://wiki.aspire.ow .org http://www.rifidi.org http://www.ibm.com http://www.sybase.com http://mobitec.ie.cuhk.edu.hk http://wiki.aspire.ow .org http://www.rifidi.org draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . readers have limited options to communicate to master software using ethernet network or serial communications. but there are prototypes using optical fiber or wireless communications. . quantity information: one of the more significant common points between scada and rfid acqui- sition software is that they receive a relevant quantity of dynamic information. scadas can receive data with milliseconds of resolution from a big quantity of data signal to monitor a complete instal- lation. the same situation may be produced inside a rfid monitoring installation. this data should be received and processed to obtain relevant information. . from fine-grain information to relevant information: both systems acquire data characterized from a very fine grain. scada basic data are status signal, analog value, and may be rate counters. rfid acquire system receive data about auto identification with a code. the two type of systems have the possibility to manage additional basic information like time–stamp, read quality, read origin, etc. all this individual information has very little meanings and it has a very low meaning relevance. the relevant information is obtained when it is processed with others data to obtain significant information. . real time process: scada systems are typical real time systems. data is acquired and processed to obtain the relevant alarms or calculation that give information to the operators. of course, there are differences about the meaning of real time depending on specific situations. it is not the same to speak about real time in electrical scada systems that in water scada. time scale it is not comparable between electrical processes with water processes systems. the same situation it is reproduced in rfid acquisition system. auto identification could be time critical as the process where it has been using. . consolidating information: the scada and rfid middleware process data is defined to obtain sig- nificant information not only real time alarms or events. this processed data is structured, stored, and summarized to be available for other systems. all this operations allow filtering managing the information that is received in the system. . external interfaces: the consolidated information should be offered to others systems. modern scada software systems manage external structured information to define and configure the own process. this information has in-out components. scada receives information about configuration, geographic information system (giss), maintaining systems or any internal warehouse. scada offers information about the process results to any back-end company software. this situation is repeated in rfid middleware environments. depcas figure depicts the overall depcas architecture. the scheme proposed here is based on the architecture of modern scada (supervisory control and data acquisition) systems. the basic issues in depcas are: . producing rfid processed data. . hiding heterogeneous rfid deployment systems with an homogeneous layer approach. . translating business (async or sync) needs to rfid systems. . providing management capabilities in rfid middleware environment. to solve these main targets the structure of the acquisition system that we propose is organized in four great layers or subsystems: draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . figure : architecture of depcas draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . . middleware device manager (mdm). it has the following basic functions. first, the communication management with one or more devices: rfid readers. second, implementing communication pro- tocol with the readers and event processor implementation (ale: application level event used in the epcglobal standard ). and third, topological configurations support defined by the acquisition equipment: antennas, edge readers and readers. . middleware logic manager (mlm). it supports the process to obtain rfid summarized data from auto-identification raw acquired data. the logical process to resolve will depend on each particular scenario [ ]. this process will generate information of a permanent nature which is the result of the process in every scenario instance. the basic scenarios on the proposed work are related with: monitoring and tracking operations, aggregation of information items to generate new information, sorting of items, format raw rfid data conversion, and state machine delivery information. the concept of scenario is equivalent to a logical process that generally can be applied in all situations where information is used for self-identification. at each stage information is received, processed, summarized and accumulates, and generate and consolidate the data according to an specific logic. . graphical user viewer (guv). it is the mechanism that depcas supplies to help in the installation and management of a rfid system from the engineering point of view. this subsystem supports two basic operations: the supervisory capability to the rfid system and certain data manageability (configuration and acquisition data). . information system exchange (epcis: epc information services). epcis services should be the gateway of information between the data provided by the mdm and different mlms to external business. these services will enable both receiving and sending information from the depcas system to other systems. following subsections describe in detail mdm, mlm, guv and epcis. . mdm one of the main purposes of rfid middleware is to manage and hide the broad range of device readers existing in acquisition rfid networks. there are two basic solutions to override this question. the first one is to define a specification that should be accomplished for every rfid reader inside a network. the second solution is to define an abstraction layer that can translate proprietary protocols and specific characteristics from a vendor to a general solution. mdm is the specific depcas layer to support all these management operations. the study of different rfid device management components allows obtaining a set of features that are included to solve it in this kind of systems. among these features we include: . device agnostic: the device management should include a specific function to hide the alternatives of existing devices related with: physical parametrization, communication protocol, communication options, etc. . configuration and parametrization of rfid devices: an agnostic approach to existing devices sets the options to manage from an homogeneous point of view the configuration and parametrization of every equipment in the rfid acquisition network. . topological rfid devices: the rfid device management can include some network organization that give additional capabilities to the acquisition network as: redundant point of reading, input charac- teristics, out characteristic, parallel reading devices, etc. see “application level events standard v. . . ” at http://www.gs .org/gsmp/kc/epcglobal/ale http://www.gs .org/gsmp/kc/epcglobal/ale draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . . start-up devices: the set of operations to startup and configure an rfid equipment inside a network are known as starting up protocols. some of the existing rfid middleware support a starting up protocol giving instructions about how to include rfid reader in the network. . statistical and report rfid devices data: the deployments of rfid reader require to introduce mainte- nance operations in the network. the statistical and reporting information are essential to optimize these operations. . monitoring devices in live operation: some existing rfid networks can manage critical operations thus the rfid middleware supports a supervisory management about working devices. . diagnosis of devices in live operation: the rfid middleware includes operations like device signals, examining logs, updating software devices, etc. . llrp implementation: the low level reader protocol is the standard proposal from epcglobal network to specify the interface between rfid readers and any software client. . osgi technology: the increase of software complexity in heterogeneous hardware environment has promote the development of frameworks to support the continuous changes. osgi framework is a module and service platform that implements a complete and dynamic model. any framework that implements osgi provides an environment for modularization of application in small pieces tightly-coupled, dynamic loadable and configuration files to declare external dependencies. the osgi architecture is one of the most common specification framework used in our days. the osgi specifi- cation enable components to hide their implementation from others components while communicat- ing through services. these features and functions define the operation of the middleware device manager in rfid mid- dleware. for instance, device agnostic feature is provided when middleware infrastructure acts as an integration point for rfid readers, printers, and other infrastructure devices, such as any sensors and actuators. rfid reader and printer manufacturers typically provide complex configuration options, pro- prietary interface languages, and communication protocols. if rfid device infrastructure eliminates the need for rfid integrators to be aware of these complexities and provides a device agnostic interface with which supported devices can be configured hiding the complexity. other features are related to scope and technological implementation of middleware device manager. table compares the features supported by depcas mdm to the related work summarized in section . the main goals of mdm in depcas are: . managing the topology rfid reader’s network. . agnostic management of different rfid vendors. . homogeneous process of different rfid tag reads. . supporting llrp and ale specification process. . providing configurable rfid network. . hiding operational environment. . minimum tag management operations. . monitoring and controlling rfid readers. see “llrp standard v. . ” available at http://www.gs .org/epcglobal http://www.gs .org/epcglobal draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . to support these objectives, the mdm definition includes the concept of minimum abstract reader commands (marc) and the rfid reader topology language (rrtl). the marc allows hiding the particu- larities of every rfid reader vendor through a set of operations that are mapped to specific vendor api. the rrtl includes the description of rfid topologies including reader control and management. the marc is the minimum set of instructions that an rfid reader must supply to support the operational instructions inside a depcas network. the set of marc is closed and give an independent layer of rfid management. the implementation of marc is related to rfid reader api and gives an independent hardware layer. the rrtl is a language that includes primitives to define depcas network. the sentences of this language include the necessary semantic to explain how the readers are organized inside the acquisition network. f o s s t r a c k r if id i a s p ir e o r a c l e ib m d e t e g o o a t r f id p la t. d e p c a s device agnostic yes/no no yes yes yes yes yes yes device configuration yes yes yes no no yes yes yes topological configuration no no yes yes no yes yes yes needs programming starting-up device yes yes yes no no no no yes/no statistical data from device yes no yes no no yes yes yes monitoring data from device no no no yes yes yes yes yes diagnosis capabilities yes no yes no yes no yes yes llrp implementation yes yes yes no no no no no osgi technologies no yes yes no no no no yes using marc sentences to configure and connect rfid devices. using rrtl. depending on the api of the chosen reader. statistical operations are included in marc sentences. marc sentences include data quality evaluation if it is possible to obtain information from readers. there are marc sentences to produce device diagnosis and there are operations included in rrtl about topological diagnosis. to process the parallel rfid readers. table : depcas mdm support compared to related word . mlm current rfid middleware systems capture raw readings from readers, filter and aggregate the data, and finally yield cleansed rfid raw events. however, rfid raw events only provide low-level information such as epc [ ], reader identification and time–stamp, which are not directly related to business processes. therefore, the ability to quickly transform these raw events into business events useful for decision-making becomes a central issue for realizing the maximum value from rfid technology in business. the key aspects to take in count in processing rfid information are: . process data: the rfid raw process is more than a basic filtering, it is a cleansing, consolidation and summarization set of operations that have as objective to create and consolidate useful rfid data for business. . alarm: one of the purposes of rfid middleware is to make difference between relevant and not relevant data in acquisition. the alarms are the first processed data that are generated inside the middleware. rfid middleware’s can generate two types of alarms: infrastructure alarming and pro- cess alarming. draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . . event: the rfid acquisition system should digest the meaningful events. the alarms are also event from this point of view but not only. the rfid middleware can define other relevant situations that are interesting for external application or to support the system exploitation. . filter: the filter rfid process refers to the removal of certain tag read events based on any dis- criminator parameter as tag code repetition in a short period or the same tag code in two different readers. . consolidate: the rfid middleware process will create a new rfid data that should be consolidated and maintained inside any database systems. this consolidation opens the way to share and to transmit to external applications the cooked rfid information. table compares the features supported by depcas mlm to the related work summarized in section . existing rfid middleware includes a basic rfid data process. this basic process is related with event pro- cess and filter reads in the rfid middleware. some of the existing middleware also propose an additional operation, called aggregation or summary reads. the basic idea under this operation is to reduce the tag reads using any criteria. for instance, aggregating operation can detect input/output read, selecting just the first and the last read when a tag appear under read range. other approach is to count the tag reads and accumulate it. other rfid include specific business modules. these modules perform business explicit process from the raw acquisition systems and offers direct connectors with the back end application. o r a c l e ib m f o s s t r a c k r if id i t ru e v u e d e t e g o d e p c a s alarm no yes no yes yes yes yes event no yes yes yes yes yes yes filter process yes yes yes yes yes yes yes aggregation process no yes no yes no yes yes/no configurable data process no no no no yes yes yes consolidate data yes yes no yes yes yes yes specific business modules no no no yes no yes no supply chain process yes yes yes yes yes yes yes epc network support yes yes yes yes yes yes yes depending on every scenario process. scenario configuration. scenario process results. scenario configuration applies to every business case. using the scenarios according to the supply chain process definition. using the scenario process. table : depcas mlm support compared to related word the main target in the conception of depcas is to think about the use of homogenous rfid middleware that can be introduced in working systems that already use identification (and want to change to use auto identification). for instance, we are habituated to use the bar code in many circumstances in daily life, like shopping, luggage, personal cards, etc. all these heterogeneous systems that use identification in a half term will transform into systems that use the auto identification. the mechanism that solves this homogenization in depcas is the scenario concept, which is equivalent to a logical process with a general application that it is possible to be used in a set of situations where the information takes advantage of auto identification. in each scenario there are rfid data that is received, processed, summarized and accumulated, according to their own logic. the basic defined scenarios in depcas are: draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . . tracking of items: this scenario corresponds to an application in which it is wanted to make the tracking of some object in a chain that is supervised with rfid readers. the objects when happening through each point generate the identification events that are registered from an origin to a destiny. the different events will depend on the parameters that are desired to check: time, distance, speed, step/no step, complete route, etc. section details this scenario. . items classification: this scenario consists of a classification of elements after an identification of an object. the read event from an object generates the classification according to certain criteria. the application of this scenario would be applicable to an automatic robotically classification system. or to a system of shipments classification. . aggregation of items: this scenario would allow adding auto identified objects to generate a new information registry. this scenario can be applied in those cases in which a product is develop by parts that are added, like for example the creation of baskets of gifts, pallets, etc. . format conversion: this scenario would allow generating information from auto identification codes. for instance, it could be applied for systems in which hybrid identifications are received and wanted to make uniform. this scenario allows to turn formats epc to other formats of particular objects like the isbn, issn, etc. . execution of machine of states: the scenarios that represent state machines would allow associating information of auto identification to the identified object states. for instance, it could be used to make the track of a rent of vehicles or the management of loans in a library. . guv guv is a combination of configurable components, that graphically reflect the dynamic state of the simple components (tags, readers...) or complex ones (routes, adders, state machines...) obtaining different types of information from data servers. the creation of a specific application will be carried out by the grouping of these components, properly configured like a client data or on a web node for an internet browser. the study of different rfid guvs allows obtaining a set of characteristics that are implemented to solve the user interface in the rfid middleware. among these features we include: . web client: the graphical component can be supported as lightweight software using web client applications. . collection of dynamic information: guv should include dynamic graphical components to show the status and quality of represented data. . managing the database infrastructure of rfid: the database management tool is a graphical com- ponent that simplifies the database operations (i.e., search, create, delete or update) about rfid infrastructure (e.g., readers, tags...) . managing the process database: the database management tool is a graphical component that simpli- fies the database operations (i.e., search, create, delete or update) about rfid process configuration. . lists of information: the list graphical resource presents the database information like a summary where the user can choose the rows and the columns shown for an specific table. . editing and management of graphical elements: the option to edit and manage the graphical resources allows building specific user environment depending on installation or user authorization policy. draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . . management and configuration of alarms: alarms are one of the most important resources on rfid middleware. guv includes several mechanisms to present the alarming information: tables, sum- maries, flashing and sound signals, etc. . user management and privileges: these kinds of options allow defining graphical user organization and policy to show/not to show data to the users. table compares the features supported by depcas guv to the related work summarized in section . the main objectives of guv are: . including a single common environment and information management in depcas middleware. this environment is common to all management information system configuration data regardless of being related to infrastructure, the processing logic sets, or the epcis modules. . managing a common graphical object to support information lists. these lists can be used like dynamic queries with real time information or static queries to show database information. . supporting graphical displaying information about a particular object in real time. you may include in new graphics another graphical object with real time attributes. . representing alarm system that can be configured. this treatment includes alarm management levels and the realization of automated operations . defining the process and system events for the acquisition of auto identification data. . managing graphical objects that are used and which allows their reuse in different graphical inter- faces. o r a c l e ib m s y b a se t ru e v u e f o s s t r a k w in r f id r if id i d e p c a s web client yes yes no no yes no no no dynamic information yes yes yes yes yes yes yes yes configurable dynamic information no yes yes yes no yes yes yes rfid database management yes yes yes yes yes yes no yes rfid process configuration no yes yes yes no yes no yes lists yes yes yes yes yes yes no yes graphical edition no no yes yes no no yes yes alarms no yes yes yes no yes no yes events yes yes yes yes yes yes no yes user management yes yes no no yes no no yes using graphical connection to real time depcas database and depcas repository. using ediguv to create and compound graphics. using a uniform database management tool (dmt). dynamic and static graphical representation. like an special dynamic list and connected to the real time database. like an special static list and connected to the repository database. using user and user groups to define operational capabilities. table : depcas guv support compared to related word draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . . epcis the epc information service provides a specification framework to solve the external rfid middleware applications to capture, secure and access rfid data. the original epc network architecture was defined in [ ] like an “omniscient entity” to provide relevant information required from business application to rfid middleware. the main four functions included in epcs in this initial proposal were: providing a long term repository of epc event data, supplying an access point to higher-level information obtained from epc process acquisition, solving the storage and access to detailed business information associated to epc process acquisition, and providing transparent on-demand access to data attribute held in others systems. epc data was interchanged using the physical markup language (pml) [ ], an xml schema for representing epc–related data including auto identification information (e.g., codes, readers, antennas, time–stamp...) and business information (e.g., size, quantities, width...). these functions were so broad that a general confusion about how to implement them happened. these pending questions are got pass in the definite epcis specification definition [ ] including an standard in- terface to allow epc data interchange. this standard interface is characterized by: managing only historical data (real-time epc data is decoupled from epcis), operating with cooked epc related information (raw epc data is processed in low epc network layers), and dealing only with enterprise systems environment. all this specific functions are solved using a three layer interface: the abstract data model layer, the data def- inition layer and the service layer. the abstract data model is a closed and defined interface that proposes a specification of general requirements for creating data definitions in the data definition layer. this data definition is called core event type. this structural organization in the interface allow to use standard vo- cabulary elements used by inter organizations and user specific vocabulary elements defined and meaning inside intra organizations. despite the objective of all these implementations is the same: to connect business with rfid middle- ware, there are a set of characteristics that differ between specific solutions. this analysis introduces the following features: . standard compliance. what is the compliance degree in the existing implementation according to epc network standard? . epcis permanent data model implementation. what is the existing repository implementation sup- ported by epcis solutions. . security epcis control. there is a large range of security control levels in epcis data, from general user access to low granular data access. . metadata management. the business process defines process data and specific attributes that are introduced in the epcis repository. . supply chain support. the epcis implementation has specific elements to solve supply chain ques- tions. . reports. the epcis includes report facility to data access using xml exporters. table compares the features supported by depcas epcis o the related work summarized in section . the main functionalities of epcis in depcas are: . publish/subscription manager. . permanent historical data information. draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . o r a c l e ib m w in r f id f o s s t r a c k d e p c a s standard compliance partial total partial partial no data model oracle oracle oracle mysql yes postgresql db sqlserver web server jini websphere unknown apache yes security webserver propietary remote authentication no security security objects and authorization for query operations metadata using specific xml no support yes data in metadata framework for new repository model vocabulary type supply chain support yes yes specific yes yes adaptor reports not business not not not reports included intelligence included included included reporting tool epcis events yes events rule yes yes and manager alerts exceptions to solve the relation between business and mlm scenario. as a publish/subscription server to the depcas repository. metadata captures the relationships between the rfid product-process data. as a middleware connector. to manage the external connection and to supervise the way connections are working. table : depcas epcis support compared to related word draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . . transfer in/out data between depcas and business applications. . security distribution of depcas information. . xml data mapping. . specific application connector with depcas a depcas prototype depcas basic implementation includes a prototype module for all its components. . mdm implementation the depcas mdm core architecture prototype has been developed using tcp and http for transporting reader protocol messages, while the message content can be either xml or text. in addition, it supports synchronous and asynchronous messaging (through the reader’s protocol notification channels mecha- nisms). the mdm marc sentences implementation includes the following operations: . setantennasequence: set an order sequence to read from multiple antennas installed in one reader . setautofalseoutput: order to autonomous function mode to ignore tag reads . setautofalsepause: order to autonomous function mode to pause working . setautomode: start/stop autonomous reader functions . setautostarttrigger: activate the trigger sending functions in reader autonomous function . setautostarttimer: activate periodic timer reader function in reader autonomous function . setautostoptimer: reset periodic timer reader function in reader autonomous function . setautostoptrigger: reset the trigger sending functions in reader autonomous function . setautotrueoutput: order to autonomous function mode to process tag reads . setautotruepause: order to autonomous function mode to start reading process . setconnecttcp: establish a tcp connection . setnotifyaddress: init a listening tcp port to instruct the reader to send notification messages . setnotifyformat: configure the format message to be used by the reader . setnotifytime: establish the period that the reader notifier should use . setpassword: identify the password reader access . setrun: start to execute a configuration reader mode . settagtype: establish the expected tag type property . settime: synchronize reader time draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . . setusername: identify the user reader access the mdm prototype includes marc implementation for rfid reader alien , feig , skywraem and we are working in marc implementation for thingmagic mercury , intermec if and impinj speed- way r . . mlm implementation the general process in any of mlm configuration follows the general flow presented in figure . mlm uses two main groups of configuration tables: the scenario configuration tables and the static configuration ta- bles. the scenario configuration tables include data structure about the specific scenario that is processed. for instance, in the tracking scenario (see section the most relevant table is the “routes” table that in- cludes information about the itinerary that items are going to be monitoring. the static configuration tables for mlm includes data about the general elements processed. for instance, what codes are going to be supervised (the who’s that depcas is going to process) or the reader points in the systems (the where’s that depcas logic manages). the mlm initial trigger is always a rr (relevant read) that always include a who (rfid code), a where (rfid reader point) or a when (rfid read time–stamp). rrs are processed one by one or can be processed by set of data by the rr process. these processes generate data logs information from the processing result of incoming new rr and static configuration. for instance, if there is a rr with a who (rfid code) that is not inside the static configuration table the process generate a not_ok log and create the necessary data inside alarm, event or statistical tables. figure : general process of depcas mlm the scenario loader initializes the configuration parameter of each scenario. for instance, in the state machine scenario process the initial operation starts every state machine defined to the initial state, pro- ducing the ignorance of every read associated to other states. the dynamic scenario tables define pro- prietary data structures for every particular scenario. these tables are managed by the dynamic scenario draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . process and they are scenario dependent. in the aggregation scenario, there are several scenario depen- dence process, for example the compositions table receives the items that are made up of the complete aggregation. when the items are processed they area accumulated to an specific aggregation. the notifier process communicates with the business event process (bep) to trigger it the events that are configured in every scenario. . guv implementation based on the concepts of high-level design have been shown we have developed a guv prototype envi- ronment that allows building graphical environment for a specific domain. the prototype has a graphic editing system that allows creation, modification and deletion of graphical elements. this environment, called ediguv defines a core container referred as depcasguv in which may include other edited elements. the prototype operates two types of graphic elements: the basic elements that are provided directly by the window sdks and the compound depcas elements that are the result of basic elements compositions for a specific function in depcas middleware. the basic graphic elements in the actual depcas version are: . windows . buttons . edit fields . text fields . check boxes . combo boxes depcas compound graphic elements are: . panel database (add-delete-modify data) . list of dynamic database . list of static database . rfid / epc label . rfid reader point . dynamic graph (route / forecast) . epcis implementation epcis is implemented using the following four subsystems: . publish/subscription manager supports two kinds of relationships: relationships that are simple data representations of some actual relationship that exists between business atomic data and depcas data, and relationships between tables that have their own properties. data representations of rela- tionships can contain a web service reference to access the relationship web service. the epcis imple- mentation defines interfaces to query a manageable resource about the relationships it participates in directly, such as a containment relationship. the publish/subscription manager also defines an draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . interface for a service that offers relationship information for many manageable resources enabling a relationship registry. manageable resources do not have to be aware that they are participating in relationships. . converter manager implements a global repository that captures the relationships between the rfid product-process data, that we call scenario in depcas terminology, and existing business data. the converter manager layer offers the possibility to control the level of aggregation by customized com- bination of several built-in filters as well as those developed by specific environments. hence a wide range of variation is at hand. attaching meta-data from backend-systems to depcas scenarios re- quires connecting specific existing external data to depcas scenarios tables. . feeder epcis interface works using a java message service (jms) to exchange data from depcas real time processing scenarios to epcis process. this capturing interface distributes the information to the consolidate database and to synchronous ecpis connectors. the message structure is defined using an xml schema. this schema structures the general data field of any message according to a specific scenario. for example, the jmstype for a tracking scenario includes: general position, specific place, building, area, reader used, department, auto id read. this message is instantiated as showed in figure . . query epcis server supports the depcas servers/clients connections. it solves the different data interchange (http, web services or xml) between depcas and external business applications. in the same way, epcis server is designed as a platform with a uniform query and update interface to ap- plications, while the actual implementation solves the details and data binding to existing databases and information systems. epcis server supports simultaneous binding to multiple databases and information systems from multiple vendors, as well as epcis as a managed application service. in order to accommodate various types of data relationships such as routes, predictions, reader points, associations with particular transactions, epcis server is implemented as a modular framework, with a very lightweight core functionality, which merely described which “profiles” or relationship types are implemented on any particular instantiation providing an epcis interface. startup time: .jan. : : : .jan. : : : queue name: jms/queue consumer queue selector: jmscorrelationid = ’uned’ and jmstype like ’issi_ ’ to end program, type q or q, then recv message: .jan. . . : message header: .jan. : : : producer- , message order: data message: uned, campus , edificio a, zona reader_a issi_ recv message: .jan. : : : message header: .jan. : : : producer- , message order: data message: uned, campus , edificio a, zona reader_a issi_ figure : example of epcis message an example of depcas scenario the tracking of people or goods with identification mechanisms is repeated in heterogeneous form in diverse situations. the tracking operation abstraction can be considered in terms of a route to monitor draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . and the parameters to process in this route, that we call predictions. the fundamental software function is tracking objects that travel by the defined routes. the process for each case depends on the type of parameter assigned to the route. the different types of routes make the corresponding verifications in each case: . bdr (basic data routes): they are fixed routes in which the passage in a certain absolute time is verified. . bdtr (basic data time routes): they are fixed routes in which the passage in a certain relative time relative to the origin is verified. . adr (alternative data routes): they are alternative routes in which one verifies the origin and the destiny in a certain absolute time. . aditr (alternative data intermediates time routes): they are alternative routes in which one verifies the alternative origin, destiny and intermediate steps in an absolute time. . adrtr (alternative data relative time route): they are alternative routes in which one verifies the origin and the destiny according to a time relative to the origin. . adrtitr (alternative data relative time intermediate time routes): they are alternative routes in which one verifies the alternative origin, destiny and intermediate steps in a time relative to each made step. . upur (undefined predictions undefined routes): routes of any type for which any type of prediction algorithm is not defined. the different processes generate the warnings that determine if the readings correspond to the pre- dicted thing or they do not correspond and why. between the system warnings we can distinguish three groups. the first group are the warnings generated in correct tracking, we denominated log_ok. a second group is related with a correct tracking in which some type of incidence takes place, like for example not fulfilling a certain part of a route or for those cases in which the route is fulfilled with some time delay error. the third group of warnings are those generated in case of breaks of established routes, an example of this type are the log_lr_nok warning that is generated in the case of appearing labels in routes in which they would not have to appear. in order to verify the operation of this type of scenarios, a monitor has been developed that shows all information about the labels tracking. conclusions research on rfid middleware is oriented to find new alternatives to efficiently include auto–identification into productive processes. new topological networks of tag readers and heterogeneous business applica- tions need to be connected through rfid middleware, solving all the established requirements. in this paper, we have summarized the features that must be supported by any “real” rfid middleware. as a result, we propose the depcas middleware, which starting point is the well known scada architec- ture. after describing the theoretical basis of our approach, i.e., the architecture of depcas, its current implementation and an application scenario have been reported. draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . acknowledgements this work has been supported by (i) the spanish government under the cicyt projects dpi- - , dpi- - and dpi- - , (ii) the comunidad de madrid under the robocity -ii excellence research network s /dpi- , and (iii) the universidad nacional de educacion a distancia under grant v/puned/ references [ ] j. j. jung, ubiquitous conference management system for mobile recommendation services based on mobilizing social networks: a case study of u-conference, expert systems with applications ( ) ( ) – . [ ] x. su, c.-c. chu, b. prabhu, r. gadh, the internet of things: from rfid to the next-generation per- vasive networked systems (wireless networks and mobile communications), auerbach publications, , ch. on the creation of automatic identification and data capture infrastructure via rfid and other technologies. [ ] c. floerkemeier, c. roduner, m. lampe, rfid application development with the accada middleware platform, ieee systems journal ( ) ( ) – . [ ] q. z. sheng, x. li, s. zeadally, enabling next-generation rfid applications: solutions and challenges, computer ( ) ( ) – . [ ] a. ustundag, m. s. kilinç, e. cevikcan, fuzzy rule-based system for the economic analysis of rfid investments, expert systems with applications ( ) ( ) – . [ ] c. lee, t. chan, development of rfid-based reverse logistics system, expert systems with applications ( ) ( ) – . [ ] c. bornhövd, t. lin, s. haller, j. schaper, integrating automatic data acquisition with business pro- cesses experiences with sap’s auto-id infrastructure, in: proceedings of the thirtieth international conference on very large data bases - volume , vldb ’ , vldb endowment, , pp. – . [ ] j. curtin, r. kauffman, f. riggins, making the śmostŠ out of rfid technology: a research agenda for the study of the adoption, usage and impact of rfid, information technology and management ( ) – . [ ] m. he, s.-j. horng, p. fan, m. k. khan, r.-s. run, j.-l. lai, r.-j. chen, a fast rfid tag identification algorithm based on counter and stack, expert systems with applications ( ) ( ) – . [ ] a. brewer, n. sloan, t. l. landers, intelligent tracking in manufacturing, journal of intelligent manu- facturing ( ) – . [ ] c.-s. lin, l.-s. chen, c.-c. hsu, an innovative approach for rfid product functions development, expert systems with applications ( ) ( ) – . [ ] y.-c. lee, s.-s. lee, the valuation of rfid investment using fuzzy real option, expert systems with applications ( ) ( ) – . [ ] n. kefalakis, n. leontiadis, j. soldatos, d. donsez, middleware building blocks for architecting rfid systems, in: mobilight, , pp. – . draft version of the paper published in expert systems with applications, doi: . /j.eswa. . . [ ] zafer, aydogmus, implementation of a fuzzy-based level control using scada, expert systems with applications ( , part ) ( ) – . [ ] c.-c. hsu, p.-c. yuan, the design and implementation of an intelligent deployment system for rfid readers, expert systems with applications ( ) ( ) – . [ ] c. cheung, c. cheung, s. kwok, a knowledge-based customization system for supply chain integration, expert systems with applications ( ) ( ) – . [ ] c. lee, w. ho, g. ho, h. lau, design and development of logistics workflow systems for demand management with rfid, expert systems with applications ( ) ( ) – . [ ] j. m. ko, c. kwak, y. cho, c. o. kim, adaptive product tracking in rfid-enabled large-scale supply chain, expert systems with applications ( ) ( ) – . [ ] m.-c. kim, c. o. kim, s. r. hong, i.-h. kwon, forwardűbackward analysis of rfid-enabled supply chain using fuzzy cognitive map and genetic algorithm, expert systems with applications ( ) ( ) – . [ ] x. huang, s. le, d. sharma, a taxonomy for rfid systems, in: ieee international conference on signal processing and communication systems, gold coast, australia, . [ ] f. bo, l. jin-ta, z. ping, g. jun-bo, d. zhen-hua, study of rfid middleware for distributed large-scale systems, in: information and communication technologies, . ictta ’ . nd, vol. , , pp. – . [ ] b. s., x. su, h. ramamurthy, c. cheng chu, r. gadh, winrfid ű a middleware for the enablement of radio frequency identification (rfid) based applications, in: in mobile, wireless and sensor networks: technology, applications and future, john wiley & sons, inc, , pp. – . [ ] h. k. chow, k. l. choy, w. lee, k. lau, design of a rfid case-based resource management system for warehouse operations, expert systems with applications ( ) ( ) – . [ ] eun-jun, yoon, improvement of the securing rfid systems conforming to epc class generation standard, expert systems with applications ( ) ( ) – . [ ] m. harrison, epc information service–data model and queries, tech. rep., auto–id centre institute for manufacturing, university of cambridge ( ). [ ] c. floerkemeier, d. anarkat, t. osinski, m. harrison, pml core specification . , tech. rep., auto–id centre institute for manufacturing, university of cambridge, ( ). [ ] epcglobal, epc information services (epcis) version . specification, tech. rep. ( ). introduction related work depcas mdm mlm guv epcis a depcas prototype mdm implementation mlm implementation guv implementation epcis implementation an example of depcas scenario conclusions credit scoring using the clustered support vector machine credit scoring using the clustered support vector machine terry harris ⇑ credit research unit, department of management studies, the university of the west indies, cave hill campus, p.o. box , barbados a r t i c l e i n f o article history: available online september keywords: credit risk credit scoring clustered support vector machine support vector machine a b s t r a c t this work investigates the practice of credit scoring and introduces the use of the clustered support vec- tor machine (csvm) for credit scorecard development. this recently designed algorithm addresses some of the limitations noted in the literature that is associated with traditional nonlinear support vector machine (svm) based methods for classification. specifically, it is well known that as historical credit scoring datasets get large, these nonlinear approaches while highly accurate become computationally expensive. accordingly, this study compares the csvm with other nonlinear svm based techniques and shows that the csvm can achieve comparable levels of classification performance while remaining relatively cheap computationally. � elsevier ltd. all rights reserved. . introduction in recent years, credit risk assessment has attracted significant attention from managers at financial institutions around the world. this increased interest has been in no small part caused by the weaknesses of existing risk management techniques that have been revealed by the recent financial crisis and the growing demand for consumer credit (wang, yan, & zhang, ). address- ing these concerns, over past decades credit scoring has become increasingly important as financial institutions move away from the traditional manual approaches to this more advanced method, which entails the building of complex statistical models (huang, chen, & wang, ; zhou, lai, & yu, ). many of the statistical methods used to build credit scorecards are based on traditional classification techniques such as logistic regression or discriminant analysis. however, in recent times non-linear approaches, such as the kernel support vector machine, have been applied to credit scoring. these methods have helped to increase the accuracy and reliability of many credit scorecards (bellotti & crook, ; yu, ). nevertheless, despite these advances credit analyst at financial institutions are pressed to continually pursue improvements in classifier performance in an attempt to mitigate the credit risk faced by their institutions. however, many of the improvements in classifier performances remain unreported due to the proprietary nature of industry led credit scoring research which attempts to find more efficient and effective algorithms. in the wider research community, the recent vintages of non-linear classifiers (e.g. the kernel support vector machine) have received a lot of attention and have been critiqued for, inter alia, their large time complexities. in fact the best-known time complexity for training a kernel based support vector machine is still quadratic (bordes, ertekin, weston, & bottou, ). as a result, when applied to credit scoring substantial computational resources are consumed when training on reasonably sized real world datasets. accordingly, efforts to develop and apply new classifiers to credit scoring, which are capable of separating nonlin- ear data while remaining relatively inexpensive computationally, are well placed. this paper investigates the suitability for credit scoring of a recently developed support vector machine based algorithm that has been proposed by gu and han ( ). their clustered support vector machine has been shown to offer comparable performance to kernel based approaches while remaining cheap in terms of computational time. furthermore, this study makes some novel adjustments to their implementation and explores the use of radius basis function (rbf) kernels in addition to the linear kernel posited by gu and han. the remainder of this paper is presented as follows. section outlines a brief review of the literature concerning the field of credit scoring and sets the stage for the proposed cvsm model for credit scoring that is presented in section . the details of the historic clients’ loan dataset and modeling method are highlighted in section . section presents the study results, and section dis- cusses the findings, presents conclusions, and outlines possible directions for future research. http://dx.doi.org/ . /j.eswa. . . - /� elsevier ltd. all rights reserved. ⇑ tel.: + ( ) ; fax: + ( ) . e-mail address: terry.harris@cavehill.uwi.edu this has been applied because credit-scoring data is often not linearly separable. expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.crossref.org/dialog/?doi= . /j.eswa. . . &domain=pdf http://dx.doi.org/ . /j.eswa. . . mailto:terry.harris@cavehill.uwi.edu http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa . background and related works . . overview credit scoring has been critical in permitting the exceptional growth in consumer credit over the last decades. indeed without accurate, automated credit risk assessment tools, lenders could not have expanded their balance sheets effectively over this time. this section presents a brief review of the relevant literature that has emerged in this space. . . what is credit scoring? credit scoring can be viewed as a method of measuring the risk attached to a potential customer, by analyzing their data to deter- mine the likelihood that the prospective borrower will default on a loan (abdou & pointon, ). according to eisenbeis ( ), hand and jacka ( ), and hand, sohn, and kim ( ) credit scoring can also be described as the statistical technique employed to con- vert data into rules that can be used to guide credit granting deci- sions. as a result, it represents a critical process in a firm’s credit management toolkit. durand ( ) posited that the procedure includes collecting, analyzing and classifying different credit ele- ments and variables in order to make credit granting decisions. he noted that to classify a firm’s customers, the objective of the credit evaluation process, is to reduce current and expected risk of a customer being ‘‘bad’’ for credit. thus credit scoring is an important technology for banks and other financial institutions as they seek to minimize risk. . . related works over the years, the demand for consumer credit has increased exponentially. according to steenackers and goovaerts ( ), this increase in the demand for credit can be attributable to the increased levels of consumption and the reliance on credit to sup- port this activity. in the united states, this rising level of consum- erism followed the introduction of the first modern credit card in s, so that by the s over % of american households owned a credit card. crook, edelman, and thomas ( ) posited that by this time, in the us, the total amount of outstanding con- sumer credit was over $ billion. comparatively, at the end of june this figure had risen to a staggering $ billion, a % increase (bgfrs, ). henley ( ) noted that the increasing demand for consumer credit has led to the development of many practical the scoring models, which have adopted a wide range of statistical and non- linear methods. similarly, mays ( ) posited that a number of various techniques have been used to build credit scoring applica- tions by credit analyst, researchers, and software developers. these techniques have included; discriminant analysis, linear regression, logistic regression, decision trees, neural networks, support vector machines, k-means, etc. in recent times, the use of more complex non-linear techniques, such as neural networks, and support vector machines, to build credit scoring applications has seen significant increases in the reported accuracy and performance on benchmarking datasets (baesens et al., ). irwin, warwick, and hunt ( ) and paliwal and kumar ( ) both provide evidence that advanced statistical techniques yield superior performance when compared to traditional statistical techniques, such as discriminant analysis, probit analysis and logistic regression. masters ( ) also provided evidence that the use of sophisticated techniques, such as neural networks, was essential because they had the capability to more accurately model credit scoring data that exhibits interactions and curvature. however, as pointed out by hand ( ) the increased performance of these more advanced tech- niques could be illusionary and if real, diminished due to shifts in the class distribution over time. the following sub-sections pres- ent a brief discussion concerning some of the classical and advanced statistical models used for credit risk assessment. . . discriminant analysis in his seminal paper, fisher ( ) proposed the use of discrim- inant analysis to differentiate between two or more classes in a dataset. since that time, durand ( ) and altman ( ) have both applied fisher’s ( ) discriminant analysis to credit scoring. durant used discriminant analysis to assess the creditworthiness of car loan applicants, while altman used it to explore corporate bankruptcy proposing his popular z-scores (altman, ). in works published separately by desai, crook, and overstreet ( ), hand and henley ( ), hand, oliver, and lunn ( ), sarlija, bensic, and bohacek ( ), and abdou and pointon ( ), they showed that discriminant analysis is indeed a valid technique for credit scoring. hand and henley ( ) noted that discriminant analysis, a parametric statistical technique, was well suited to credit scoring because it was designed to classify groups and variables into two or more categories or discriminate between two groups. however, saunders and allen ( ) noted that with this type of method certain assumptions about the data must be met. these assumptions include, normality, linearity, homoscedas- ticity, non-multicollinearity, etc. falbo ( ) and sarlija et al. ( ) posited that despite these limitations, over the years this technique has been frequently applied to build credit scoring applications, and it remains one of the most popular approaches taken today when classifying customers as creditworthy or un- creditworthy. several authors have criticized the use of discriminant analysis in credit scoring. eisenbeis ( ) point-out a number of the statis- tical problems in applying discriminant analysis to credit scorecard development. these problems include the following: group defini- tion, classification error prediction, estimating population priors, and the use of linear functions instead of quadratic functions, to mention a few. nevertheless, greene ( ) and abdou ( ) noted that despite these limits, discriminant analysis is one of the most commonly used techniques in credit scoring. . . . linear regression another popular classical statistical technique applied to credit scoring is linear regression. this method has developed into an essential component of data analysis in general and is concerned with describing the relationship between a dependent variable and one or more independent variables. thus, customers’ historical payments, guarantees, default rates and other factors can be ana- lyzed using linear regression to set up a score for each factor, and compare it with the bank’s cut-off (threshold) score. hence, only if a new customer’s score exceeds the bank’s cut-off score will credit be granted (hand & jacka, ). in its basic form, linear regression used for credit scoring requires the establishment of a threshold score. this threshold credit score is derived from the relationships between the firm’s historic clients’ features and their associated weights. as can be seen in the linear equation, z ¼ h þ h x þ h x þ�� �þ hnxn, where the variable n denotes the number of features collected from past and potential clients. these features are represented by the x’s, which are multi-dimensional vectors in rm, where m denotes the number of clients in the historical clients’ database. the h’s repre- sent the weights, and the feature variables and their weights used to calculate a credit score, z r, thus when an applicant scores t. harris / expert systems with applications ( ) – https://isiarticles.com/article/ southern journal of agricultural economics december, xlayer: an expert system providing management advice to commercial layer managers ed schmisseur and john pankratz abstract expert or knowledge-based system pro- xlayer, an expert/knowledge-based mi- grams contain domain-specific knowledge and crocomputer program, was designed and de- use complex inferential reasoning to reach veloped to diagnose layer management prob- conclusions human experts would reach if faced lems and recommend expert remedial man- with a comparable problem (hart; weissand agement advice. the program also provokes kulikowski). the basic knowledge contained management action by calculating the eco- in these programs consists of rules, facts, re- nomic loss attributed to major management lationships, reasons, and heuristics obtained problems. it analyzes data generated by a from human experts who can solve problems commercially marketed layer performance fi- in a particular domain of expertise. expert nancial microcomputer program and has dem- systems embody techniques for solving prob- onstrated the ability to emulate poultry man- lems, manipulating stored knowledge, coping agement experts in the diagnoses of indi- with uncertain information, and explaining how vidual layer management problems. the pro- an inference is proceeding or a conclusion has gram provides scarce expert poultry manage- been reached (simons). ment advice to poultry layer managers regard- expert or knowledge-ased system program less of size and scale of operation. developments are a new area of applied poul- try science research. the results of one effort to apply expert system technology to poultry key words: expert system, knowledge-based layer management are reported here. system, rule-based, economic loss, layer management.l, lr m. problem-solving foundations expert or knowledge-based computer pro- the practical and empirically verifiable di- grams are becoming an important operations agnostic rules of poultry layer management and management component of many fortune as practiced by a cooperating poultry manage- companies (simons; herrod) and are being ment consulting firml served as the problem- actively explored to assist smaller businesses solving paradigm used in the xlayer pro- (harmon et al.). they are beginning to emerge gram. these domain-dependent methods were as a field of research, development, and use in developed by consulting firms and have proven commercial agriculture (mckinion and useful over a period of years through practical lemmon; michalski et al.; boulanger; roach consulting experience. collectively, these diag- et al.; fermanian et al.; spahr and puckett; nostic rules implicitly seek to maximize layer levins and varner; oltjen et al.). they also flock profitability by identifying problems represent the next logical step in the progres- which impact production performance, costs, sion of computer-supported information and and returns. they utilize problem-solving tech- decision-aid systems for farm managers first niques and data manipulations which explic- initiated by agricultural researchers in the itly and rigorously critique these separate but early s. (schmisseur and doluschitz). interrelated areas of layer management. i the management expertise contained in the xlayer program represents that of mel gehman, president and founder of heritage pms, inc., rd # box -b, annville, pennsylvania . ed schmisseur is an associate professor, and john pankratz is a courtesy appointment, department of agricultural and resource economics, oregon state university. approved for publication as technical paper no. of the oregon agricultural experiment station. this project was financed in part by the department of agricultural and resource economics and the oregon agricultural experi- ment station, oregon state university and heritage pms, inc., of annville, pa. copyright , southern agricultural economics association. the general problem-solving approach em- financial performance, and egg gradeout. flock bodied in the xlayer program is relatively performance data contain livability, egg pro- straightforward. it attempts confirmation of duction, and nutrition information such as mor- all management problems contained in the know- tality, body weight, house temperature, egg ledge base by matching a list of management case weight, shell thickness, metabolizable problem symptoms with indicators of substan- energy, and protein, and the content of spe- dard flock performance. when a critical num- cific amino acids in the layer ration. financial ber of matches occurs, a management problem performance indicators report itemized income is confirmed and a management recommen- and expense categories and net profit. egg dation is issued. the search for management gradeout data show egg size and grade, qual- problems begins with the evaluation of pro- ity, quantity, and price information. duction decisions followed by decisions about the xlayer program uses approximately egg marketing, and then production costs. different data observations generated by the diagnostic rules, general problem- the layer performance program. these data, solving approach, and search strategy embod- automatically stored in a file called "record," ied in the xlayer program were solicited represent both actual layer performance and from the cooperating consultants. a series of equivalent performance potentials or stan- interviews with these experts dards. performance potentials or standards are ) identified important management problems based on standards published by commercial impacting layer flock profitability; genetic companies and breeder farms. they ) dictated the data and the heuristics needed reflect the different layer strains, age of birds, to diagnose these problems; molting phase, and environmental conditions. ) established the analytical or step-by-step other data required by the xlayer pro- procedure followed to determine problems; and gram, but not produced by layer perform- ) ascertained what management recommen- ance, include information on pullet flock his- dation would be issued. tory, layer house equipment, marketing ar- after each interview, new insights about the rangements and prices, and criteria by which diagnostic process were encoded in the flock performance are compared to perform- xlayer program and then tested using test ance standards. pullet history information in- case data. in subsequent interviews, previ- cludes pullet weights, feed consumption, uni- ously discovered problem-solving rules were formity and shank index information at se- confirmed, new problem-solving capabilities lected weeks of rearing, and general pullet were added, and the depth of the program's rearing conditions. this information is manu- diagnostic capabilities was increased. ally entered and permanently stored in a file called "flock profile" when a pullet flock starts data requirements production. the xlayer program was developed to information about layer house equipment analyze weekly layer flock performance data includes the type of feeding, egg gathering, generated by a commercially-marketed poul- and manure handling system in the layer try management microcomputer program, house. these data, also manually entered, are layer performance. this integration was permanently stored in a file called "housing/ important since the consulting firm developed equipment profile." layer performance to collect and calculate information about marketing and prices in- data useful in their diagnoses of layer man- eludes egg selling/processing arrangements, agement problems. furthermore, many of egg prices by egg grade or case weight, and xlayer's data needs could be obtained di- feed ingredient prices. these data are speci- rectly from data files produced by layer per- fled in a file called "price profile." price infor- formance rather than through additional key- mation is manually updated weekly. board entries. the various criteria by which flock perform- the major types of data produced by layer ance are compared to flock performance stan- performance and utilized by the xlayer dards are contained in a file called "evalu- program are illustrated in figure . the layer ative criteria." these criteria were obtained performance program produces a report which from the cooperating consulting firm. they are provides weekly data on three major areas of based on experience in analyzing flock per- poultry management-flock performance, formance records. layer performance is a registered trademark of heritage pms, inc., rd # box -b, annville, pennsylvania . layer report flock egg performance gradeout egg income profit size/price production grade livability nutrition expenses quality figure . heritage's layer performance report. program overview infer management problems based on other known information. at each query point, us- the xlayer program: ers can ask the xlayer program to explain ) identifies major management problems its reasons for the query. also, at the end of judged to be impacting layer flock profitability; the consultation, users can query the program ) calculates associated economic losses at- to explain how it arrived at its conclusions. tributed to major management problems; and after the xlayer program has completed ) issues specific management recommenda- its analysis, an "executive management re- tions designed to eliminate the management port" is compiled. this report includes a brief problem. it is written using a commercially- description of each management problem, its marketed expert development and delivery associated economic loss to provoke manage- shell microcomputer program, m. . the ment action and facilitate quick partial bud- xlayer program' s code was written in the geting of the decision recommendation, and "prolog"-like production rule language con- specific management recommendations. if no tained in the m. program. the shell program management problems exist, the xlayer requires an ibm-pc compatible microcom- program generates a brief report indicating puter equipped with a graphics card and a this finding. minimum of k ram. the xlayer program's relationship to its diagnostic framework data needs and users is illustrated in figure . users execute and communicate with the a generalization of xlayer's diagnostic xlayer program by interacting with m. . framework is presented in figures through the m. program automatically calls the . these diagnostics are contained in some xlayer program and immediately begins individual production rules. diagnostics per- the consultation. during consultation, the tain to egg production, egg blend price, vari- xlayer program automatically seeks rele- able production costs, and fixed costs. the di- vant flock performance, performance stan- agnostic framework, or fault tree, illustrates dards, and other needed data. should data not the complexity of the inference process. it also be found, the natural language user interface shows the path of impact of management of m. queries the user. users then must re- decisions made in each specific area of pro- spond. a response of "unknown" is acceptable duction. as the xlayer program has the ability to management decisions about layer nutrition layer performance report <- arecords & standards ^ —pro |flock profile u se r— housing/equipment profile m. m. - |price profile xlayer[_ itxlayer ievaluative criteria a- i' questions i . | management report figure . xlayer's relationship to users and data. m. is a registered trademark of teknowledge, inc., p.o. box , palo alto, california . depict one of the complex inferences within plementation problem can manifest itself in the xlayer program. nutritional decisions several performance indicators. impact egg production directly and also indi- when a methionine problem is confirmed, rectly through feed consumption (figure ). another xlayer rule is invoked which actu- but, nutrition decisions also indirectly impact ally issues the management alert and pre- egg blend price through both small eggs and scribes a remedy. this rule is: egg quality (figure ) and indirectly impact variable production cost through feed costs f methionine and display ([verify suspected (figure ). specifically, egg quality is directly methionine imbalance in the layer ration. affected by shell quality, which in turn is im- methionine intake should be within per- pacted by nutrition decisions, and feed cost is cent of your flock's potential intake. if con- directly affected by feed consumption, which firmed reformulate the rations protein and in turn is impacted by nutrition. similar com- methionine content. plex inferences exist for management decisions then problem. affecting environmental conditions in the layer however, before the primary rule can be house, disease, pullet quality, and stress. confirmed, either low feed consumption and the complexity of the diagnostics contained low production or small egg sizes must also be in the xlayer program is more easily seen confirmed from other production rules con- in its diagnostic rules. for example, the pri- tained in the xlayer program. mary rule confirming a methionine supplemen- the low feed consumption, low production, and tation problem exhibits a complex inference. small egg sizes diagnostic rules are relatively this rule and others presented also provide simple. the low feed consumption rule is: insights into the basic analytical process or problem-solving approach employed in the iffeedperhenay = fphd xlayer program. the rule, taken directly and goalfeedperhenday = gfphd from the xlayer code, is: and critical_feed_perhen_day = cfphd and fphd < gfphd * cfphd if ((low_feed_consumption and low_production) then low_feed_consumption. or (low_feed_consumption and small_egg_sizes ois unkn( onsumption and smthis rule confirms low feed consumption whenis unknown) and methionineperhenday = mphd the amount of feed consumed per hen day is and goalmethionineperhenday = gmphd less than the goal or standard for feed con- and uppertwidebounds p uw b sumption per hen day adjusted downward by and louer wide bounds = lwb coefficient of tolerance for feed consumption. and mphd > gmphd * uwb the low production rule is: or mphd < gmphd * lwb) if eggs_per_hen_housed = ephh or (small_egg_sizes and goal_eggs_perhen_housed = gephh and methionine_per_hen_day = mphd and critical_eggs_per_hen_housed = cephh and goal_methionine_per_hen_day = gmphd and ephh < gephh * cephh and upper_moderate_bounds = umb and egg_blend_price = ebp and lower_moderate_bounds = lmb and starting_number_hens = snh and mphd > gmphd * umb and (((gephh - ephh) * snh)/ ) * ebp = or mphd < gmphd * lmb) prodl then methionine. and display( [' the rule's interpretation is: economic losses attributed to low produc-the rule's interpretation is:the rule '£ mterpretat ' . n tion are: ',tab( ), $(prodl)] ) ) if low feed consumption and low produc- i r tion exists and the methionine level fed per then lowproducton. hen day exceeds either an upper or lower it confirms low production when egg produc- bound, tion per hen housed is less than the standard ) or small egg sizes exists without either for production adjusted downward by a coeffi- low feed consumption or low production and cient of tolerance for egg production, and it the methionine level fed per hen day ex- calculates the associated economic loss. this ceeds either a narrower upper or lower loss is based on the difference between the bound, egg production standard and actual egg pro- then a methionine supplementation problem duction per hen housed adjusted by the total is confirmed. it shows that a methionine sup- number of hens housed. egg production quality conditions consumption disease pullet disease fright nutrition parasites quality parasites environmental ration mechanical environmental cannibalism conditions formulation equipment conditions prolapse figure . egg production diagnostic framework of xlayer. egg blend price market small practices egg quality eggs i shell interior quality quality equipment stress disease candling stress off taste damage error discoloring mnvroma imin ral [ handling & environmentalnutrition disease equipment conditions absorption problems low feed parasites pullet figure . egg blend price diagnostic consumption nutrition diseases quality framework of xlayer. high variable cost medication high energy miscellaneous labor wage labor wage costs costs feed costs costs rate hours costs feed ingredient consumption costs i interest & repairi | + | +ldepreciation costs environmental feather conditions wasteage loss nutritional theft figure . variable and fixed cost diagnostic framework of xlayer. the small egg sizes rule is: in new grains gradually even if the cost if caseweight = cw per pound is higher. gradually move to and goal_caseweight = gcw the lower cost grain substitute.'] ) and critical_case_weight = ccw then problem. and cw < gcw * ccw and caveragewblendrice = abp notedly absent from the diagnostic frame- and dozen_eggs_graded = deg work are interrelationships between such fac- and potential blend_price = pbp tors as nutrition, disease, and stress. although and (pbp - abp) * deg = smalleggs these interrelationships are theoretically pos- and display( [' sible, the cooperating consultants did not re- economic loss attributed to small egg sizes quire analyzing these type of interrelationships is approximately,',tab( ), $(small in their consulting work and did not include eggs), per week.'] ) them in the xlayer program. they argued then small_egg_sizes. in today's skillfully managed commercial poul- try flocks, management problems attributed this rule confirms small egg sizes when egg to these type of interrelationships are rare. case weight is less than a case weight stan- dard adjusted downward by a tolerance level program outputs for egg case weight. it also calculates the as- t x p sociated economic loss due to small egg sizes . the xlayer program possesses the abil based on the difference between the egg blend t to dagnose ultiple management problems price received and the potential egg blend andrecommendmanagementactions formore price multiplied by the number of eggs graded than individual layer production manage- another type of diagnostic rule contained in ment problems. this capability is exhibited in the xlayer program reflects a complex in- a management report displayed on screen and/the xlayer program reflects a complex in- or printed. the report identifies management ference requiring the user to respond to a prte. the reort idifis management query about current production conditions. a problems, their associated economic loss, and simple example of this type of rule is: specifc management recommendations. an example management report appears in fig- if (low_feed_consumption and low_production) ure . specific management problems and rec- or (low_feed_consumption and ommendations related to housing and equip- small_egg_sizes) ment management, nutrition, diseases, eco- and grain_change = yes nomics, marketing practices, and general then change_grain. management practices are contained in the as this rule is being confirmed and the pre- program. conditions of either low feed consumption and approximately percent of the xlayer low production or low feed consumption and program's diagnostics are about layer nutri- small egg sizes exist, the user is automatically tion nutritional diagnostics include such queried about possible recent changes in the things as excessive or restrictive metabolizable type of grain used in the layer ration. the energy, calcium, and sodium supplementation, query involves two separate rule statements methionine, methionine-cystine, lysine, thre- onine, and trytophane amino acid imbalance, question(grain_change)=' improper vitamin and/or trace mineral mix have you significantly changed the type supplementation, and poor egg shell quality of grain used in your layer ration?'. induced by improper ration formulation. legalvals(grain_change) = [yes,no]. some percent of the xlayer program's the first of these two query rules issues a diagnostics directly address general manage- prespecified question to the user, while the ment practices. improperly trained egg can- second provides acceptable answers and error dlers, a poorly managed poultry waste sys- checking capability to the query. tem, cage overcrowding or under-utilization, if the user responds to the questions by an- sudden changes in layer house environment, swering yes, another xlayer rule is invoked excessive egg handling time, and egg or feed which states the management problem and of- theft represent some of the type of general fers a remedy. this rule is: management problems considered by the xlayer program. if changegrain and display( [' housing/equipment diagnostics account for a major change in the grain ration is about percent of the xlayer program's suspected to be causing production prob- management expertise. the xlayer pro- lems. reformulate your ration and phase gram identifies problems like improperly ad- justed or operating mechanical egg gathering program testing/validation and feeder equipment, high or low layer house the xlayer program has been subject to temperature, low or high water consumption controlled testing with numerous test case due to malfunctioning waterers or poor well data. during initial case testing, the xlayer water quality, limited lighting, and inadequate program's diagnostic capabilities were further ventilation. veno dti crlastion. p enhanced, and its depth of knowledge was in- economic diagnostics, representing per- cresed. more than individual cases, rep- cent of the xlayer program's capability, resenting a flock's complete weekly data set, address traditional economic performance were subjected to analysis by the xlayer problems. these include such things as high program. these cases, although carefully con- repair and maintenance, pullet rearing, labor, structed, confirmed that the program could interest, and energy costs. successfully diagnose each, as well as various disease and marketing problems represent practical combinations, of the more than the remaining diagnostic powers of the pro- layer management problems included in the gram. high broker fees, inappropriate class xlayer program weight sales contract, respiratory diseases, currently, it is being intensively field tested parasites, infectious coryza, fowl cholera, and with flock data from both the east and west mg disease are some of the disease and mar- coasts. in these tests weekly flock perform- keting diagnostics addressed by the xlayer ane data are being analyzed by the xlayer program. program and its diagnostics are compared to program limitations pthose independently developed by the poultry management experts being emulated by the the xlayer program has the potential to xlayer program. in these limited field tests be a useful management tool, yet it has limita- with some , hens in various flock sizes, tions. major limitations relate to the program's representing approximately flock weeks of knowledge base. xlayer's knowledge base production data, the xlayer program has is static since the program does not possess exhibited the promising ability to perform on the ability to learn from experience. as a re- a level consistent with that of its counterpart sult, xlayer's knowledge base must be con- poultry management experts. that is, in ap- tinually updated as new diagnostic techniques proximately percent of the weekly tests, are developed and possibly new genetic bird the xlayer program's diagnostics were strains are introduced. the magnitude of this identical to those independently prescribed by limitation is unknown. it may be mitigated the poultry management experts. because new knowledge can be added to ex- additional field testing and more intensive pert system programs much more easily than validation with more participating flocks is modifying code in algorithmic programs. continuing. field testing is scheduled to be xlayer's knowledge base is also limited completed by the summer of . to the diagnostic knowledge of the cooperat- ing consultants. no measure of their poultry program availability management diagnostic skills has been made. the xlayer program will be made avail- the magnitude of this limitation is also un- able to poultry managers by the cooperating known. consulting firm after extensive field testing is economic losses attributed to low production are: $ per week. verify suspected high metabolizable energy level in layer ration. if confirmed, reformulate ration to reduce metabolizable energy level to your flock's standard recommended in the layer performance financial report. also note a methionine-cystine imbalance in the layer ration. methionine-cystine should be within percent of your flock's potential intake. if confirmed, reformulate the ration's protein content and methionine-cystine supplementation. water intake appears high. figure . xlayer's management report: an example. completed in the summer of . the cost of gram indicates that it can substitute for scarce the program has not been established. pro- and relatively expensive human, poultry layer gram code, excluding layer performance and management expertise. its successful adoption m. , can be obtained by contacting the authors and widespread use should increase the pro- of this article. duction efficiency of layer operations and im- rconclusions prove producer profits. the xlayer program also has shown that the present level of performance of the it can provide valuable management informa- xlayer program suggests that expert sys- tion in a readily useable form to a range of tem programs have potential applied applica- flock sizes. relative to economies of size is- tion in poultry layer management. further- sues, this expert system program appears to more, it has been demonstrated that an ex- be size neutral. that is, it equally benefits dif- pert system program can be linked to a stand- ferent sizes of operation. if any unequal bene- alone agricultural, computerized decision- fit should occur, it is hypothesized benefits support program and can automatically and favor smaller operations because of the econo- routinely access information generated by this mies associated with making available scarce program to diagnose management problems and expensive human expertise in the more and provide insightful management advice. economical form of an expert system micro- the level of performance of the xlayer pro- computer program. references boulanger, a.g. "the expert system plant/cd: a case study in applying the general pur- pose inference system advice to predicting black cutworm damage in corn." m.s. thesis, computer sci. dept., univ. ill., champaign-urbana, il, . fermanian, t.w., r.s. michalski, and b. katz. "an expert system to assist turfgrass manag- ers in weed identification." in proc. summer computer conference. july - , chicago, il, american society of agricultural engineers, st. joseph, mi, . - . harmon, p., r. maus, and w. morrissey. expert systems tools and applications. new york: john wiley and sons inc., . hart, a. knowledge acquisitionfor expert systems. new york: mcgraw-hill book co., . herrod, r.a. "industrial applications of artificial intelligence." in proc. ieee western conference on computer systems. june - , anaheim, ca, ieee computer society press, washington, dc, . - . levins, r.a., and m.a. varner. "an expert diagnostic aid for reproduction problems in dairy cattle." comput. electron. in agric., ( ): - . mckinion, j.m., and h.e. lemmon. "expert systems for agriculture." comput. electron. in agric., ( ): - . michalski, r.s., j.h. davis, v.s. bisht, and j.b. sinclair. "a computer-based advisory system for the diagnosis of soybean diseases in illinois." plant diseases, ( ): - . oltjen, j.w., l.g. burditt, and g.e. selk. "expert system for culling management of beef cows." okla. agric. exp. stn. rpt., mp- , . - . roach, j.w., r.s. virkar, m.j. weaver, and c.r. drake. "pomme: a computer-based consul- tation system for applied orchard management using prolog." expert syst., ( ): - . schmisseur, e., and r. doluschitz. "expert system insights: future decision tools for farm managers." j. am. soc. farm managers and rural appraisers, ( ): - . simons, g.l. expert systems and micros. manchester, england: the national computing centre limited, . spahr, s.l., and h.b. puckett. "electronics and livestock production." illinois review, ( ): - . weiss, s.m., and c.a. kulikowski. a practical guide to designing expert systems. totowa, nj: rowman and allanheld, . microsoft word - .doc configuration irrigation schedule based on expert systems and operations research iman hassen*, mohamed rasmy**, ahmed rafea*** *central lab for agriculture expert system, arc email: iman @ mail.claes.sci.eg **faculty of computer and information, cairo university email: m_h_rasmy @ hotmail.com ***computer science dept., auc email: rafea@aucegypt.edu abstract: this paper presents configuration design knowledge to solve irrigation scheduling problem in agriculture domain and a general approach for configuration management knowledge. it also introduces a configuration agriculture irrigation schedule (cais) tool that produces the configuration irrigation schedule design according to the commonkads methodology. a configuration that satisfy the user requirements, and that additionally satisfy some optimality criteria is required. by integrating dynamic programming and inference knowledge, with demand including different objective, that it is used in generating the inference steps (ppsm) to select the suitable components of the inference steps minimizing the number of components and/or maximizing the accuracy of the result of each inference step. operations research techniques greatly improved the understanding of the interactions of a large number of components and enhanced the decision selection under uncertain outcomes. copyright ® ifac keywords: configuration, operations research, expert systems, knowledge-based systems, knowledge acquisition, artificial intelligence, dynamic programming. . introduction recent approaches integrate both the operations research (or) and expert systems (es) together under a unify framework. although the benefits of applying or and es technology can be significant, developing real world extremely time consuming task. modeling an agriculture domain requires incremental modifications to candidate schedules need to be designed and implemented to generate a configuration design includes these two approaches. configuration design is a form of design where a set of pre-defined components is given and an assembly of selected components is sought that satisfies a set of requirements and obeys a set of constraints (mittal, s., & yu, b, ). sometimes an optimality criterion may be given that defines an ordering upon possible solutions. following chandrasekaran ( ) one could argue that a specification of the technology to be applied in the design process is also part of the input. in addition to the problem specification—knowledge is needed about the design domain (domain theory), additional requirements and/or constraints that are not explicitly given as input but that are implicitly given in regulations, common sense etc., optimality criteria and optionally certain preferences. preferences are design choices which are not the outcome of the design process itself, but which are given beforehand and constrain the space of design options. such design choices may be changed when no solution can be found within the constrained search space. we propose configuration agriculture irrigation schedule (cais) tool according to the commonkads methodology (wielinga, ) which could overcome some of the limitation found in previous research efforts in the following section, a configuration agriculture irrigation schedule (cais) components are described, in section three the processes of cais tool are presented, the conclusion is in section four. . configuration agriculture irrigation schedule (cais) componets a configuration agriculture irrigation schedule (cais) tool consist of three main components, namely agriculture schedule ontology, or library, and expert systems library. this tool produce one intermediate result called 'irrigation knowledge base' through the knowledge acquisition process, and it generate a configuration irrigation schedule design through the generation process as depicted in figure . the following subsections describe each component of the three main components in addition to the intermediate result and the final product of the tool. dt . agricultuure schedule ontology an ontology for a knowledge based system is an explicit specification for the objects, concepts and other entities that are presumed to exist in some area of interest as well as the relationships that hold among them (gruber t.r., ). in cais agriculture schedule ontology classified into five parts: resources, irrigation domain ontology (activity), irrigation schedule (product), user requirement (demand), and constraint. resources are an entity that supports or enables the execution of activities (stephen f. ). resources are generally in finite supply and their availability constraints when and how activities execute. in cais, we conceptualize the resource as consisting of objects (i.e. soil, water, climate, farm, previous crop, and plantation) that are elementary entities such as integers and string (i.e. soil salinity, soil texture). the activity is a set of processing steps required to produce or provide the product (irrigation schedule). it is the set of activities that fulfill the demand. the sequences of activity are expansion, et , etcrop, shwc, water requirement, and frequency/interval to achieve the user requirement. an irrigation domain ontology (domain model) defines declaratively the set of terms and relations of a domain independent of any problem solving method. the domain model contains static information about the application environment. . operations researsh library operations research library is a collection of algorithms on dynamic programming, linear programming, and pert/cpm techniques that used in irrigation schedule. the dynamic programming is used in generating the inference steps (ppsm) to select the suitable components of the inference steps minimizing the number of components and/or maximizing the accuracy of the result of each inference step. a dynamic programming formulation that used in generating inference steps described as follows: ft(st) = max/min (r(dt, st) + [ft (s t+ )]) subject to s t+ = gt (dt, st) st � st, dt � dt, t = , , ,….n where ft(st) is the maximum expected total return r(dt, st), t= , ,…….n finite stages of tasks, st is a finite state of a primitive problem solving method, dt, is a decision for each task, the state of the system in t+ is described by the generation function gt (dt, st). the objective function can be changed to minimize the components as follows: ft(st) = minimum component of ppsm (inference step) when it is in ppsm (state) s with n more tasks (stages) to generate the inference step. another dynamic programming formulation that used in a dialy irrigation task described as follows: ft(st)= max[(yt*pt - (cw*qt*it + it* ce * cl) + ft (s t+ )] t= , , , ………….t subject to ot = et (d, t, r) ( ) et = etcrop(o) ( ) user requireme knowledge engineer domain expert irrigation knowledge base knowledge acquisition psm library task library ppsm library dynamic pro. linear prog. pert/cpm operations research library meta d.m. re. library generation configuration irrigation schedule design agriculture schedule ontology irrigation library expert system. library operations research library operations research library operations research library figure : generating configuration irrigation schedule design d= , st = shwc ( ) wt = et + et- ( ) st + rt – wt = qt ( ) e = ( ) it = ( ) where ft(st) is the maximum profit of expected yield in ton, t= , ,….n finite stages of plant age per day, yt is the expected yield at certain stages, pt is the price per ton, cw is water cost; wt is the water requirement , it is an irrigate variable taking the value for irrigate or if not., ce represents energy costs for each irrigation, cl denote labor costs. ot is a value of et inference step that determined by current date, daily temperature and relative humidity that denoted by d, t, r respectively. et is a value of etcrop inference step that determined by plant age and et that denoted by t and ot respectively. st is a value of shwc soil available water inference step that determined by plant age. r is a rain, and qt is a dummy variables to take decision in { , }, means start irrigation today, means wait until to next day. the whole describtion of those formulations will be explained in section . . expert systems library expert system library is a collection of problem solving method (psm), tasks, primitive problem solving method (ppsm), meta domain knowledge representation, and irrigation libraries. the roles that need to be filled by a psm prescribe what domain knowledge is expected from the expert. the domain knowledge is, therefore, no longer something that is acquired and represented independently of how it will be used in concrete problem solving. however, it is still explicitly and separately represented, giving the advantage of maintainability and systematic ness in knowledge acquisition. this library contains nine psms irrigation schedule the task reflects the general problem solving method for scheduling. a problem solving method can be defined as a knowledge-use-level characterization of how a task might be solved. it can also be seen as an abstract model of how to solve certain problem. in irrigation schedule design, a generic problem solving method called “propose and revise� is based on mathematics model and heuristic model. irrigation task library contains tasks for irrigation schedule for crop. irrigation is the main task of the irrigation generates design. the goal of the irrigation task is to get irrigation schedule of the case description. its inputs are the case description. its output is the irrigation schedule of the case description. . irrigation knowledge base the irrigation knowledge base is the result of the knowledge acquisition process. it contains all the related knowledge for the crop, evapotranspiration (et ) method used, etcrop methods used, available water in soil methods used, and water requirement methods used. . irrigation schedule design irrigation schedule design is the final product to satisfy the user requirement. there three main components of the irrigation schedule design: task knowledge, inference knowledge, and domain knowledge as described in the following subsections. . . irrigation task knowledge the design of task knowledge consists of two parts: the task definition and the task body. the task definition describes the main system goal, as well as the input and the output roles. the task body describes the task type (composite or primitive), subtasks (if any), additional roles (which are not included in either input roles or output roles), and the control over these subtasks. figure shows a generated part of task knowledge of irrigation. it shows that the irrigation task contains only one task, namely: 'propose mathematic irrigation schedule'. the input is case-description and the output is proposing irrigation schedule. type of task is composite. the irrigation schedule subtasks involves two transfer tasks, one primitive task, and one composite task namely: 'get case description', 'display irrigation schedule', 'initialize irrigation parameters', and 'compute propose irrigation schedule' respectively. there are also two task static primitive (inference step) are expand and irrigation units. qt > and et = qt <= , task: irrigation schedule , task-definition: goal: the main goal of the irrigation is to determine the water requirement during cultivation. input: case-description output: irrigation schedule task_body type: composite sub_tasks: propose mathematic irrigation schedule control_structure: propose mathematic irrigation schedule. task: propose mathematic irrigation schedule task-definition: goal: generating an propose irrigation schedule input: case-description output: propose irrigation schedule task_body type: composite sub_tasks: get case description , determine plant parameters, initialize irrigation parameters, compute propose irrigation schedule, irrigation units, display irrigation schedule control_structure: get case description , case description, expansion model -> expanded case description, initialize irrigation parameters, compute propose irrigation schedule, water requirement, unit model -> irrigation unit, display irrigation schedule. figure : task knowledge of irrigation . . inference knowledge the inference describes how domain knowledge is used in elementary reasoning steps (or inferences). the inference knowledge is described using two views namely: inference structure and inference specification (breuker j. et al., ). the design of inference knowledge consists of two main parts namely: inference structure and inference specification. an inference structure diagram is used to describe this type of knowledge. three types of components are used in this diagram, namely: inference steps, dynamic roles, and static roles. an inference specification contains five parameters: name, static role, input role, output role, and specification. a name of inferences represent the role that these inferences play in solving the problem, static roles indicate which domain model should be accessed, input and output roles are supposed to be a part of the overall working memory of the problem solver and are thus not directly linked to specific domain model. the first inference expand (see figure ) uses the expansion model to derive a new data using the available data included in case description role. . . domain knowledge the domain knowledge layer consists of two parts: the set of ontology related to the application domain and domain models. domain ontology is considered as the most fundamental constituent of domain knowledge. it defines the language of the domain, i.e., the terms used to describe the domain. this language consists of a number of constructs such as concepts, expressions, and relations (gruber t.r., ). the word “ontology” can be defined as �what exist in the real life application". generally speaking, the scheduling ontology should be mapped to five classes of concepts: resources, irrigation domain ontology (activity), irrigation schedule (product), user requirement (demand), and constraint. a domain model is defined as a coherent collection of expressions about a domain that represents a particular viewpoint defined in ontology. the domain model may therefore embody certain assumptions that are specific for the ontology that is used. domain model have an internal structure for knowledge representation. irrigation schedule domain model includes domain models. the irrigation domain model are expansion model, evapotranspiration(et ) model, etcrop model, water requirement model , shwc model, frequency_interval_model, and constraint model. each domain has its classification (i.e evapotranspiration (et ) model classify into et penman method, eto modified penman method, et penman mountance method, et hargrve method, et blaney method, et radiation method, et penman monteith method, et pan evaporation method, et reference based on sector, eto reference based on governor, and et reference based on directora.). each domain model classification contains a parts which describe the knowledge representation for this method. for example, et hargrve method consists of four parts: )eto hargerve for current month, ) eto hargerve for next month, ) smooth factor, and ) adaptive eto hargerve . configuration agriculture irrigation schedule (cais) processes there are two main processes in the cais tool, the first one is a knowledge acquisition, and the second is generation as shown in figure . the first process produces the irrigation knowledge base, and the second one produce the irrigation schedule design. the next two subsections describe those processes. . knowledge acquisition the irrigation knowledge base is the result of the knowledge acquisition process. in the knowledge acquisition process, the knowledge engineer collecting the user requirement and eliciting the knowledge using the agriculture schedule ontology and the expert system library by using a special knowledge acquisition tool that consists of five parts crop parameters, evapotranspiration (et ) method, etcrop methods, available water in soil methods, and water requirement methods. after the knowledge engineer collecting the knowledge relate to crops, the domain expert uses this tool to verify the crop parameters. . generation the generation starts with the requirements definition where several questions are used to guide the generations of the irrigation schedule design for any crop. the criteria for generating design describe by the property (attribute) and the configuration design for irrigation system is used to collect data on: o generation model (e.g. mathematical, heuristic...) name :expand goal :derive parameters used in the system according to the available data involved in the case description role static role : expansion model input role :case description output role:expanded case description spec :soil.type : relation plant age : function figure : inference knowledge of irrigation o schedule type(daily, weekly, each- -day, monthly) o the objective of irrigation management (i.e. maximize net return, minimize irrigation costs, maximize yield, optimally distribute a limited water supply) o scheduling problem solving containing propose and revise (i.e., propose, revise) o events for reactive scheduling (i.e., forest, wind, rain...) o disease (i.e., fungal...) o yield response to water (i.e. forecasting expected yield) o minimize the component of design (yes, no) o evaporation method (et penman, eto modified penman, et penman mountance, et hargrve, et blaney, et radiation, et pan evaporation, et reference sector, eto reference governor) the process generation is decomposed into four processes namely: generate tasks, generate inference, generate domain model, and generate ontology to produce the irrigation schedule design that contains four parts: tasks, inference, domain model, and ontology as shown in figure . the next subsections describe each one of them. . . generate task irrigation schedule task has sub-tasks. the set of sub- tasks of a specific method have some sort of part-of relationship with the method. each sub-task might itself be realized by different methods. thus, the task-method decomposition is recursive until primitive inferences (or primitive tasks in our terminology) are reached, calling the mechanism of generating task by the "irrigation task schedule". there are four type of tasks: primitive, transfer, static primitive, and composite task. in case of composite task, the generate task mechanism calls itself recursively to get the sub of problem solving method or primitive problem solving method. the irrigation scedule task contains two psms are 'propose and revise irrigation schedule mathematical model' and 'propose and revise irrigation schedule heuristic method'. it must select only one psm based on the user requirement. the mechanism generate a set of task using psm library, task library, or library, and user requirement. consultring the task library and match each selected task stored in the dynamic list of generated tasks to write its knowledge in the task knowledge. . . generate inference by using one of the or library called dynamic programming to select inference step. the list of primitive tasks that generated in the priviouse section are considered as a stages. given a sequence of task (stages) t through tn to be executed in linear order, where n is the number of task for specific psm [t = {et , etcrop, pawc, interval, water requirement,…..}], where ti � t. primitive problem solving method (inference step) is used to describe the natural property of stages (states), each task contains more than one ppsm, it must select only ppsm (inference step). the finite ppsm of the system is denoted by st [set = { et penman method, eto modified penman method, et penman mountance method, et hargrve method, et blaney method, et radiation method, et penman monteith method, et pan evaporation method, et reference based on sector, eto reference based on governor, and et reference based on directorate.}]. the decision at each task (stage) is � psm library task library� ppsm library meta d.m. re. library generate task user requireme nt� knowledge� engineer domain expert agriculture schedule ontology irrigation library generate inference generate domain model knowledge acquisition set of tasks task knowled ge inference knowled set of inference domain models irrigation knowledge base ontology or library (dynamic programming) generate ontology. figure : detail process of generating configuration irrigation schedule design function domain_model_ generation. begin get list_of_inference from inference_db(ppsm) while(not empty list_of_inference) begin get name_ppsm from list_of_inference get static_role of name_ppsm store static_role in domain_model_db(domain_model) end{ while } get production_schedule.system = system get list_of_dm from domain_model_db(domain_model, system) while(not empty list_of_dm) begin get domain_model from list_of_dm get system from list_of_dm getdomain_model_represention(system,domain_model, _, representation, case representation relation: procedure generate_relation_dm(table name(item), input_output, position text, text) table: procedure enerate_table_dm(domain_model, system) function: procedure generate_function_dm( domain_model, system) end{case} end{ while } figure : domain model algorithm denoted by dt. when the process comes to ppsm (states) of some task, it can make different decision to determine the ppsm of the next task. the value of dt(st) is used to denote the decision made of ppsm st of task t. the state generation inference knowledge of the cais system in t+ is described by the function gt(dt, st) and the return in each period rt(dt, st). the objective function is dynamic according to the user requirement. (i.e. minimize the component of ppsm result; maximize the accuracy of ppsm result,…). each ppsm (state) has the path (arcs) to represent the return value of the path and next ppsm (state). we add for each ppsm (state) the knowledge base concerning with each return value based on the objective function. constraints are the most important part of every scheduling problem. there are two type of constraint. first, a set of constraints on the value of resources. second, constraint model, a set of constraint on the propose irrigation schedule according to objective function to adaptive the irrigation schedule to satisfy on of the following objective maximize net return, minimize irrigation costs, maximize yield and optimally distribute a limited water supply. . . generate domain model generate domain model uses irrigation knowledge base, meta domain model representation, irrigation library, and set of inference as input and produces the domain model. figure depict the algorithm that used to generate the domain model. . conclution integrating operations research and expert systems under a unify framework in agriculture domain is very important task specially in scheduling system like irrigation system to generate a configuration design includes these two approaches. this paper presents configuration design knowledge to solve irrigation scheduling problem in agriculture domain and a general approach for configuration management knowledge. we present a configuration agriculture irrigation schedule (cais) tool that produces the configuration irrigation schedule design according to the commonkads methodology. this configuration satisfes the user requirements, and some optimal required criteria. by integrating dynamic programming and inference knowledge, including different objective, to select the suitable components and minimize the number of components and maximize the accuracy of the result. . references breuker, j., and van de velde, w. ( ). commonkads library for expertise modeling: reusable problem solving components. ios-press, amsterdam, august . chandrasekaran, b. ( ). design problem solving: a task analysis. ai magazine, winter issue, pages – . gruber t.r., ( ). a translation approach to portable ontology specifications, knowledge acquisition. – . gruber t.r. ( ). ontolingua: a mechanism to support portable ontologies. technical report ksl - (revision), knowledge systems laboratory, stanford university, march . mittal, s. & frayman, f. ( ). towards a generic model of configuration tasks. in proceedings of the th ijcai, pages – , san mateo, ca, morgan kaufman. smith, s.f., o. lassila and m. becker ( ). a scheduling ontology: informal concept descriptions , icll working paper, january, . stephen f. smith and marcel becker ( ). an ontology for constructing scheduling systems, aaai spring symposium on ontological engineering, stanford, ca, march, (aaai press). yu, b. and maccallum, k. ( ). modelling of product configuration design and management by using product structure knowledge. int. workshop on knowledge intensive cad, finland, . wielinga, bob j. ( ) expertise model definition document, esprit project p kads-ii, document id.: kads-ii/m /uva/ / . , university of amsterdam, . eswa.eps expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a improving the interest operator for face recognition yong xu a,*, lu yao a, david zhang b, jing-yu yang c a bio-computing research center, harbin institute of technology, shenzhen graduate school, shenzhen, guang dong , china b biometrics research center, the hong kong polytechnic university, china c school of computer science and engineering, nanjing university of science and technology, china a r t i c l e i n f o keywords: face recognition feature extraction interest operator - /$ - see front matter � elsevier ltd. a doi: . /j.eswa. . . * corresponding author. tel.: + ; fa e-mail address: laterfall @yahoo.com.cn (y. xu). a b s t r a c t when the conventional interest operator is used as the feature extraction procedure of face recognition, it has the following two shortcomings: first, though the purpose of the conventional interest operator is to use the intensity variation between neighboring pixels to represent the image, it cannot obtain all vari- ation information between neighboring pixels. second, under varying lighting conditions two images of the same face usually have different feature extraction results even though the face itself does not have obvious change. in this paper, we propose two new interest operators for face recognition, which are used to calculate the pixel intensity variation information of overlapping blocks produced from the original face image. the following two factors allow the new operators to perform better than the conventional interest operator: the first factor is that by taking the relative rather than absolute variation of the pixel intensity as the feature of an image block, the new operators can obtain robust block features. the second factor is that the scheme to partition an image into overlapping rather than non-overlapping blocks allows the proposed operators to produce more representation information for the face image. experi- mental results show that the proposed operators offer significant accuracy improvement over the con- ventional interest operator. � elsevier ltd. all rights reserved. . introduction since the interest operator was first proposed, it has been wildly used to represent image or target feature. for example, moravec ( ) used an interest operator to compute the intensity vari- ances in the horizontal, vertical and both diagonal directions for each pixel point in an image and selected the minimum of these values as the variance of that point. nasrabadi and choo ( ) used the interest operator in the stereo vision correspondence. in nasrabadi and choo ( ), the interest operator was applied to all image blocks and the point having the local maximal variance in each local block was selected as the so-called interesting point of that block. căleanu, huang, gui, tiponu, and maranescu ( ), căleanu ( ), căleanu ( ), zhao, huang, and sun ( ) and zhao ( ) used the interest operator to extract fea- tures from face images. the interest operator was also exploited for target recognition (haber & modersitzki, ). generally speaking, the interest operator for recognition can be viewed as a feature extraction algorithm that calculates variation information in different directions of the image pixel intensity. note that when the interest operator is applied to face recognition, the first step is to divide each image into the same number of blocks with a fixed ll rights reserved. x: + . size. then feature extraction is performed for each block by using the interest operator. finally, one can treat the feature extraction results of all the blocks of a face image as the representation of this image and can classify face images using the representation and a classifier. it should be pointed out that complex imaging conditions such as varying pose, facial expression and lighting conditions may re- strict the ability of the interest operator. actually, complex imaging conditions degrade performances of most of face recognition tech- niques (adini, moses, & ullman, ). in addition, the human face is a deformable object and a face can produce different deforma- tions at different times, resulting in various facial expressions. con- sequently, there may be much difference between a block of one image of a face and the same block of another image of the same face, if the two images are associated with different poses or facial expressions. face recognition using the conventional interest oper- ator, therefore, has the following shortcoming: the feature extrac- tion results of different images of the same face might have much difference. another shortcoming of the conventional interest oper- ator stems from the algorithm itself. under varying lighting condi- tions the algorithm usually also produces different feature extraction results for the same face, even if the face itself does not have any change in pose and expression. this is because vary- ing lighting conditions produce different intensity values for the image pixel. for example, high intensity lighting usually produces mailto:laterfall @yahoo.com.cn http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa y. xu et al. / expert systems with applications ( ) – high intensity value for the pixel and consequently the variation between two neighboring pixels usually also appears to be high. on the other hand, low intensity lighting usually produces low intensity value for the pixel and consequently the variation be- tween two neighboring pixels is also low. as a result, when we compare the feature extraction results of two images that are ob- tained under the high and low intensity lighting conditions, a low similarity may be produced. the third shortcoming of the con- ventional interest operator is that it cannot obtain all variations be- tween neighboring pixels. indeed, since the conventional interest operator is applied to only pixels within an image block, the con- ventional interest operator is not able to compute the variations between two neighboring pixels that are located in two blocks, respectively. this is illustrated by fig. . in this paper, in order to overcome the shortcomings of the con- ventional interest operator, we propose to exploit new operators and the overlapping block partition scheme for face recognition. the rationale of the proposed approach is as follows: first, because the improved interest operator takes the relative rather than abso- lute variation of image pixel intensity as the features of the face image, this approach can produce more stable face features for the same face than the conventional interest operator, especially under varying lighting conditions. this is helpful for obtaining a satisfactory classification accuracy. second, since in our approach different blocks overlap one another, our approach allows intensity variations between all pixels to be provided whereas the conven- tional interest operator cannot do so. experimental results illus- trate that our approach can produce a higher accuracy in comparison with the conventional interest operator. experiments also show that a linear feature extraction approach can be com- bined with the proposed interest operators to obtain further per- formance improvement. . the conventional interest operator as mentioned above, the conventional interest operator evalu- ates the variation of image pixels in the horizontal, vertical and both diagonal directions for each block of an image. the conven- tional interest operator can be described as follows: it calculates the mean l and the center variance r of a block using ( ) and ( ), respectively. it calculates r ; r ; r and r , which respec- tively stand for the intensity variations of block pixels in horizon- tal, diagonal �, vertical and diagonal � directions, using ( )– ( ), respectively. fig. . illustration of the non-overlapping block partition scheme used for the conventional interest operator. this figure shows that an image is divided into nine non-overlapping blocks each consisting of a number of pixels. for a block, we denote the boundary pixels that adjoin another block by ‘�’ and we denote the other pixels by ‘‘�”. note that for two ‘�’ pixels respectively located in two blocks, even if they are neighbors, the pixel intensity variation of the two pixels cannot be obtained by the conventional interest operator. u ¼ p � q xp x¼ xq y¼ pðx; yÞ ð Þ r ¼ p � q xp x¼ xq y¼ ½pðx; yÞ� u� ð Þ r ¼ p � q xp� x¼ xq y¼ ½pðx þ ; yÞ� pðx; yÞ� ð Þ r ¼ p � q xp� x¼ xq� y¼ ½pðx þ ; yÞ� pðx; y þ Þ� ð Þ r ¼ p � q xp x¼ xq� y¼ ½pðx; y þ Þ� pðx; yÞ� ð Þ r ¼ p � q xp� x¼ xq� y¼ ½pðx þ ; y þ Þ� pðx; yÞ� ð Þ hereafter we suppose that the size of each block is p � q and p(x, y)( x p, y q) represents the pixel intensity of the point (x, y) in a block. fig. shows the feature extraction results of a face image obtained using the conventional interest operator. from fig. , we can clearly see that if two original images of the same face are obtained under varying conditions, their resultant images may have obvious difference. . description of our approach our approach consists of two components, an overlapping block partition scheme and an improved interest operator. the goal of our approach is to improve image presentation ability of the con- ventional interest operator. the basic idea for improving the inter- est operator is to take the relative variation of the gray intensity as features of the image block. based on this idea, we develop two new interest operators as shown in the following. . . improved interest operator improved interest operator is defined as follows: r ¼ p � q xp x¼ xq y¼ ½pðx; yÞ� u� ðu þ c Þ ð Þ r ¼ p � q xp� x¼ xq y¼ ½pðx þ ; yÞ� pðx; yÞ� ðu þ c Þ ð Þ r ¼ p � q xp� x¼ xq� y¼ ½pðx þ ; yÞ� pðx; y þ Þ� ðu þ c Þ ð Þ r ¼ p � q xp x¼ xq� y¼ ½pðx; y þ Þ� pðx; yÞ� ðu þ c Þ ð Þ r ¼ p � q xp� x¼ xq� y¼ ½pðx þ ; y þ Þ� pðx; yÞ� ðu þ c Þ ð Þ where u is also defined by ( ) and c is a positive constant. indeed, the new operator as defined in ( )–( ) produces the relative rather than absolute variation of the pixel intensity by dividing the result of the conventional interest operator by a quantity associated with the mean of the pixel intensity. the difference between improved interest operator and the conventional interest operator is as fol- lows. as presented above, varying lighting condition is one typical factor that makes two corresponding pixels from two images of the same face have quite different intensities. as a result, under varying lighting condition, two corresponding image blocks from two images of the same face usually appear to have different inten- sity variations. the conventional interest operator, therefore, usu- fig. . the original face and resultant images obtained using the conventional interest operator. each row shows the face and resultant images. for example, (b), (c), (d), (e) and (f) respectively stand for the resultant images on r ; r ; r ; r ; r of the face image shown in (a). the original face image is first divided into a number of non- overlapping by blocks and then the conventional interest operator is implemented for each image block. y. xu et al. / expert systems with applications ( ) – ally does not perform well in producing stable features for the same face under varying lighting condition. however, by using the rela- tive variation of the pixel intensity as shown in ( )–( ) as the fea- ture of the face image, improved interest operator is able to produce more stable features for the same face under varying light- ing condition. the use of constant c can prevent the operator from being unfeasible in the case where the mean of the pixel intensity is zero. . . improved interest operator improved interest operator is designed as follows: r ¼ p � q xp x¼ xq y¼ jpðx; yÞ� uj u þ c ð Þ r ¼ p � q xp� x¼ xq y¼ jpðx þ ; yÞ� pðx; yÞj u þ c ð Þ r ¼ p � q xp� x¼ xq� y¼ jpðx þ ; yÞ� pðx; y þ Þj u þ c ð Þ r ¼ p � q xp x¼ xq� y¼ jpðx; y þ Þ� pðx; yÞj u þ c ð Þ r ¼ p � q xp� x¼ xq� y¼ jpðx þ ; y þ Þ� pðx; yÞj u þ c ð Þ where u is still defined by ( ) and c is a positive constant. differing from improved interest operator , the second improvement opera- tor takes as the feature of an image block the result of the conven- tional interest operator divided by the sum of the mean of the pixel intensity and c . the use of c enables the operator to be workable in the case where the mean of the pixel intensity is zero. improved interest operator also has the similar advantages to improved interest operator . . . overlapping block partition scheme as mentioned above, though the conventional interest operator was developed to calculate the pixel intensity variation, not all var- iation information between neighboring pixels could be obtained. indeed, the conventional interest operator cannot calculate the variations between two neighboring pixels respectively located in two blocks as shown in fig. . in order to overcome this draw- back of the conventional interest operator, we propose to partition an image into a number of overlapping blocks and to extract fea- tures from each block by using either of the two interest operators respectively presented in sections . and . . note that in our partition scheme two horizontally neighboring blocks overlap one another and so do two vertically neighboring blocks. fig. shows that how an original � image is partitioned into four overlapping image blocks each having the size of � . here each rectangle unit represents a pixel. the overlapping region is represented by the shadow, which shows that half of each block is overlapped by a neighboring block. our approach implements improved interest operator (or improved interest operator ) for each image block produced by the overlapping block partition scheme. advantages of our approach can be described as follows. first, because our approach takes the relative rather than absolute vari- ation of the pixel intensity as the feature of an image block, the fea- ture extraction result appears to be less affected by the lighting fig. . illustration of the scheme to partition an image into overlapping image blocks. the overlapping region is represented by the shadow. each image block is of size of � . half of every block is overlapped by a neighboring block. for example, half of the � image block located in the left top is overlapped by its right neighbor image block. for the same block, its down neighbor � block also overlaps half of its size. y. xu et al. / expert systems with applications ( ) – condition than the conventional interest operator. this helps our approach obtain more robust features than the conventional inter- est operator. second, our approach allows information of the inten- sity variation of all neighboring pixels to be obtained due to the overlapping partition scheme, whereas the conventional interest operator cannot do so. fig. shows an original face image and the feature extraction result of this image obtained using improved interest operator . the overlapping block partition scheme can be shown more clearly by the following example: if the original image is repre- sented by an � matrix and we partition the original image fig. . the original face image and the resultant images obtained using improved interes example, (b), (c), (d), (e) and (f) respectively stand for the resultant images on r ; r ; r divided into overlapping by blocks and then improved interest operator is implem image block. into a number of � overlapping blocks and half of each block is overlapped by a neighboring block, we will obtain overlap- ping image blocks. when we calculate each of the r ; r ; r ; r ; r values for all the image blocks of the original image, the result that we obtain is a -dimensional vector. the vector can be also shown as a � matrix as illustrated in fig. . since each block has its own five values r ; r ; r ; r ; r , the feature extraction result of the original image can be regarded as five -dimensional vectors or five � matrices. concatenating these one-dimensional vectors, we can obtain a -dimensional vector. more generally, if the original image is an m � n matrix and we partition the original image into a number of s � t overlapping blocks and half of each block is overlapped by a neighboring block, ( m/s � ) � ( n/t � ) image blocks will be generated. as a result, for this original image, the feature extraction result produced by our approach will be a ( m/s � ) � ( n/t � )-dimensional vector. similar to the features obtained using the conventional interest operator, the features obtained using our approach are also usually very high-dimensional. however, as shown in section , linear fea- ture extraction procedures such as dpca (two-dimensional pca) or dfld (two-dimensional fisher discriminant analysis) (cho, chang, kim, & lee, ; kongsontana & rangsanser, ; mutelo, khor, woo, & dlay, ; nhat & lee, ; sanguansat, asdornw- ised, jitapunkul, & marukatat, ; wang, wang, zhang, & feng, ; xu, zhang, yang, & yang, ; yang, zhang, frangi, & yang, ; yang et al., ) can be used to transform the feature extraction results of the interest operator into lower-dimensional features. t operator with c = . each row shows a face image and the resultant images. for ; r ; r of the face image shown in (a). note that the original face image is first ented for each image block. half of an image block is overlapped by a neighboring . . . . . . . . . . dpca io+ dpca( × ) io+ dpca( × ) (a) . . . . . . dfld io+ dfld( × ) io+ dfld( × ) (b) . . . . . . . . pca io+pca( × ) io+pca( × ) (c) fig. . face recognition results on the ar database obtained by different linear feature extraction procedures combined with or without the conventional interest operator. (a) classification right rate associated with dpca, (b) classification right rate associated with dfld and (c) classification right rate associated with pca. the horizontal coordinate represents the number of the transforming axes of the linear feature extraction procedure used for feature extraction and the vertical coordinate represents the classification accuracy. the horizontal and vertical coordinates of figs. , , – are same as this figure. table experimental result comparison of our approach and other approaches on the ar face database. best accuracy(%) mean of the accuracy(%) size of the block dpca . . dfld . . pca . . io + dpca . . � io + dfld . . � io + pca . . � oiio + dpca . . � oiio + dfld . . � oiio + pca . . � oiio + dpca . . � oiio + dfld . . � oiio + pca . . � y. xu et al. / expert systems with applications ( ) – . more discussion on improved and the conventional operators in this section we will compare improved interest operators with conventional interest operator and some other algorithms that exploit gradient information of the image. first, we analyze the relationship and difference between the conventional interest operator and the gradient operator. as presented above, the inter- est operator computes the pixel intensity variation in different directions. gradient operators also evaluate the pixel intensity var- iation. the difference between the interest operator and the gradi- ent operator are as follows: a gradient operator produces a vector i.e. gradient vector that uses two components to indicate the gra- dient whereas the interest operator obtains only scalar values. the square-root of the squared sum of the two components of the gradient vector is usually used to denote the magnitude value of the gradient. the similarity between the interest operator and the gradient operator is as follows: for the conventional interest operator, r and r as defined in ( ) and ( ) act as the mean of the squares of the two components of the roberts gradient oper- ator (gonzalez & woods, ), respectively. additionally, accord- ing to the definition of the gradient, r and r as defined in ( ) and ( ) can also be respectively regarded as the mean of the squares of two components of a gradient operator. consequently, we can conclude that the interest operator and the gradient oper- ator provide image gradient information in different ways. the interest operator aims at calculating ‘average’ gradient information in different directions of an image block, whereas the gradient operator provides gradient information of each image pixel in the form of a vector. gradient operators and improved gradient opera- tors have been applied to recognition problems (gao & leung, ; haber & modersitzki, ; takaÂcs, ; wei & lai, ) and image matching problems (ando, ; wolfson & rigoutsos, ). approaches exploiting image gradient information for recogni- tion can be classified into two classes. the first class of approaches exploits the image gradient information itself to perform recogni- tion. the approaches proposed in haber and modersitzki ( ) and wei and lai ( ) as well as the conventional interest oper- ator are examples of the first class. in haber and modersitzki ( ) and wei and lai ( ), ‘normalized gradient’ and ‘relative gradient’ were respectively used as image features. in wei and lai ( ), the magnitude of the conventional gradient of a pixel point was first divided by the sum of a constant and the maximum mag- nitude of the gradient in an image block. then the division result was taken as ‘relative gradient’ of the pixel point. the second class of approaches firstly exploits the image gradient information to produce image edges and then employs image edges to perform recognition. the rationale of the second class of approaches is that edge information is a useful object representation feature (gao & leung, ). the approaches used in gao and leung ( ), ta- . . . . . . dpca iio+ dpca( × ) iio+ dpca( × ) oiio( / )+ dpca( × ) oiio( / )+ dpca( × ) (a) . . . . . . . . . . . dfld iio+ dfld( × ) iio+ dfld( × ) oiio( / )+ dfld( × ) oiio( / )+ dfld( × ) (b) . . . . . . . . . . pca iio+pca( × ) iio+pca( × ) oiio+pca( × ) oiio+pca( × ) (c) fig. . face recognition results on the ar database obtained by different linear feature extraction procedures combined with or without improved interest operator . (a) classification right rate associated with dpca, (b) classification right rate associated with dfld and (c) classification right rate associated with pca. . . . . . . dpca iio+ dpca( × ) iio+ dpca( × ) oiio( / )+ dpca( × ) oiio( / )+ dpca( × ) (a) . . . . . . dfld iio+ dfld( × ) iio+ dfld( × ) oiio( / )+ dfld( × ) oiio( / )+ dfld( × ) (b) . . . . . . . . pca iio+pca( × ) iio+pca( × ) oiio( / )+pca( × ) oiio( / )+pca( × ) (c) fig. . face recognition results on the ar database obtained by different linear feature extraction procedures combined with or without improved interest operator . (a): classification right rate associated with dpca. (b): classification right rate associated with dfld. (c): classification right rate associated with pca. y. xu et al. / expert systems with applications ( ) – fig. . some face images of a same face. these images are obtained under different illumination conditions. (a) face images obtained in the cases where the azimuth angle and the elevation angle of the light source with respective to the camera axis are small. (b) face images obtained in the cases where the azimuth angle and the elevation angle of the light source with respective to the camera axis are quite large. . . . . . . . . dpca io+ dpca( × ) io+ dpca( × ) (a) . . . . . . . . . dfld io+ dfld( × ) io+ dfld( × ) (b) . . . . . . . . pca io+pca( × ) io+pca( × ) (c) fig. . face recognition results on the yaleb database obtained by different linear feature extraction procedures combined with or without the conventional interest operator. (a): classification right rate associated with dpca. (b): classification right rate associated with dfld. (c): classification right rate associated with pca. y. xu et al. / expert systems with applications ( ) – kaÂcs ( ) and wang et al. ( ) are three typical examples of the second class. in takaÂcs ( ), the binary coding result of so- bel edge detection operator was used as face edge feature. in gao and leung ( ), the line edge map approach detected the edge feature and then classified faces using the edge feature. it was as- sumed that image edge detection using gradient operators was al- most not influenced by varying lighting condition (gao & leung, ); however, the approaches in gao and leung ( ), takaÂcs ( ) and wang et al. ( ) are all not able to obtain truly invari- ant facial features with respect to lighting condition. this is be- cause the pixel intensity value varies with the lighting condition and consequently the gradient value seems to be also variable with respect to the lighting condition. the interest operator appears to be also helpful for indicating salient image edge information in different directions as shown in fig. . actually, an interest operator obtains ‘average’ directional edge of an image block since it sums gradient information in a cer- tain direction. improved interest operators have the following advantage: by improving the conventional interest operator, im- proved operators enable the obtained ‘average’ directional edge information of a face image block to be less variable with respect to the lighting condition. this is very beneficial to face recognition. in addition, since the average relative variation of the pixel inten- sity appears to in some extent be robust to facial expression and pose, improved interest operators may also produce more stable face feature than the conventional interest operator under the con- ditions of varying facial expression or pose. . experimental result in this section we compare our approach and the conventional interest operator using the ar and yaleb face databases. we eval- uate the similarity between the feature extraction results of two face images by using sðx; yÞ¼ xt y jjxjj � jjyjj ; ð Þ where x, y stand for two vectors respectively corresponding to the feature extraction results of the two face images. after the similar- ities between a testing sample and each of all training samples are computed, we select out the training sample that has the maximum similarity to the testing sample and classify the testing sample into the class that the selected training sample belongs to. . . experiments on the ar face database the ar face database includes more than face images showing faces with different facial expressions, in varying lighting . . . . . dpca iio+ dpca( × ) iio+ dpca( × ) oiio( / )+ dpca( × ) oiio( / )+ dpca( × ) (a) . . . . . . . . . dfld iio+ dfld( × ) iio+ dfld( × ) oiio( / )+ dfld( × ) oiio( / )+ dfld( × ) (b) . . . . . . . . . pca iio+pca( × ) iio+pca( × ) oiio( / )+pca( × ) oiio( / )+pca( × ) (c) y. xu et al. / expert systems with applications ( ) – conditions and occluded in several ways (yang et al., ) . each subject have face images. for each subject, the filenames of the images contain the numbers from to , respectively. if the number contained in the filename is , we call the image the first image of the subject, and so on. we use the computer generates random integers in the range of from to . for every subject, we take the images whose filenames contain the numbers from the random integer sequence as the training samples and consider the remaining samples as test samples. the generate random integer sequence is , , , , , , , , , , , , . fig. shows face recognition results on the ar database ob- tained by different linear feature extraction procedures combined with or without the conventional interest operator. note that here- after by io (interest operator) we denote the conventional interest operator. by iio we denote improved interest operator based on the non-overlapping partition scheme. by iio we denote im- proved interest operator based on the non-overlapping partition scheme. in addition, we use oiio (overlapped improved interest operator ) to represent improved interest operator based on the overlapping partition scheme. we also use oiio to represent improved interest operator based on the overlapping partition scheme. if the interest operator is followed by a linear feature extraction procedure, we add some indicators to the denotation. for example, ‘io + dpca’ means that feature extraction is imple- mented by the conventional interest operator followed by dpca. we use a bracket to show the size of each block or the overlapping region. for example, ‘io + dpca( � )’ means that the image is partitioned into a number of � non-overlapping image blocks; whereas ‘oiio( / ) + dpca( � )’ means that the image is parti- tioned into a number of � overlapping image blocks and / re- gions of two neighboring blocks are overlapped each other. according to fig. , the combination of pca and the conventional interest operator is not able to obtain a higher accuracy than pca. the combination of dpca and the conventional interest operator may produce a higher or lower accuracy than dpca. so does the combination of dfld and the conventional interest operator. table shows experimental result comparison of our approach and other approaches on the ar face database. from table and figs. – , we can see that the combination of either of the two im- proved interest operators and dpca or dfld can produce a high- er accuracy than dpca or dfld. the combination of an improved interest operator and dpca or dfld also performs better than the combination of the conventional interest operator and dpca or dfld. fox example, the combination of the conventional inter- est operator and dpca obtains the mean of recognition accuracy of . % and the best accuracy of . %. when oiio is combined with dpca, the mean of the accuracy and the best accuracy are . % and . %, respectively. for the combination of oiio and dpca, the mean of the accuracy and the highest accuracy are . % and . %, respectively. this means that compared to the combination of the conventional interest operator and dpca, the combination of oiio and dpca improves the mean accuracy and the highest accuracy . % and . %, respectively. . . experiments on the yaleb face database the images in the yaleb database are obtained with varying illuminations and unfixed poses and there exists a wide rang of illumination cases. to focus on face recognition with varying illu- minations, we select and use face images with pose of every person. we crop each of these images to obtain a � image fig. . face recognition results on the yaleb database obtained by different linear feature extraction procedures combined with or without improved interest operator . (a): classification right rate associated with dpca. (b): classification right rate associated with dfld. (c): classification right rate associated with pca. h t t p :/ / c o b w e b . e cn . p u rd u e . e du / ~ a l e i x / a l e i x _ f a c e _ d b . h t m l ; h t t p : // c o b - web.ecn.purdue.edu/~aleix/ar.html. http://cvc.yale.edu/projects/yalefacesb/yalefacesb.html. http://cobweb.ecn.purdue.edu/~aleix/aleix_face_db.html http://cobweb.ecn.purdue.edu/~aleix/ar.html http://cobweb.ecn.purdue.edu/~aleix/ar.html http://cvc.yale.edu/projects/yalefacesb/yalefacesb.html . . . . . . . dpca iio+ dpca( × ) iio+ dpca( × ) oiio( / )+ dpca( × ) oiio( / )+ dpca( × ) (a) . . . . . . . . . dfld iio+ dfld( × ) iio+ dfld( × ) oiio( / )+ dfld( × ) oiio( / )+ dfld( × ) (b) . . . . . . . . . pca iio+pca( × ) iio+pca( × ) oiio( / )+pca( × ) oiio( / )+pca( × ) (c) fig. . face recognition results on the yaleb database obtained by different linear feature extraction procedures combined with or without improved interest operator . (a): classification right rate associated with dpca. (b): classification right rate associated with dfld. (c): classification right rate associated with pca. table experimental result comparison of our approach and other approaches on the yaleb face database. best accuracy(%) mean of the accuracy(%) size of the block dpca . . dfld . . pca . . io + dpca . . � io + dfld . . � io + pca . . � oiio + dpca . . � oiio + dfld . . � oiio + pca . . � oiio + dpca . . � oiio + dfld . . � oiio + pca . . � y. xu et al. / expert systems with applications ( ) – (xu, yang, jin, & zheng, ). we test our approach using the ob- tained images. some face images of a same face are shown in fig. . fig. shows face recognition results on the yaleb database ob- tained by different linear feature extraction procedures combined with or without the conventional interest operator. fig. shows face recognition results on the yaleb database obtained by differ- ent linear feature extraction procedures combined with or without improved interest operator . fig. shows face recognition results on the yaleb database obtained by different linear feature extrac- tion procedures combined with or without improved interest oper- ator . table shows experimental result comparison of our approach and other approaches on the yaleb face database. table and figs. – show that the combination of improved interest operator or and a linear feature extraction procedure performs better than the combination of the conventional interest operator and the same linear feature extraction procedure. fox example, the combination of the conventional interest operator and dfld obtains the mean of recognition accuracy of . % and the highest recognition accuracy of . %. when oiio is combined with dfld, the mean of the accuracy and the highest accuracy are . % and . %, respectively. for the combination of oiio and dfld, the mean of the accuracy and the highest accuracy are . % and . %,, respectively. this means that compared to the combination of the conventional interest operator and dfld, the combination of oiio and dfld improves the mean accuracy and the highest accuracy . % and . %, respectively. . . more experimental analysis in this subsection, we further investigate our approach by ana- lyzing the similarity of different images of the same face for the ar face database. analysis is performed for the first, second, seventh, eighth and eleventh face images of each subject. the used five images of one subject are shown in fig. . when compared with image (a), images (b) has different expression, image (c) is ob- tained under different lighting condition, images (d) and (e) are two occluded face images. note that after the conventional interest operator or the pro- posed interest algorithm is applied to all the five images of a face, we can also evaluate the similarities between the resultant images fig. . illustration of the five images of a face used for similarity analysis. table means of the similarities between the resultant images of (a) and those of (b), (c), (d), (e). each row shows the means of the similarities between the resultant images associated with an interest operator of image (a) and images (b), (c), (d), (e). iio and iio respectively denote improved interest operators and applied to the non- overlapping partition scheme. similarity of (b) and (a) similarity of (c) and (a) similarity of (d) and (a) similarity of (e) and (a) io . . . . iio . . . . oiio . . . . iio . . . . oiio . . . . y. xu et al. / expert systems with applications ( ) – of image (a) and each of images (b), (c), (d), (e) by using s ðx ; y Þ¼ xt y jjx jj�jjy jj , where x , y represent the two one-dimen- sional vectors of the resultant images of the two original face images, respectively. note that an interest operator produces five resultant images for each original image, so the corresponding one-dimensional vector used for similarity computation is ob- tained by treating the five resultant images of the original image as an integral image and by concatenating the rows or columns of the integral image. for each subject, we respectively compute the similarities be- tween the resultant images of image (a) and those of images (b), (c), (d), (e) produced by an interest operator. then we calculate the means of the similarities based on all the subjects. the means obtained in the case where each image block is of the same size of � are shown in table . in the overlapping block partition scheme, half of each block is overlapped by a neighboring block. the second to fifth columns of table respectively show the means of the similarities between the resultant images of image (a) and those of images (b), (c), (d), (e). from table , we can see that, when improved interest opera- tors and are applied to non-overlapping image blocks, the ob- tained similarities between the resultant images are higher than the similarities of the resultant images obtained using the conven- tional interest operator. moreover, oiio and oiio produce higher similarities respectively than iio and iio . this implies that the combination of the improved interest operator and the overlapping block partition scheme is more helpful for obtaining high similar- ities than the combination of the improved interest operator and the non-overlapping partition scheme. therefore, we can conclude that both the improved interest operator and the overlapping par- tition scheme are useful for resulting in higher similarities for the resultant images of image (a) and those of images (b), (c), (d), (e). this also means that our approach can effectively reduce the differ- ence between different images of the same face, which is very ben- eficial to face recognition. . conclusion the new approach proposed in this paper differs from the fea- ture extraction approach using the conventional interest operator in the following two aspects. the first aspect is that the proposed approach takes the relative rather than absolute variation of the pixel intensity as the feature of an image block. this allows the proposed approach to obtain robust block features that are less af- fected by varying imaging conditions such as varying lighting and facial expression. the second aspect is that the proposed approach adopts the overlapping block partition scheme. this enables the proposed approach to fully reflect the pixel intensity variation information and to produce more representative feature extraction results. as a result, the proposed approach can do better in face recognition than the conventional interest operator. the analysis and experimental results illustrate the feasibility and the satisfactory performance of the proposed approach. exper- imental results show that the use of the proposed approach can ob- tain more than percent accuracy improvement. acknowledgement we wish to thank program project (no. aa z ), national natural science foundation of china (nos. , ) and natural science foundation of guangdong prov- ince, china (no. ) for supporting. references adini, y., moses, y., & ullman, s. ( ). face recognition: the problem of compensating for changes in illumination direction. ieee transactions on pattern analysis and machine intelligence, ( ), – . ando, s. ( ). image field categorization and edge/corner detection from gradient covariance. ieee transactions on pattern analysis and machine intelligence, ( ), – . cho, d. u., chang, u. d., kim, b. h., & lee, s. h. ( ). d direct lda algorithm for face recognition. in proceedings of the fourth international conference on software engineering research, management and applications, wa, united states (pp. – ). căleanu, c. d. ( ). facial recognition using committee of neural networks. in proceedings of the fifth seminar on neural network applications in electrical engineering, neurel, belgrade, yugoslavia (pp. – ). căleanu, c. d. ( ). face recognition using parallel neural processing and interest operator method. ph.d. thesis, university politehnica timisoara. căleanu, c. d., huang, d.-s., gui, v., tiponu, v., & maranescu, v. ( ). interest operator versus gabor filtering for facial imagery classification. pattern recognition letters, , – . gao, y., & leung, m. k. h. ( ). face recognition using line edge map. ieee transactions on pattern analysis and machine intelligence, , – . gonzalez, rafael c., & woods, richard e. ( ). digital image processing ( nd ed.). boston, ma, usa: addison-wesley longman publishing co., inc.. haber, e., & modersitzki, j. ( ). intensity gradient based registration and fusion of multi-modal images. technical report, department of mathematics & computer science, emory university, atlanta. kongsontana, s. & rangsanser, y. ( ). face recognition using dfld algorithm. in proceedings of the eighth international symposium on signal processing and its applications (vol. , pp. – ). moravec, h. ( ). robot rover visual navigation. ann arbor, mi: university of michigan research press. mutelo, r. m., khor, l. c., woo, w. l., & dlay, s. s. ( ). a novel fisher discriminant for biometrics recognition: dpca plus dfld. iscas , ieee, – . nasrabadi, n. m., & choo, c. y. ( ). hopfield network for stereo vision correspondence. in m. gupta & g. knopf (eds.). neuro vision systems principles and applications (vol. , pp. – ). ieee. nhat, v. d. m., & lee, s. ( ). improvement on pca and dpca algorithms for face recognition. in proceedings of the fourth international conference on i singapore image and video retrieval (civr), singapore (pp. – ). sanguansat, p., asdornwised, w., jitapunkul, s., & marukatat, s. ( ). two- dimensional linear discriminant analysis of principle component vectors for face recognition. ieice transactions on information and systems, e -d( ), – . takaÑcs, b. ( ). comparing face images using the modified hausdorff distance. pattern recognition, , – . wang, l. c., der, s. z., & nasrabadi, n. m. ( ). automatic target recognition using a feature decomposition and data decomposition modular neural network. ieee transactions on image processing, , – . wang, l., wang, x., zhang, x., & feng, j. ( ). the equivalence of two-dimensional pca to line-based pca. pattern recognition letters, , – . wei, shou-der, & lai, shang-hong ( ). robust and efficient image alignment based on relative gradient matching. ieee transactions on image processing, , – . wolfson, h., & rigoutsos, i. ( ). geometric hashing: an overview. ieee computational science & engineering magazine, ( ), – . xu, y., yang, j.-y., jin, z., zheng, y.-j. ( ). local correlation classification and its application to face recognition across illumination. in the international conference of machine learning and cybernetics, dalian (pp. – ). xu, y., zhang, d., yang, j., & yang, j.-y. ( ). an approach for directly extracting features from matrix data and its application in face recognition. neurocomputing, , – . yang, j., zhang, d., frangi, a. f., & yang, j. y. ( ). two dimensional pca: a new approach to appearance-based face representation and recognition. ieee transactions on pattern analysis and machine intelligence, , – . zhao, t. ( ). several approaches for feature extraction and selection for face recognition. m.s. degree thesis, university of harbin institute of technology (in chinese). zhao, z. q., huang, d. s., & sun, b. y. ( ). human face recognition based on multi- features using neural networks committee. pattern recognition letters, , – . improving the interest operator for face recognition introduction the conventional interest operator description of our approach improved interest operator improved interest operator overlapping block partition scheme more discussion on improved and the conventional operators experimental result experiments on the ar face database experiments on the yaleb face database more experimental analysis conclusion acknowledgement references an expert system approach to quality control an expert system approach to quality control e.p. paladini* departmento de engenharia de produção e sistemas, universidade federal de santa catarina, cp , trindade, - , florianopolis, sc, brazil abstract this paper describes the basic actions proposed for quality inspection. these actions involve structuring a decision supporting expert system (dses) to help with those decisions related to the preliminary activities for inspection development —most of them relating to determining of the need or convenience of carrying out the inspection itself. once the opportunity to carry it out is defined, the expert system helps the user to select the type of inspection to adopt from amongst: ( ) automatic or sensorial inspection; ( ) inspection by samples or complete; ( ) acceptance or rectifying and, in the most relevant module, ( ) inspection by attributes or by variables. the complementary documentation of the dses contains the directions to operate it, the rules and qualifiers that make up the system, as well as the results achieved through its experimental implementation. q elsevier science ltd. all rights reserved. keywords: expert system; quality control; total quality management . introduction having the current concept of quality in mind (and the context of total quality management itself), the feasibility process of basic quality control actions is nowadays consid- ered to take place within a well-defined structure, known as quality evaluation system. inspection, in turn, is the most important activity in the quality evaluation system of an industrial process. when correctly developed, the inspection makes possible to carry out a precise analysis of how the process operates and serves as a basis for a set of decisions that directly affect it, such as corrective and preventive actions which must be complied with in order to guarantee acceptable quality levels. quality inspection has got a number of effective techni- ques with a wide range of applicability. since there are several techniques available, there are many options for those who intend to devise an inspection process. never- theless, this situation poses difficulties as to the correct use of the different techniques, since most of such techniques have their own utilization particularities and yield results valid only within certain contexts. the need of choosing the most adequate technique for each situation shows that it is necessary to organize the information related to quality inspection, or else the whole quality evaluation process could be seriously compromised. the present paper highlights this question, which is considered to be a relevant restriction to the perfect use of quality evaluation. a more efficient way of optimizing qual- ity evaluation development is proposed. we dedicate special attention to quality inspection by attributes, a much more difficult area when it comes to making the correct decision about quality evaluation of services, products and processes. the literature about quality inspection is plentiful and easily accessible. more often than not, classical texts on quality control deal with this subject by focusing on proce- dures of sampling for acceptance, whereas rectifying inspection is rather rarely analyzed. the development of the theoretical models of the operating characteristic (oc) curve associated with basic concepts, such as accep- table quality level (aql), lot tolerance percent defec- tive (ltpd) and producer’s and consumer’s risk, is always presented as a means of motivating the analysis of sampling schemes or sampling plans and inducing the procedures which aim at structuring such plans and making feasible their performance analysis, mostly in terms of reliability and costs. tables suggesting models of sampling plans are just as relevant and are present as well in almost all books of the area. the following authors can be cited as classical in this field: besterfield ( ), charboneau and webster ( ), juran ( ), and montgomery ( ), to name a few. texts of renowned authors, such as dodge and deming can be found in more specialized journals. a collection of articles by dodge about inspection by attributes was made available by the american publication called journal of quality technology in (see dodge, ). as for the work of deming, articles on this subject written by him can expert systems with applications ( ) – pergamon expert systems with applications - / /$ - see front matter q elsevier science ltd. all rights reserved. pii: s - ( ) - www.elsevier.com/locate/eswa * tel.: - - - ; fax: - - - . e-mail address: paladini@iaccess.com.br (e.p. paladini). be found in the same journal published in (papadakis, ). the types of inspection are manifold, each of them bear- ing their particularities, specific purposes and restricted adequacy to certain contexts. such is the case of sampling inspection or complete inspection, as well as inspection by attributes or by variables, for instance. the choice of the correct type of inspection is extremely relevant for quality evaluation, since the adequacy of the inspection process to the context where it takes place has been considered in many situations and its importance has always been stressed. let us consider the inspection of raw materials as an example. an incorrect development of this type of inspec- tion can give the suppliers the impression that quality is not important to their customers and of course the suppliers will not have to regard it as relevant. such a situation, which can be seen in practice, is to be found in several texts. horsnell ( ), for example, draws attention to this fact, by showing that the use of inspection by attributes when products enter the factory is of such usefulness that it transcends inspection proper, in addition to being an easily applicable technique. in a broader analysis, one can see that the quality of the process can be significantly (and positively) affected by the mere inclusion of inspection in the process (it has been considered many years ago—see, for instance, whittle, ). quality inspection is widely used nowadays. nonethe- less, several problems have been encountered upon using inspection systems and such problems derive for the most part from an inadequate application of the prescribed meth- odology. this inadequacy results from decisions which ought to be made as regards the way the inspection is to be carried out. such decisions are considered here to be intuitive. however, they do deserve detailed investigation. we under- stand that the lack of an adequate methodology for planning the decision-making process is a fundamental restraint to the use of quality inspection. this situation justifies the elaboration of a general deci- sion-making support system. due to its peculiarities, the system was organized under the format of an expert system, which benefits from the advantages of artificial intelligence (ai) in order to carry out an evaluation. it should be noted that since inspection is the most impor- tant part of quality evaluation, the expert system proposed here plays a fundamental role in the total quality manage- ment area. it is easy to see that this paper describes a process of interaction between users and computers in industrial companies. the process becomes feasible through the development and application of an expert system related to the area of quality control. this paper describes the devel- oped expert system, its operating methodology and lists the results obtained after its application in the companies we have studied. the interaction between the human resources of the organizations and the computational resources was intense in the last few years. with the advent of ai this interaction has acquired new, rather specific characteristics. . artificial intelligence applied to quality evaluation a brief review of the specific technical literature reveals that there are many cases of successful applications of al techniques to the quality evaluation area. such applications were helpful in the resolution of relevant problems. below are some examples thereof: (a) hosni and elshennavy ( ) reported the develop- ment of a quality control system based on knowledge, suitable to specific procedures of inspection by variables and to the selection of control graphs, both by attributes and by variables. (b) eyada ( ) developed an expert system aimed at auditing procedures of quality assurance involving both suppliers and products in the process. some years earlier, in the same field, gipe and jasinski ( ) conducted an analysis of expert system adequacy to quality assurance problems, showing the feasibility of this application. another interesting expert system was developed by crawford and eyada ( ) with the purpose of planning the allocation of resources for the quality assurance program. (c) there are several expert systems used to select control graphs (see alexander & jagannathan, ; dagli, ; dagli & stacey, ). an expert system designed for process control at a general level was developed by moore ( ). (d) evans and lindsay ( ) reported the development of an expert system for statistic quality control, which not only selects control graphs, but also offers interpreta- tions of such graphs and provides conclusions about the control status of a process. (e) brink and mahalingam ( ) developed an expert system which evaluates quality at manufacturing level, so as to detect and correct defects occurring during the productive process. (f) pfeifer ( ) reported the development of an expert system to detect defects during the productive process and described its successful application in germany. (g) an expert system which makes use of pattern recog- nition in order to ‘visualize’ pieces under inspection was developed in in the united states (see ntuen, park & kim, ). this system, called kims, was success- fully tested in a variety of experiments (one of them is reported by ntuen, park & sohn, ), where the perfor- mance of the kims in image recognition tasks is assessed. (h) fard and sabuncuoglu ( ) developed an expert system which seeks to select sampling by attributes, determining which type of sampling is the most adequate for each case: simple, double or multiple. a project of a e.p. paladini / expert systems with applications ( ) – https://isiarticles.com/article/ personal identification with iris patterns raf. j. of comp. & math’s. , vol. , no. , design a fuzzy expert system for liver and pancreas diseases diagnosis baydaa s bhnam baydaa_sulaiman@uomosul.edu.iq college of computer sciences and mathematics university of mosul, mosul, iraq received on: / / accepted on: / / abstract fuzzy logic is a branch of artificial intelligence techniques, it deals with uncertainty in knowledge that simulates human reasoning in incomplete or fuzzy data. fuzzy relational inference that has applied in medical diagnosis was used within the medical knowledge base system to deals with diagnostic activity, treatment recommendation and patient's administration. in this research, a medical fuzzy expert system named (liv&panfes) has been developed for diagnosis and decision making of general liver and pancreas diseases. the (liv&panfes) is a rule based fuzzy expert system, results of laboratory analysis are inserted into the system. this system can define the probable diagnosis on these data, and later on it can pick out the most probable one for disease. keywords: artificial intelligence, fuzzy logic, fuzzy expert system تصميم نظام خبير مضبب لتشخبص امراض الكبد والكلى بيـداء سليمان بهنام العراق ،جامعة الموصل، كلية علوم الحاسبات والرياضيات / / تاريخ قبول البحث: / / تاريخ استالم البحث: الملخص المنطققا المبققو رققو ققري مققل كانيققات القق فاو اسصققطناعم، مققو لمققبة م المعلومققات المبققوبة و ققر إن العالقققات المبققوبة كمقققبة م ققم البطةيةققات الطويقققة اعبمققا ا علققق إن. اإلنمقققانعمققل أسققلو الكاملققة ققم محاكقققا ات الةاصة بالطةص المعنم.المعلومات الممبةلةة مل الفحوص نظقام مبقو يو قر لق يل قم الملقاي الطوقم ويمقاع قم (liv&panfes) م ر ا البحث كم كةميم الكو والونكرياس. أمراضاكةاذ الارار فيما لبعلا ببطةيص قققم كح لققق نوميقققة المقققرض المةقققا بققق (liv&panfes)لعبمققق كطقققةيص النظقققام الةو قققر المبقققو إل الطائعة البم كعو األمراضائج البحال ل المةبورية، ويمبة م النظام م اكةاذ الارار وكطةيص المريض عل نب لكل مما. أوالونكرياس أوالكو .، نظام يو ر مبو ال فاو اسصطناعم، المنطا المبو الكلمات المفتاحية: . artificial intelligence in medicine artificial intelligence (ai) is a study to emulate human intelligence into computer technology. the potential of ai in medicine has been expressed by a number of researchers [ ][ ][ ] and [ ] summarized the potential of ai techniques in medicine as follows: ▪ proved a laboratory for the examination, organization, representation and cataloging of medical knowledge. ▪ produces new tool to support medical decision-making, training and research. ▪ integrates activities in medical, computer cognitive and other sciences. baydaa s. bhnam ▪ offers a content-rich discipline for future scientific medical specialty. . fuzzy control in medicine fuzzy control techniques have recently been applied in various medical processes. fuzzy control compared to classical control theory, which is a fuzzy logic approach to control, offers the following advantages: ▪ the main advantage of using fuzzy control system is the simplicity of the approach , the capacity of dealing with the complex data and solves a problem-control system methodology [ ][ ]. ▪ it can be used in systems, which don't need to measure the parameters of the model and cannot be easily modeled mathematically. in particular systems with non-linear responses that are difficult to analyze may respond to a fuzzy control approach. hence, many complex systems can be controlled without knowing the exact mathematical model of them [ ][ ]. ▪ using fuzzy control to identify the diseases form the symptoms which helps to develop fuzzy rules that can be stored in the knowledge base and can be fired during further decision process [ ]. ▪ continuous variables may be represented by linguistic constructs that are easier to understand, making the controller easier to implement and modify [ ]. ▪ fuzzy controllers may be less susceptible to system noise and parameter changes, in other words, they will be more robust. ▪ complex process can be controlled by relatively few logical rules permitting an easily comprehensible controller design and faster computation [ ]. ▪ in other word, fuzzy control can be best applied to production tasks, that heavily rely on human experience and intuition, and which therefore rule out the application conventional control methods [ ][ ]. . medical knowledge as a fuzzy relation the knowledge base, which is the central to any fuzzy logic system, represents the domain knowledge. a knowledge base contains two types of knowledge: facts and rules. the facts represent various aspects of specific domain that are known before consultation of the system. a rule is an if-then structure that logically relates information contained in the if part to other information contained in the then part [ ]. the relationship between symptoms and diagnosis by the concept of medical knowledge was introduced. "in a given pathology, we denote s a set of symptoms, d a set of diagnoses and p a set of patients. what we call medical knowledge is a fuzzy relation, generally denoted by r, from s to d expressing of diagnoses". zadeh's max- min compositional rule adopted as an inference mechanism. it accepts fuzzy descriptions of the patient's symptoms and infers fuzzy descriptions of the patient's diseases by means of the fuzzy relationships. if a patient's symptom is si then the patient's state in terms of diagnoses in a fuzzy set dj with the following membership function [ ][ ][ ]:   ddssdssd rs ss d ij =  ,,),();(minmax)(  ),( ds r  is the membership function of the fuzzy relation "medical knowledge". with p, a set of patients, and a fuzzy relation q from p to s, and by "max-min" composition" we get the fuzzy relation t=q×r with the membership function [ ][ ]:   ddssppdsspdp rq ss t =  ,,,),();,(minmax),(  design a fuzzy expert system for liver and pancreas diseases diagnosis . the designed liver & pancreas fuzzy expert system(liv&panfes) the design structure, application and working principles of liver& pancreas fuzzy expert system (liv&panfes) has been described for diagnosis of general liver & pancreas diseases. the (liv&panfes) is parallel rule-firing system, in which all fireable rules are fired effectively at one time. during the process with the medical expert system (liv&panfes), laboratory test results are converted into fuzzy compatibility values reaching from zero to unity by consideration of the linguistic medical concepts. these fuzzified data are used to infer diagnosis with knowledge contained in a knowledgebase. fuzzy relations were calculated for all linguistic medical concepts between test results and diagnosis by using the obtained fuzzy sets with the given set of patient data. . medical knowledgebase the examination and laboratory data of the system assigned for the liver&pancreas patients were collected from al-jamhoree teaching hospital. the data then inserted into the system (liv&panfes), for the design process: the total serum bilirubin (tsb), (tsb-direct), serum glutamic pyruvic transaminase (sgpt), serum glutamic oxaloacetic transaminase (sgot) and serum alkaline phosphataste (sap) examinations are used for diagnosis hemolytic anemia (ha), jaundice (jau), hepatitis (hep), myocardial infarction (myi) and osteomalacia (ost) diseases; which are related with liver diseases. the fasting blood sugar (fbs) and random blood sugar (rbs) examinations are used for diagnosis hypoglycemia (hypogl), hyperglycemia (hypergl) diseases; which are related with pancreas diseases. the unit that used for all examinations is (mg/dl ). . inference methodology the fuzzy control application structure as shown in figure ( ) has seven crisp input variables , these input parameters represent the patients data. the mamdani model of inference was used. all (liv&panfes) fuzzy rules were of the form: " if x is a and y is b then z is c ", where a , b and c are fuzzy sets. the max-min operators were used for implication throughout for implementing (liv&panfes). it was necessary to obtain crisp output for the purposes of evaluation of the fuzzy model , center-of- gravity (centroid) defuzzification was used to produce crisp values on an arbitrary scale of the fuzzy output variable (outfes) . . linguistic variables and fuzzy terms each of the seven input parameters was assigned a linguistic variable and examination of the data and rules showed that each could naturally be divided into two or three fuzzy terms corresponding to meanings of low (l), normal (n), and high (h) for the types laboratory examinations. one output fuzzy variable was used, from the rules it was determined that the output fuzzy variable (outfes) has eighteen linguistic terms: (normal, ha, jau, hep, myi, ost, hypogl, hypergl, ha&hypogl, ha&hypergl, jau&hypogl, jau&hypergl, hep&hypogl, hep&hyperl, myi&hypogl, myi&hypergl, ost&hypogl, ost&hypergl), these terms represent the general liver and pancreas diseases. baydaa s. bhnam . rule set liv&panfes must contain a set of rules that can deal with fuzzy tags, fuzzy sets and relations and must provide an appropriate output which corresponds to a particular input. the rules for the fuzzy expert system were obtained by means of knowledge expert-doctor, and had been carefully refined to form a complete and consistent set of classifiers. part of the developed (liv&panfes) fuzzy knowledge base rules are shown in table ( ), total of rules are formed as shown in the appendix. table( ): liv&panfes fuzzy rules outfes rbs fbs sap sgot sgpt tsb (direct) tsb rule no. normal n n n n n n n … … … … … … … … … myi n l n h n n n … … … … … … … … … ost n h h n n n n … … … … … … … … … hep&hypogl l l n h h h h … … … … … … … … … ost&hypergl h h h n n n n `for example rule can be interpreted as follows (as in table ( )): rule : if patient's tsb is high, patient's tsb-direct is high, patient's sgpt is high, patient's sgot is high and patient's sap is normal, patient's fbs is low, patient's rbs is low, then patient has hepatitis (hep) and hypoglycemia (hypogl) diseases; which are related with liver and pancreas diseases. . fuzzy operations the probabilistic family of operators was chosen for the reason that all the rules features conjunction of the input linguistic variables use the fuzzy conjunction operators. the min operator is used for conjunction , the overall truth value of such a rule will obviously by determined solely by the reasoning of the clinician in considering all the parameters. after the truth degree of each rule is determined, we calculate the truth degree of all rules by taking max between working rules. . membership functions in this paper, triangular and trapezoidal membership functions are used. fuzzy membership functions for the medical factors for the mentioned laboratory parameters are calculated by the following functions as shown in figures( ) and ( ). fbs sgpt tsb(direct) rbs tsb tsb(indirect) sap sgot sgo t figure( ): the structure of the liver&pancreas fuzzy expert system(liv&panfes) design a fuzzy expert system for liver and pancreas diseases diagnosis normal ( , ) ; ost ( ,, ) ; jau&hypogl ( , ) ; myi&hypergl( , ) ha ( , ) ; hypogl ( , ) ; jau&hypergl ( , ) ; ost&hypogl( , ) jau ( , ) ; hypergl ( , ) ; hep&hypogl ( , ) ; ost&hypergl( , ) hep ( , ) ; ha&hypogl ( , ) ; hep&hypergl ( , ) myi ( , ) ; ha&hypergl ( , ) ; myi&hypogl ( , ) figure( ): the membership functions of output term (outfes) for liv&panfes figure( ):the membership functions of input terms for liv&panfes baydaa s. bhnam . experimental results for the output factor outfes, the linguistic expressions represent the patients disease. the truth degree of the rules are determined for each rule by aid of the min and then by taking max between working rules. for example, for the input patient laboratory tests values tsb= . , tsb(direct)= , sgpt= , sgot = . , sap= , fbs = and rbs= . the rules ( ) and ( ) is fired as shown in figure ( ). figure( ): calculation of the value outfes represents the diagnosis of a patient affected by hep depending on the following laboratory test values (tsb= . , tsb(direct)= , sgpt= , sgot = . , sap = , fbs = , rbs= ). from mamdani max-min inference we will obtain the membership function of our system, max (rule , rule ) = rule , then we calculate the crisp output. the crisp value of the output is calculated by the method of center of gravtity defuzzifier. as it is seen from figure ( ), the value of outfes= , this means that the patient has the hepatitis (hep) disease (see figure( )). another examples for implementation of the (liv&panfes) system for patients disease diagnosis shown in figures ( , , , , , , , and ) with patients laboratory tests values, fired rules ( ), ( ), ( ), ( and then rule is max), ( ), ( ), ( and then rule is max), ( ) and ( ) are values of output term (outfes) that refer to the diagnosis of the patient complaint. the performance of the model is tested for patients, the results show that the all diagnosis using (liv&panfes) system is correct. figure( ): calculation of the value outfes represents the diagnosis of a patient affected by hypergl depending on the following laboratory test values (tsb= . , tsb(direct)= . , sgpt= . , sgot = , sap = , fbs= , rbs= ) design a fuzzy expert system for liver and pancreas diseases diagnosis figure( ): calculation of the value outfes represents the diagnosis of a patient affected by ost&hypogl depending on the following laboratory test values (tsb= . , tsb(direct) = . , sgpt = , sgot = , sap = , fbs = , rbs= ). figure( ): calculation of the value outfes represents the diagnosis of a patient affected by normal depending on the following laboratory test values (tsb= . , tsb(direct)= . , sgpt = , sgot = . , sap = , fbs = , rbs= ). figure( ): calculation of the value outfes represents the diagnosis of a patient affected by jau depending on the following laboratory test values (tsb= , tsb(direct)= . , sgpt = , sgot = . , sap = , fbs = , rbs= ). baydaa s. bhnam figure( ): calculation of the value outfes represents the diagnosis of a patient affected by hep&hypergl depending on the following laboratory test values (tsb= . , tsb(direct)= , sgpt= , sgot = , sap = , fbs= , rbs= ). figure( ): calculation of the value outfes represents the diagnosis of a patient affected by ha depending on the following laboratory test values (tsb= . , tsb(direct)= . , sgpt= . , sgot = , sap = , fbs= , rbs= ). figure( ): calculation of the value outfes represents the diagnosis of a patient affected by ost&hypergl depending on the following laboratory test values (tsb= . , tsb(direct)= . , sgpt= , sgot= , sap= , fbs= , rbs= ). design a fuzzy expert system for liver and pancreas diseases diagnosis figure( ): calculation of the value outfes represents the diagnosis of a patient affected by myi&hypergl depending on the following laboratory test values (tsb= . , tsb(direct)= . , sgpt= , sgot= . , sap= , fbs= , rbs= ). figure( ): calculation of the value outfes represents the diagnosis of a patient affected by jau&hypogl depending on the following laboratory test values (tsb= . , tsb(direct)= . , sgpt= , sgot= . , sap= , fbs= , rbs= ). . conclusions the developed diagnosis module (liv&panfes) consists of expert system and fuzzy logic techniques to perform diagnostic tasks. a set of rules will be defined using the patients disease database as well as the expert knowledge on the disease domain (from doctors). the designed expert system uses the rules to diagnose patient's illness base on their laboratory tests. in addition , fuzzy logic is integrated to enhance the reasoning when dealing with fuzzy data. the combination of expert system and fuzzy logic that forms a hybrid (expert-fuzzy) system could increase the system performance and has been implemented successfully. baydaa s. bhnam references [ ] adlassing k.p. ( ), "fuzzy set theory in medical diagnosis", ieee transactions systems man cybernetics , vol. , , pp. – . [ ] adlassing k.p. and g.kolarz, ( ), "cadiac - : computer – assisted medical diagnosis using fuzzy subsets", in approximate reasoning in decision analysis, mm gupta and e sanchez, eds., pp. – , north-holland, new york . [ ] balwinder, s., rajneesh, g., rakesh ,k. and singh, r.p.p, ( ), "design and vlsi implementation of fuzzy logic controller", ijcns) international journal of computer and network security, vol. , no. , december, pp. - . [ ] chinniah, p. and muttan, s., ( ), "icd based medical expert system using fuzzy temporal logic", (ijcsis) international journal of computer science and information security, vol. , no. , pp. - . [ ] cohen m.e. and hudson d.l., ( ), "the use of fuzzy variables in medical decision making", in fuzzy computing m.m. gupta and t. yamakawa, eds., pp. – . elsevier science, north holland. [ ] halim m., k. m. ho, and a.liu, ( ), "fuzzy logic for medical expert systems", annals academy medicine, vol. , no. , pp. - . [ ] hamid, m., dan, i., jerome, b. and bernadette, d., ( )., "human activities of daily living recognition using fuzzy logic for elderly home monitoring", fuzz-ieee, korea, august - , pp. - . [ ] hoong, n.k., ( ), "medical information science - framework and potential", international seminar and exhibition computerization for development -the research challenge, university pertain malaysia: kuala lumpur, pp. – . [ ] hughes, c., ( )., "the representation of uncertainty in medical expert systems", medical informatics, vol. , pp. - . [ ] khalil, s., ( ), "performance tuning of novell netware based on fuzzy reasoning", international journal of computers, issue , vol. , pp. - . [ ] klement, e.p. and slany, w., ( ), "fuzzy logic in artificial intelligence", proceedings of the th austrian artificial intelligence conference flai , berlin. [ ] mamdani, e.h., qstergaard, j.j. and lembessis, e., ( ), "use of fuzzy logic for implementing rule-based control of industrial processes", in: p.p.wang: advances in fuzzy sets, possibility theory, and applications, new york: plenum press, pp. - . [ ] nguyen hoang phuong and vladik kreinovich, ( ), "fuzzy logic and it's applications in medicine", institute of information technology, national center for natural science and technology of vietnam. [ ] oscar, m., roberto, s. and patricia, m., oscar, c., miguel a. p. and iliana, m.m, ( ),“performance of a simple tuned fuzzy controller and a pid controller on a dc motor”, in proceedings of the ieee symposium on foundations of computational intelligence, pp. - . design a fuzzy expert system for liver and pancreas diseases diagnosis [ ] sanchez, e., ( ), "resolution of composite fuzzy relation equations", information and control. , pp. - . [ ] sanchez, e., ( ), "medical diagnosis and composite fuzzy relations", advances in fuzzy set theory and applications, amsterdam: north-holland, pp. - . [ ] serena, h. c., anthony j. j. and john p. n., ( ), "artificial intelligence techniques: an introduction to their use for modelling environmental systems", mathematics and computers in simulation, vol. , pp. - . [ ] shivaraj, g., prakash, b.d., vinayak, v.h., avinash, a k.m., sonal, n.,v. and shruthi, s., k., ( ), "a review on laboratory liver function tests", panafrican medical journal, published november, pp. - . [ ] ulrich h. l.,( ),"humanoid knowledge-based decision support systems (kbdss)- subjective or objective data query", journal of advanced computattional intelligence and intelligent informatics, vol. , no. ,pp. - . [ ] watanabe h., w.j. yakowenko, y. m. kim, j.ande, and t.tobi, ( ), "application of a fuzzy discrimination analysis for diagnosis of valvular heart disease", ieee transactions fuzzy systems, vol. , no. , pp. - . [ ] william siler and james j. buckley, ( ), " fuzzy expert systems and fuzzy reasoning", wiley inerscience. [ ] zadeh a.l.,( ),"the role of fuzzy logic in the management of uncertainty in expert systems ", fuzzy set systems, vol. , pp. - . baydaa s. bhnam appendix: rule-base knowledge for (liv&panfes) outfes rbs fbs sap sgot sgpt tsb (direct) tsb rule no. normal n n n n n n n ha n n n n n n h jau n n n n n h h hep n n n n h h h hep n n n h h h h hep n n h h h h h hep n n n n n h n hep n n n n h n n myi n n n h n n n ost n n h n n n n normal n l n n n n n normal n h n n n n n hypogl l n n n n n n hypogl l l n n n n n hypogl l h n n n n n hypergl h n n n n n n hypergl h l n n n n n hypergl h h n n n n n ha n l n n n n h jau n l n n n h h hep n l n n h h h hep n l n h h h h hep n l h h h h h hep n l n n n h n hep n l n n h n n myi n l n h n n n ost n l h n n n n ha n h n n n n h jau n h n n n h h hep n h n n h h h hep n h n h h h h hep n h h h h h h hep n h n n n h n hep n h n n h n n myi n h n h n n n ost n h h n n n n ha&hypogl l l n n n n h jau&hypogl l l n n n h h hep&hypogl l l n n h h h hep&hypogl l l n h h h h hep&hypogl l l h h h h h hep&hypogl l l n n n h n hep&hypogl l l n n h n n myi&hypogl l l n h n n n ost&hypogl l l h n n n n ha&hypogl l n n n n n h jau&hypogl l n n n n h h hep&hypogl l n n n h h h hep&hypogl l n n h h h h hep&hypogl l n h h h h h hep&hypogl l n n n n h n hep&hypogl l n n n h n n myi&hypogl l n n h n n n ost&hypogl l n h n n n n ha&hypogl l h n n n n h jau&hypogl l h n n n h h hep&hypogl l h n n h h h design a fuzzy expert system for liver and pancreas diseases diagnosis hep&hypogl l h n h h h h hep&hypogl l h h h h h h hep&hypogl l h n n n h n hep&hypogl l h n n h n n myi&hypogl l h n h n n n ost&hypogl l h h n n n n ha&hypergl h l n n n n h jau&hypergl h l n n n h h hep&hypergl h l n n h h h hep&hypergl h l n h h h h hep&hypergl h l h h h h h hep&hypergl h l n n n h n hep&hypergl h l n n h n n myi&hypergl h l n h n n n ost&hypergl h l h n n n n ha&hypergl h n n n n n h jau&hypergl h n n n n h h hep&hypergl h n n n h h h hep&hypergl h n n h h h h hep&hypergl h n h h h h h hep&hypergl h n n n n h n hep&hypergl h n n n h n n myi&hypergl h n n h n n n ost&hypergl h n h n n n n ha&hypergl h h n n n n h jau&hypergl h h n n n h h hep&hypergl h h n n h h h hep&hypergl h h n h h h h hep&hypergl h h h h h h h hep&hypergl h h n n n h n hep&hypergl h h n n h n n myi&hypergl h h n h n n n ost&hypergl h h h n n n n elsevier required licence: © < >. this manuscript version is made available under the  cc‐by‐nc‐nd  .  license http://creativecommons.org/licenses/by‐nc‐nd/ . /    a disaster management knowledge repository (dmkr) and sharing system siti hajar othman, ghassan beydoun abstract. disaster management (dm) is collaborative and complex. roles involved in dm processes often cut across many organizational boundaries and are dynamic. knowledge involved is enormous and diverse. it includes information related to varieties of disasters, roles, plans and operations. many practices may also vary across different regions and authorities. structuring, storing and reusing dm knowledge is the challenge tackled in this paper. the paper presents a metamodel-based disaster management knowledge repository (dmkr) tool, dmkr . , to facilitate collaboration and dm knowledge sharing. the knowledge representation that underpins the repository structure is a disaster management modeling language that makes statements about what can be expressed in various dm models. using the tailored dm language, the tool developed offers a flexible structure to allow the storage and retrieval not only of observed and measured data, but also interpretative and inferred information of the disaster management knowledge. the paper describes the development, functionality and implementation of dmkr. the validation of dmkr illustrates how users can easily instantiate dm models to communicate and generalize their knowledge for the benefit of sharing it within their community. keywords: knowledge sharing; decision making; disaster management; metamodels a disaster management knowledge repository (dmkr) and sharing system siti hajar othman, ghassan beydoun abstract. disaster management (dm) is collaborative and complex. roles involved in dm processes often cut across many organizational boundaries and are dynamic. knowledge involved is enormous and diverse. it includes information related to varieties of disasters, roles, plans and operations. many practices may also vary across different regions and authorities. structuring, storing and reusing dm knowledge is the challenge tackled in this paper. the paper presents a metamodel-based disaster management knowledge repository (dmkr) tool, dmkr . , to facilitate collaboration and dm knowledge sharing. the knowledge representation that underpins the repository structure is a disaster management modeling language that makes statements about what can be expressed in various dm models. using the tailored dm language, the tool developed offers a flexible structure to allow the storage and retrieval not only of observed and measured data, but also interpretative and inferred information of the disaster management knowledge. the paper describes the development, functionality and implementation of dmkr. the validation of dmkr illustrates how users can easily instantiate dm models to communicate and generalize their knowledge for the benefit of sharing it within their community. . introduction disaster management (dm) processes involve many interacting elements e.g.: people, authority, emergency teams, resources, procedures, uncertain environmental situations and many more. modeling coordination of dm activities is tremendously hard and complex. dm activities often extend across various government sectors, non-governmental organizations/industry, from international levels down to state or region levels and may also include individual people and various facets of society. it is often unclear what are the exact tasks and responsibilities before, during or after a disaster strikes. the complexity of dm and the failures of dm agencies can be observed in many recent examples, e.g. the management of the swine-flu (h n ) pandemic hitting australian shores in large numbers through cruise ships in (larcombe, moloney et al. ), or the devastating communication failures in the victorian bushfires in victoria in (cordner, woodford et al. ). observed failures often surface as the required expertise not being available in a timely manner or a systemic inability to timely recognize and identify the required expertise. reusing expertise is overlooked as it is often perceived as too specific to kinds of events such as floods, bushfires, tsunamis, pandemics or earthquakes. this research advocates the use of a middle knowledge layer to enable dm practitioners to discern disaster dependent and disaster-independent features in the challenges that they face. for this purpose, the metamodel-based disaster management knowledge repository (dmkr) is constructed. the structure of the repository is an essential factor in obtaining good retrieval results for knowledge retrieval system (shiva and shala ). dmkr structure follows a tailored dm metamodel that recently appeared in (othman, beydoun et al. ). dmkr facilitates knowledge reuse and timely identification of required dm knowledge sources. it allows a unified access to various dm experiences and offers a common dm description language. the work advocates that combining expertise used in various disasters will enable the best approach in managing a new disaster. it provides the following benefits:  facilitating communication among different disaster emergency users;  simplifying the teaching of new dm approaches through a set of semantic rules;  providing guidelines for creating dm models which can cover specific phases of dm (e.g: response phase earthquake emergency response model, mitigation phase bushfire risk reduction model);  highlighting scope for improvement in a dm practice through validation against other existing metamodels in dm field. we developed and validated a dm language as a middle layer of knowledge in the form of a disaster-independent metamodel to unify knowledge from different disaster experiences in (othman, beydoun et al ). this required tuning the metamodelling process to dm to ensure that appropriate knowledge sources (candidate models) are identified and used as input to the process. the resulting language, dmm, was theoretically validated showing how dm models from the literature can be generated from the dmm (othman, beydoun et al ). in this paper, we deploy the language by demonstrating a viable and a unifying dm knowledge repository. this paper develops an actual knowledge sharing system, the dmkr system which stores dm solution modelling components and guides users on how to reconstruct solutions as new contexts arise. to store dm knowledge as reusable modelling components, dmkr has a metamodel driven interface to guide dm experts to articulate their knowledge appropriately. to illustrate the expressive power of the system, a number of case studies to describe real disaster situations are used in the dmkr implementation. besides constructing dm solution models based for specific disaster problems presented to the system, dm stakeholders are also provided with the solutions from a range of disaster categories (e.g.: earthquake, landslide, nuclear meltdown and etc.). this enables mixing and matching solutions across different disasters. typical, users of dmkr can be disaster managers (local, state or federal), monitoring personnel, coordinators, aid agencies or even researchers who may wish to study the domain. the rest of this paper is organized as follows: section describes the dm model driven background and related work that underpin the conceptual foundation of this paper. section describes the development and the architecture of a model driven dm system. section demonstrates the usefulness of the system implementing it and illustrating in real-world disaster problems. specifically, a bushfire disaster exemplar is used to illustrate the knowledge capture and reuse facilitated by the system. finally, section concludes with recommendations for future research. . background and related work a model is an abstract representation of a real world domain typically used to manage complexity (levendovszky, rumpe et al. ). it generally consists of a collection of two elements: concepts and relationships. concepts characterize domain entities and relationships characterize links between them (trabelsi, atitallah et al. ). a model explicitly expresses the structure, behaviour and other properties in a domain and ideally has a causal connection to the real world (aßmann, zschaler et al. ). our approach for dm knowledge sharing can be said to be model-driven. it is inspired by model driven systems development (colin, thomas et al. , stahl, voelter et al. ). instead of requiring software developers to detail how a system is implemented, in a model-driven approach software models specify what the functionality and the architecture of the system to be used are (colin, thomas et al. ) via the use of software models developed in a specific language. with syntactically correct models, using only allowed symbols and conforming to rules, the modelling language facilitates sharing of the outcome of the modelling process. developers are able to abstract and share knowledge using models as starting points. to formally describe the semantics underpining a formal modelling language, a metamodel is required, without which semantics of domain models can be ambiguous ( ). applying a model driven approach to share dm knowledge has the added benefit of also making knowledge accessible to non-technical specialists and newcomers to dm. similar to model-driven software development, this also requires a dm modelling language to specify all elements with which any model can be described. model transformations can then enable precise and timely knowledge sharing communication across the various phases of the process (trabelsi, atitallah et al. ). similarly, in dm, these are also required to enable the unified access to various dm knowledge models and to also communicate the knowledge across different disasters and dm activities. the metamodel we use in this paper to specify dm modelling constructs is dmm. this is a dmm specific metamodel which we synthesized and validated created in (othman, beydoun et al. ). dmm has four sets of concept classes: the mitigation (shown by figure ), preparedness (figure ), response (figure ) and recovery (figure ) class of concepts. each set represent a corresponding dm phase. this clearly describes the dm domain to its users. the relationship between the metamodel and domain models are described through model transformations (vytautas stuikys ) which convert one model to another model (omg ). in this work, we deploy the transformations prescribed in the metamodelling framework of mof, the standard for software metamodelling offered by a omg ( ). mof is essentially a meta-metamodel that defines a common way for capturing the diversity of modelling standards and interchange constructs that are used in model driven software engineering. in (cook ), the mof is defined as a framework designed for defining modelling languages. mof is based on a hierarchical architecture of four meta-layers as illustrated in figure , mof is primarily about defining information models for metadata. it uses an object-modelling framework that is essentially a subset of the uml core. the main four modelling concepts in mof are classes, association, data types and package. the main advantages of omg standard are wide acceptance, coverage of many domains and cohesion among the standard (m. picka ). mof has four layers, m , m , m and m . it has different views on modelling at different layers of details. it is strictly hierarchical. concepts at any given layer (below m ) belong to a concept from the layer above. any concept in any given layer (above m ) can be instantiated at the layer below. m -level is reserved for meta-metamodel element, comprised of the description of the structure and semantics of meta-meta data. m -level is for the metamodel layer (instance of meta-metamodel), comprised of the descriptions of the structure and semantics of metadata. m -level is designed for the model layer (instance of metamodel), comprised of the metadata that describe data in the information layer. the lowest level, m , is dedicated for user models (instance of model and also called as information layer). in this paper, m -level will cover the data that a disaster model describes at m . figure the derivation of various dm solution models in three mof modelling layers figure the dmm - mitigation-phase class of concepts the system introduced in this paper, dmkr . , represents how all the components associated the dmm are operationalised. it is inspired by method engineering, where a metamodel is used to index a software development methodology as required by a software process model (firesmith ). the design of dmkr and the theories underlying it are described in the next section. figure the dmm - preparedness-phase class of concepts figure the dmm - response-phase class of concepts . the disaster management knowledge repository system (dmkr) this section describes the architecture of the knowledge repository, dmkr . . this repository utilises dmm as a representational infrastructure to store and structure complex dm knowledge and processes. dmkr integrates the metamodel with retrieval and modeling support components. it has a model driven interface to ensure that users can easily access and reuse the stored knowledge (see figure ). the stored knowledge can be reused to develop varying dm solutions as dm contexts vary. it has a modelling tool to develop variety kinds of new dm models based on past dm problems and knowledge retrieved. new dm models can be derived through operationalisations of vertical model transformations from dmm. the transformations used are: instantiation (to instantiate a single concept), and conformance (to instantiate a model holistically from a number of concepts). for example, an emergency response model (an m -model) is a model conformance from the dmm (figure ). at this level, m -fragment is used to represent dm models. an m solution derived from dmm can in turn derive various models of real-world disasters at the m -level e.g.: response after a nuclear meltdown, rescue and search after an earthquake, flood rescue model etc. this m -level is the dm user model and represents real-world dm activities sought by dm users. concepts at this level are represented through m -fragment. the combination of all m -fragments would form the dm solution model (m -user model) from the system. figure the dmm - recovery-phase class of concepts . dmkr architecture dmkr design has three components (see figure ): (a) module a – dmm and its tools component: this contains four sub-components: consistency checker module, metamodel formulation module, modelling constructor module, dmm concepts database. each of these will be detailed in section . . (b) module b – the dm knowledge base (the “dm knowledge repository”): the repository schema of this database follows the structure of its metamodel. it stores a collection of emergency organisation and coordination activities executed by different agencies. (c) module c – the retrieval module (the “dm retrieval”): this module functions to assist in deriving the best dm model solution according to the disaster in hand. the dmkr is built using a relational database management system (rdbms). the usage of rdbmss in metamodelling applications is common, e.g. in geosciences (viswanathan and schneider ), for geographic data (shanzhen and yuntao ) and web-based interface components (panesar-walawege, knutsen et al. ). a collection of relational database tables is created to manage the variety of modelling tools element in dmkr. these include: model fragments (m -fragment, m -fragment, and m -fragment); model database (m -dmm, m - model and m -user model); concept and class elements (terminology, notations); dm knowledge (data for model fragments); user information (dm real-user); and disaster information (type of disaster). figure the main page of the dmkr system figure the dmkr system architecture . knowledge representation in dmkr in dmkr, modelling structures to support specific dm activities and related past experiences to these activities are available side by side. the experiences underpin the dm knowledge stored as model fragments. in dmkr, knowledge is populated in a structured knowledge storage unit known as model fragments - the building blocks of reuse. these reusable knowledge units can be mixed and matched as dm context demands. the use of modelling fragments is adapted from method engineering field where a fragment describes a cohesive method component (firesmith ). dmm itself is expressed as a collection of m -fragments making statements about what can be expressed in valid dm models. each concept in dmm has its own m -fragment. in this research, a fragment consists of dm operations and modelling structures that are prerequisites for the applicability of the operations. following the mof metamodelling framework, a model fragment is available at three levels of abstractions (table ): the m -metamodel level, the m -dm model, and the m -dm user model. the m -fragment is used to represent fragment for concept in the m -metamodel (dmm), the m -fragment is used to represent an instance of concept in the m -dm model, and the m -fragment is used to represent information of the instance of concept in the m -dm user model. when an m -model is developed, dm knowledge of every single unit fragments of the model can later be retrieved and reused. the user is guided to user identify relations between unit fragments and dm knowledge. this offers fast and precise knowledge derivation. object oriented inheritance links these fragments across the three levels of mof. for dmm, when a model fragment is indexed by the metamodel, the actual operational dm knowledge will need to be generated as m -fragments. to enable effective usage of concepts to create combination of concepts at the m -level and m -level, dmkr users are given explanations about domain concepts in the metamodel level. these explanations provide examples and more details of usage. table representation of dmm concept in mof modelling paradigm model level level name description elements m dmm metamodel  defines the language for specifying a dm model  consists of metaclasses, metaattributes, metaoperations, metarelationships of dmm concepts m -fragment (metamodel fragments) - contains dmmclass (e.g.: concept: “evacuation”, “aid”, “emergencyplan”, “warning”) m dm model instance of dmm metamodel  models a specific information of dm problems ( e.g.: emergency preparedness model – all the required components for this model are defined in dmm) m -fragment (data model fragments) - contains dmmobject (e.g.: a) setupevacuationcentre (from “evacuation”), b) refugeeshelterdonation (from “aid” concept), c) identifydamagecost (from “damageassessment” concept), d) disseminatealertwarning (from “warning” concept) m dm user model instance of dm model  models user’s real world data  dm model represents a real dm m -fragment (user model fragments) - contains dmminstance of real world data (e.g.: data on evacuation during flood disaster in brisbane, aid distribution during japan tsunami) the object-oriented paradigm is used to design and implement model fragment in dmkr. every dmm concept is represented as class, dmmclass. dm concept::= dmmclass (note: symbol ::= is an equivalent). each dmmclass has the following fields: name, id, terminology, attribute, operations, and relationships. that is: dmmclass::= where:  dmmclassname represents a dm concept name  dmmclassterminology represents a dm concept definition  dmmclassattribute represent a dm requirement  dmmclassoperation represent a dm task  dmmclassrelationship represents the relationship with other concept/s  dmmclassid is a unique identifier for the concept each dmm concept is represented as a dmmclass which has its own m -fragment, a unit fragment. the coupling between the dm operations and the contextual knowledge required for their applicability, in unit fragments (specific for this research, a model fragment is used to represent this unit fragments). each concept will have its own concept’s task, requirement, plan, responsible user etc. these are refined down the mof hierarchy through the corresponding unit fragments. an example of this is shown in table giving an example of a dmmclass concept, evacuation, a preparedness-phase concept and how it becomes refined from m to m according to dm context. table the dmmclass of evacuation concept class evacuation classid bp definition: an organised, phased and supervised withdrawal, dispersal or removal of civilians from dangerous or potentially dangerous areas, and their reception and care in safe areas. relations preparednessplan (bp ) relation type: association, relation name: “follows” warning (bp ) relation type: association, relation name: “activates” a tt ri b u te s evacuationname: the name of the process (e.g.: queensland flood evacuation). evacuationroad: the network of roads where evacuees will travel to reach a safe evacuation destination. evacuationcentres: the safe destination for people who live in disaster risk areas as a temporary refuge. evacuationtransport: the transportation facilities that have a capacity and need for resources (e.g.: fuel) to travel (e.g.: ambulance, helicopter, fire engine). evacuationalertprocedure: communication of public warning. evacuationauthority: leaders who have the authority to order evacuations. evacuationexercise: a program to practice the need to evacuate under certain circumstances and play specific actions to be taken when a real disaster happens. evacuationserviceteam: availability of emergency services facilities and personnel (e.g.: personnel and equipment for search and rescue, hospital and medical services). hazardzone: the zone areas where the disaster took place which has one or more locations (e.g.: north queensland during cyclone yasi ). locationatrisk: the location where people at risk should be moved (e.g.: townsville and innisfail in north queensland during cyclone yasi ). peopleatrisk: people who live in hazard zone areas who must be evacuated. evacuationplan: the plan of evacuation that needs to be followed by all stakeholders involved (e.g. “emergency management australia (ema) flood evacuation plan”). o p e ra ti o n s performpublicevacuationprocedure(): a process to implements the evacuation procedure by all victims and emergency teams. alertwithdisasterwarning(): a situation which requires people to listen to a nominated radio station or watch a nominated tv channel for further advice regarding the evacuation warnings. makeevacuationdecisionwarning(): a process of the authority making a decision as to whether to evacuate or not, which will be assisted by the availability of timely and relevant information. broadcastpublicdisasterwarning(): a process of disseminating the evacuation information and warning through a variety of media (print, electronic, community announcement, word-of-mouth, etc). evacuatepeopleatrisk(): a process of evacuating the victims by the response emergency teams with a special consideration of special needs victims. setupandorganizeevacuationcentre(): an establishment and management of the evacuation centres which provides basic human needs including accommodation, food, water, first aid, hygiene facilities, clothing, blankets, information and referral services, among others. returnevacuee(): a final stage of the evacuation process. it will be necessary to assess the disaster area to determine if return is possible and identify any special conditions which may need to be imposed. . dmkr detailed design we use a database management system to implement the components of dmkr. it has these four components to support modelling, storage and retrieval : a) consistency checker module, b) metamodel formulation module, c) modelling constructor module, d) dmm concepts database. each of these components is detailed in this section. . . consistency checker module model transformations, at m , m and m levels, are specified via a set of rules defining the modelling constructs and model composition mechanisms for its dmkr modelling design activities. the consistency checker module is a library of these rules which act as policy guidelines for the correct utilisation of the dmm modelling components. the module ensures compliance of model transformations to these rules. the rules define semantic properties of metamodel transformations as prescribed in wachsmuth ( ) and lammel ( ). thus a consistent transformation from a source model (dmm) to a corresponding target model (either an m -model or an m -user model) is guaranteed. figure shows a sample of six rules (for the rule_id of r c, rr a, r , rr a, and rr ). for example rule rr a formalises the creation of an “instance model of dmm”: dmm rule rr a rule syntax : Ґ(µdmm) := [ ι (µdmm) |= µdmm ] rule meaning : the set of all dm model (metamodel instances) conforming to a disaster management metamodel, µdmm rr a rule ensures each newly created dm model (metamodel instances) is compliant to the dmm, (µdmm). this rule will mostly be used in association with other rules according to the type of model being developed. for example, if user a wants to develop a new preparedness model from the dmm, a rule specific to initialise the element is used: dmm rule rc b rule meaning : the set of concepts for a preparedness model rule syntax : cm (µdmmp) ::=[preparednessplan, preparednessorganization, preparednesstask, suppliesregistry, warning, preparednessgoal, evacuation, beforedisaster, event, decisionmaking, administration, emergencypublicinformation, pre-position, disasterfactor, exposure, disasterrisk, training, preparednessteam, media, mutualaidagreement, publiceducation, publicawareness, resource, monitoring, aidagency] figure an extract form the rules repository . . metamodel formulation module dmm classes are stored as a concept_table and dmm corresponding phase classes are stored in phase_table. figure shows the relationship between both tables in which: a) shows a table’s relationship, and b) shows data values of each tables. in this figure, the mitigationplan is a concept in the mitigation-phase class (phase id = ). in concept_table, the concept_id of the concept has the same id with the m fragment_id, am . this shows how one m -fragment exists for every single concept in dmm. to formalise the usage of concepts in each m - fragment (a dmm class), specific rules in the consistency checker module are used. the following four rules formalise the instantiation of a concept (rule r a and r b) and a conformance of model (rule rr a and rr b) for two m -fragments (the mitigation and preparedness class): dmm rule r a rule syntax : cm (µdmmm) rule meaning : mitigation concepts of the disaster management metamodel rule construct : mitigationconcept::= () dmm rule r b rule syntax : cm (µdmmp) rule meaning : preparedness concepts of the disaster management metamodel rule construct : preparednessconcept::= () dmm rule r a rule syntax : ι(µdmmm) rule meaning : mitigation model of disaster management rule construct : mitigationmodel::= ( and ) and ( and ) dmm rule r b rule syntax : ι(µdmmp) rule meaning : preparedness model of disaster management rule construct : mitigationmodel::= ( and < preparednessm fragment>) and ( and < preparednessm fragment>) figure relationship between concept and phases; (a) concept_table and phase_table, (b) data values of the concept_table and phase_table rule r a ensures that every single mitigation-phase concept consists of a dmmclass and a model fragment. rule r a ensures that the mitigation-phase model (dm solution model of dmkr) consists of an m -model and an m -dm user model derived from the mitigation-phase class. the different fragments and their associated models are managed through the storing of data for fragment (m fragment_table, m fragment_table and m fragment_table) and model (m model_table, m model_table and m model_table). a single dmmclass has only one m -fragment and the combination of all m -fragments form the m -model (metamodel). as shown in figure (a), the dmm has many m -fragments in the model. as shown in figure (b), at any one time many m - models may consist of many m -fragments and many m -user models may consist of many m -fragments. table shows an example of m -fragments with its respective unit fragments that belongs to the masscasualtymanagement class. these fragments underpin the formulation of m and m -fragments and models. table part of unit fragments of the masscasualtymanagement class unit id unit name unit description cr a tactical procedure the procedure which provides standard guidelines for mass casualty operation in a multi-casualty emergency medical situation. cr b organization structure the organisational structure of multi-responsible teams who are responsible for the overall management of multi-casualty incidents (e.g.: patient transportation group, medical supply group, medical communication group, victims treatment group). cr c equipment the equipment needed to perform a mass casualty operation (e.g.: disaster kits containing bar coded wrist bands, envelopes for storing victim's items, sealed bags for storing dead bodies for medical treatment or mortuary). ocr a mass casualty setup the implementation of mass casualty procedure (e.g.: prepare disaster kits containing: bar coded wrist bands, different sized envelopes for personal effects, a collection of all items taken from the victims, dead or alive, and place them in the bags. sealed bags are transported with the victims for medical treatment or the mortuary). ocr b classify casualty classifying the victims of a disaster (e.g.: victims who died in the medical post (mp) certified by the mp doctor. victims who died in the field need to be taken to the hospital and certified by a qualified doctor). ocr c move patient the movement of patients from the scene of the incident to medical facilities in close coordination within the treatment dispatch manager(s) in the medical group and the medical facilities. figure shows the m -fragment of an evacuation class with its corresponding unit fragment. m -models describe a specific dm problem. an m -model is formed by many m -fragments. m -fragments can be instantiated to represent many m -fragments (real world data). examples of suitable candidates to be represented as figure the m -fragment of an evacuation class with its respective unit fragment figure the m -model of a disaster mitigation model m -models are: preparedness operations before floods, the coordination of resources during a flood response, and the reconstruction and rehabilitation following a bushfire. in section . , real-world dm problems for various bushfire problems are used to implement m -models. for example, a disaster mitigation model is the m -model conformance from the mitigation-phase class. this model is created from a combination of a few dmmclass: the structuralmitigation, the insurance, the non-structuralmitigation, the needsplanning, the riskanalysis etc. by using a few unit fragments taken from each class, this model can later be formed. the relationship between the m - figure an m -fragment of the homeradiationmonitoringdetection, is instantiated from an m -fragment of the performpublicemergencyplanmodelling decision associated with activity tasks activity tasks setupstructuralmitigationplan(), a unit fragment of the structuralmitigation concept model and m -fragment can be described as for each m -model; the model has many m -fragments. in figure , one unit fragment of a ‘setup structural mitigation plan’ is instantiated from the structuralmitigation class. for this instantiation process, a special notation is used to link a relationship between model-metamodel. the stereotype notation of (<>) is used to signify that the setupstructuralmitigationplan() is a class operation that comes from the structuralmitigation class (shown by an arrow). other instances of unit fragment taken from the dmmclass are: a ‘perform risk analyses’ from the riskanalysis, and a ‘perform hazard assessment’ from the hazardassessment class. figure shows a sample of one m -fragment, home radiation monitoring detection, in dmkr that represent dm knowledge of a nuclear radiation problem. this is instantiated from the m -fragment, performpublicemergencyplan(), which in turn is derived from the m -fragment, preparednesstask. in this m -fragment, the real activity of how a home radiation monitoring detection process should be conducted during a nuclear disaster situation is demonstrated. . . modelling constructor module the modelling constructor module contains the library of notations/symbols that can be used by system users to develop their semantic dm model in dmkr. in this paper, modelling notations offered by the unified modelling language (uml) are used. a model derived from dmm can be modelled according to uml-based representations: a) ‘class diagram-view’ (instantiation of concepts is in form of class), b) ‘state transition diagram-view’ (mapping events – start, stop, condition etc.), c) ‘use case diagram-view’ (model actors and their respective functions), or d) ‘sequence diagram-view’ (model activity sequence order). modelling constructor module is the library that provides notations for customisation in dmkr. the representation of a derived m -model during a conformance of the model from the dmm is another issue in dmkr. as described by brottier et al. ( ), a transformation offers an opportunity to improve the readability of the m -model rather than use the same notation. sometimes, improving the readability of the model may involve hiding some relationships in the m -model. therefore, when users create the m -model in a class diagram-view, the representation of the model can be seen exactly as its conformant metamodel (because the dmm is developed in this view). however, in other certain cases, users can also present their new m -model in a human-friendly representation such as the state transition diagram or in use case diagram (uml models). dmkr gives flexibility for a system user to create their model in the modelling environment provided in the system. . . dmm concepts database this database is implemented as the following three smaller databases: . the dmmclassterminology database: this contains dmm concept definitions. . the dmmclassrelationship database is the collection of relationships between concepts within and across mitigation, preparedness, response and recovery-phase class of concepts (m -level). these relationships guide users on integration of model components (concepts) as they create a new model for their dm context. for instance, when a user creates a new m -model model of the “earthquake rescue model”, one of the important elements that will be derived from the response-phase class is the rescue concept. in the m -fragment corresponding to this concept, there are seven relationships to other concepts that users need to consider in formulating their m -model e.g. “isperformedby” linked to the emergencymanagementteam concept or “follows” linked to the “emergencyplan” concept. . the third database specifies the attributes and the operations of each class. to illustrate this, figure shows three attributes for the evacuation concept: evacuationroad, evacuationcentres and evacuationalertprocedure. it also shows three operations: performpublicevacuationprocedure(), evacuatepeopleatrisk() and setupandorganizeevacuationcentre(). . demonstrating dmkr in real-world disaster problems: bushfire case study in dmkr, system users are given flexibility to prepare their model. how a model can be formed is varied according to the intention of the user for their new model. this section illustrates how knowledge can be stored and later retrieved and reused using dmkr. an focusing on bushfire knowledge is shown. in section . , the storage of bushfire knowledge is shown. in section . , retrieving this knowledge according to varying bushfire scenarios is illustrated. . storing bushfire knowledge in dmkr bushfire management knowledge is typically available in sources in text form available in reports, books and disaster-related websites. storing this knowledge begins following the structure of the m -fragment, dmmclass). for starters, knowledge for the responsetask for a bushfire is stored, based on previous bushfire experiences describing how various dm users operate during the response phase. their specific actions and roles are stored. knowledge is collected from the following sources: a website (johnston ); a report (emergency management australia (ema) ); and four books (paterson ), (troy and kennedy ) (machin and hentze ), (stephens and collins ). the stored knowledge combines knowledge of different groups of stakeholders. for example, solution taken from stephen and collins ( ) provides knowledge of three different groups: i) media, ii) emergency team, and iii) affected people and community, whereas, a solution from machin and hentze ( ) provides knowledge for the emergency team and monitoring user. the knowledge stored in the m -fragment of the responsetask concept represents tasks of different users. later other groups of users (e.g.: task of government, task of people at risk or task by media and etc.) will be able to reuse this knowledge. figure samples of data values for the attribute_table and the operation_table further bushfire knowledge is sourced from the report of the ‘ bushfire royal commission final report’ (parliament of victoria ). this knowledge is presented through figures . (a) to (c). the report describes the real experiences of the bushfire tragedy of february in victoria, australia. this was a huge tragedy, claiming lives, destroying at least homes, rendering people homeless and releasing millions of tonnes of carbon into the atmosphere. figure (a) describes knowledge of bushfire factors related to this tragedy. the knowledge is stored as an m -fragment of the disasterfactor concept. in figure (b), the recommendation from the report describes knowledge for structuralmitigation. figure (c) contains knowledge for the legislation concept captured from recommendation . unlike knowledge stored in the responsetask concept which categorises bushfire knowledge according to users: government, people at risk, media. figure categorises the knowledge according to dmm concepts: disasterfactor, structuralmitigation and legislation. data presented in figure (a) is a collection of knowledge representing identified factors in the victoria bushfire. this is stored in dmkr as illustrated in table , as dmmclassattribute of the disasterfactor concept: the gravityfactor, the triggeringfactor and the complexityfactor. (a) victorian bushfire factors knowledge data, the new case entry for the disasterfactor concept in preparedness-phase class (b) recommendation of ‘ bushfire royal commission final report’ (parliament of victoria ), the new case solution for the structuralmitigation concept (c) recommendation of ‘ bushfire royal commission final report’ (parliament of victoria ), the new case solution for the legislation concept figure bushfire experiences knowledge derived from one knowledge source: bushfire royal commission final report. the experiences knowledge are fragmented into three different dmm concepts: (a) disasterfactor (b) structuralmitigation (c) legislation concept is stored as an m -fragment of the structuralmitigation  rapid fire spread followed ignition, which responding crews could not contain.  fires crowned in forested areas, which made them impossible for ground crews to control.  powerful convection columns were generated above the fires.  extensive forward spotting occurred as a result of the fuel type, the weather conditions and the topography.  late in the day a wind change altered the direction of fire spread and extended the fire front. is stored as an m -fragment disasterfactor of preparedness phase is stored as an m -fragment of the legislation detailed disasterfactor knowledge is structured into small unit fragments of m -fragments (in the disasterfactor concept), as follows: the dmmclassattribute of the gravityfactor: ) temperature: - jan - consecutive hot days above c ( f), with peaking at c ( f) on jan – rd hottest day in the city’s history. hottest: . ( f). ) hot tropical air to south eastern australia is caused by a) heatwave cause by a slow moving high-pressure system over the tasman sea; b) intense tropical low located off the north west australian coast; c) monsoon trough over northern australia; and d) the weather conditions and the topography. late in the day a wind change altered the direction of the fire spread and extended the fire-front. ) low humidity. the dmmclassattribute of the triggerfactor: ) power lines ignition - rapid fire spread followed ignition, which responding crews could not contain, ) lightning, cigarette butts, ) sparks from a power tool - powerful convection columns were generated above the fires. the dmmclassattribute of the complexityfactor: ) major drought more than a decade, ) -year warming trend link to human induced climate change, ) extensive pyrocumulus clouds (most intense firestorm ever experienced in australia history). table storing knowledge of disasterfactor of the victorian bushfire . deriving new bushfire solution models from dmkr in this section, the functionality of dmkr as a dm modelling tool is presented. in dmkr based modelling, users are given flexibility to utilise the functionalities facilitated by dmm. in the rest of this section, the modelling capabilities of dmkr are illustrated in two illustrative bushfire scenarios. scenario presents how knowledge retrieved directly from the specific m -fragments can be used as the inputs to create an m -model of the aerial fire fighting strategy model. the models required in this scenario are developed bottom up. in scenario , knowledge of an m -model initially created in m -level is refined into m -fragments of two m -user models. the models required in this scenario are developed top-down. the victorian bushfire factor knowledge is stored to the disasterfactor concept through its m -fragment of the ‘gravityfactor’ scenario : supporting aerial fire fighting model process using dmkr (m ) knowledge in this scenario, the reuse of knowledge retrieved from various m -fragments to support the user to generate an m - model is shown. the user here aims to develop the aerial fire fighting model. the development of this model requires other knowledge that needs to be studied before it can be constructed. for instance, knowledge of many aerial fire suppression factors is crucial for planning an efficient aerial bushfire fire fighting operation as described next. scenario : “the use of aerial fire fighting resources has grown steadily over the past few decades and aircrafts now play an integral role in fire management and suppression in victoria. in - the then department of natural resources and environment operated the largest fleet of firebombing aircraft ever used in australia including, for the first time, an erickson air crane, a large helicopter that is used for firebombing. an evaluation of the effectiveness of this fleet of aircraft was commissioned by department of sustainability and environment (dse) and published in . a bushfire cooperative research centre (crc) team tries to create a framework model of the effectiveness and efficiency of aerial fire-fighting in australia. key findings from this study should include: the effectiveness of aerial fire suppression depends on many factors, including suppressant agent used (retardant, foam or water), aircraft characteristics, drop characteristics, fire intensity, fire size, fuel type, pilot skill, aircraft travel time, distance from fire, ambient conditions, availability of support ground resources, and organisational and infrastructure arrangements.” ( victorian bushfires royal commission ) for this kind of problem, a study of different aerial fire suppression factors such as “suppressant agent used, aircraft characteristics, drop characteristics, fire intensity, fire size, fuel type, pilot skill, etc” are required. dmkr can support this by providing dm knowledge of the aerial fire suppression factors from its repository. these factors can be derived through the m -fragments of related dmm concepts. for example, through an m -fragment of the responsetask concept, the dmmclassattribute of the concept, the ‘responsetaskfactor’ stores knowledge of various real-world responses task operation’s factors in the m -fragments of the concept. particularly for the ‘suppressant agent used’ factor, knowledge earlier sourced from emergency management australia (ema) of can be reused. a sample of the relevant m -fragment in dmkr is as follows: ... if fire-fighting objective = ("for common solid and combustible materials") then use = ("water") if fire-fighting objective = ("for extinguishing fires that is flammable and combustible liquids") then use = ("foam") if fire-fighting objective = ("effective as a smothering and chemical chain reaction") then use = ("extinguishing powder") if fire-fighting objective = ("for extinguishing small fires and fires involving electrical equipment") then use = ("carbon dioxide") if fire-fighting objective = ("for extinguishing chemical chain reaction-inhibiting") then use = ("vaporizing liquid") if fire-fighting objective = ("to react to form a smothering soap layer on the surface of burning fats") then use = ("wet chemical") if fire-fighting objective = ("for extinguishing combustible chemical") then use = ("special agents") … in addition to the above, dmkr contains knowledge of similar processes (choosing a right fire fighting agent) implemented by other countries. this knowledge can support the crc’s team in their study. in addition, knowledge other factors, for example, the ‘aircraft characteristics’ can also be retrieved for reuse. ‘responsetaskequipment’ is another dmmclassattribute of the responsetask concept. this knowledge is also sourced from practices of other countries and is retrieved straightforwardly as m -fragments. these include how: a) china uses “metropolitan aerial fire fighting system for the ec helicopter simplex model ”; b) usa through its los angeles county fire department uses “sikorsky s- c fire hawk”; c) italy through italian civil protection department uses “bombardier canadair cl aircrafts”; d) ukraine uses the “antonov aircraft”; and e) germany uses the “kfor helicopters etc (sources from euro-atlantic disaster response coordination centre (eadrcc) )” in their aerial fire fighting operations. efficiency of an aerial fire-fighting strategy also depends on fire behaviour analysis. through an m -fragment of the monitoring concept, the dmmclassoperation of the ‘eventanalysis’ represents knowledge of how to monitor analyses disaster data. specific for the purpose of the crc team, the following m -fragment for the bushfire fire behaviour analysis is also reusable: --------------------------------------------------------------------------------------------------- activity tasks: eventanalysis (m -fragment of the monitoring concept: ) use current bushfire events data (humidity percentage, air temperature degree celsius, average wind speed, slope position, fuels loads, fuel type, fuel mixture, steepness, vegetation communities) ) based on current events data, perform fire prediction analysis (fire prediction, fire forecasting, fire environmental model, calculating fire spread speed, estimating probability of fire ignition, fire evaluation based on weather, calculating fire danger index) and perform fire behaviour prediction analysis. ) send event analysis results to emergency operation centre. fire behavior prediction analysis: i) fire spread speed: [ d = . ( + . v) ] description: this subroutine representing fire spread speed to leeward. where v is a wind velocity meter per second (m/sec) and d is the limit of distance which fire can spread (m). ii) fire spreading probability: [ fij = a. (sij . pij). w bij. p (tckl)] description: the fire spreading judgement, that is whether or not cell ij changes from one state to another, is done by using the fire spreading probability of cell ij. the probability is given by the fire spreading judgement index fij. this is calculated for all cells ij within the neighbourhood of cell kl and defined by this equation. iii) point of fire origin: [ tx = ( + a/ + d/d) / ( + . v)] description: this equation shows the time until an adjacent house to the point of fire origin ignites after catching fire in the origin. where, a is a length of average of a side in building (m), d is a pitch of building (m) and d is the limit of distance which fire can spread (m). iv) continuation time of fire: [ty = (w/ . )/(aw /'h / af) description: the time t until cell kl burns out after catching fire is defined in this subroutine. where, w is a fire load (kg/m ), aw is a opening area (m ), af is a floor area (m ) and h is a height (m). this equation approximately represents the continuation time of fire as assuming wooden building and constant combustion speed v) process of fire spreading: [ if nij(t) = and fij > ran then nij(t+ ) = ] description: fire spreading judgement index fij of a selected cell ij is calculated in fire spread model. if fij is satisfied with the following requirements, the set of cell ij becomes nij = (catching fire). this subroutine calculates nij. where nij (t) is the state of cell ij at simulation time t and ran is the random number which takes from to vi) fire extinguished utilisation: [ ce = (vw/(cw.tf)). rw] description: the number of cells of which the fire can be extinguished with water utilisation, ce, is given in this equation. this is based on the assumption that one nozzle continues spraying water to a cell of state until extinction of the cell …. scenario : creating process models (m ) and user models (m ) for a new bushfire mitigation plan process the derivation of an m -model using dmkr to setup a bushfire mitigation plan is presented in the following scenario: “the head of operation engineer at a government-owned electric power company (swepco) in australia has been requested to prepare a bushfire mitigation plan for his company. swepco is a company which has the responsibility for power sector operations such as dispatching electricity to different public local areas, generate electricity, operates power hydro stations and many more electrical operational works. the requirement is to develop a specific bushfire mitigation plan for this company as a result of many recent reported bushfire events in several of the company’s client’s locations. therefore, a company preparedness strategy is required to overcome any unforeseen bushfire events in the future. as this will be the first time for this swepco head engineer to prepare a plan that he has no prior experience of, this user definitely requires some support for preparing this task. the information of how to start the mitigation plan process, how to set up a strategic planning committee, which rules and legislations are relevant to bushfire problems, or what are the structural or non-mitigation appropriate for the plan are important for this user.” in this scenario, the swepco’s engineer will be directed to the mitigation-phase class represented in the metamodel formulation module. this is because the ‘setup mitigation plan’ is recognised as a part of mitigation activities by the consistency checker module. the engineer is then offered information on all the concepts (dmmclass in mitigation-phase) from which he is able to derive his own dm model (m -model of setup mitigation plan for swepco company). such concepts exist. class diagram-view (figure ) is one m -model possibility which contains concepts in the mitigation-phase class selected for his model. this now becomes the m -fragments of the swepco setup bushfire mitigation plan model. based on the m -model created by this user, the next step is to use dmkr to generate data for his m -model’s m -fragments. the collection of existing dm knowledge case solutions is stored in a repository that matches the inputs (bushfire knowledge) for the m -model entered by this user that will be retrieved. for this example, two such models are available in dmkr based on sources earlier collected from powercor australia ltd (powercor australia ltd ) and the united energy company (united energy ). these provide bushfire mitigation knowledge specific for an electricity company which match the intention of the m -model (mitigation plan model - figure ) as specified by the user. generally for such users, electricity networks are a serious fire ignition hazard. prevention is very important due to enormous cost associated with such disasters. for powercor company, the bushfire mitigation model includes processes for asset inspection, maintenance, construction, upgrading, replacement, vegetation management, performance monitoring and auditing processes. this strategy plan applies to assets that could cause fire ignition in areas designated as hazardous bushfire risk areas. for the united energy company, its bushfire mitigation plan model concentrates more on the policies and procedures used to demonstrate the commitment from all levels of management within the company to the minimisation of bushfire risk due to electricity assets. figure the setup mitigation plan model (m ) derived from the mitigation-phase class the weather conditions for the two m -user models retrieved match swepco’s operating conditions. the bushfire conditions in australia faced by the two companies are equivalent to swepco’s problem situation. information derived from powercor and united energy provide relevant and significant bushfire knowledge for swepco’s plans formulation. for instance, specific to the structuralmitigation, based on the derived knowledge retrieved from the two sources, different kinds of structural mitigation processes are found to be conducted in their mitigation plans. the structural mitigation plans of powercor include : a) the management of vegetation around power lines, b) inspect private overhead electric lines, and c) the implemented new technologies to minimise risk of electricity assets causing fire ignition, whereas, for the united energy the plan stresses: a) electric lines clearance, b) asset management system, and c) procurement of equipment and services. from the retrieval of this knowledge, it is identified that the procedure used (data in the m -fragment) to check the technology and equipments standard in each plan is different. powercor uses the ‘technical standard ea - overhead - bushfire mitigation’ and united energy uses the ‘equipment materials approval document no.jen pr wi ’. this information gives richness to the bushfire mitigation plans for the swepco m -model. hence, by using the m -model created in figure , the real-world bushfire knowledge source of the two companies will be presented as the retrieval result for the swepco’s new m -model. the two new m -user models are shown in figures and . figure the retrieval of model solution (m ) for the powercor mitigation plan (m ). figure the retrieval of model solution (m ) for the united energy mitigation plan (m ). the m -user model source from powercor australia the m -user model source from united energy even though sources from powercor and united energy are chosen as a solution for the swepco case, the repository of dmkr also has a population of others similar model samples: the berlin wildfire hazard mitigation plan from new hampshire, usa (north country council ), the fire hazards mitigation plan from the navi mumbai municipal, india (navi mumbai municipal corporation ) and many other fire mitigation solutions. in this section, the functionalities of the metamodel-based dmkr system were demonstrated in real world dm scenarios. specifically, real-world dm bushfire scenarios were visualised. collecting and storing of bushfire knowledge experience from different bushfire knowledge sources showed the capability of dmkr for dm knowledge storage. the collected knowledge was later reused to solve other incoming bushfire problem scenarios faced by other users. this process demonstrated the potential of this system as a dm modelling tool. moreover, derived solutions were also reusable and this was also demonstrated. utilising dmkr, dm users in different bushfire scenarios were presented with new derived dm solutions suitable for their real world problems. . conclusion and future works this paper illustrates a new a metamodelling based approach to guide dm practitioners to store and reuse dm based on a disaster management metamodel (dmm). it presents a knowledge sharing system, dmkr . , to uniformally store and structure dm knowledge. this is facilitated by a specific a dm language. dmkr facilitates the derivation of new dm solution models using the dm language, underpinned by the metamodel dmm. dmkr successfully combines the use of dmm for storage, retrieval and modelling to describe modelling capability for any specific disaster. to dm practitioners therefore, the paper offers a coordinated approach and activities to achieve dm goals where dmkr can act as a central tool for creating, organizing and managing dm modelling knowledge. through fragmenting dm knowledge experience for reuse, dmkr guides dm users to achieve specific dm goal. many stakeholders are either directly involved (e.g.: disaster manager, emergency coordinator, fire fighters..) or indirectly involved in dm operations (e.g.: local businesses, insurance companies etc..) but they do not have a complete view of how different dm activities can be conducted. dmm through its four sets of classes (response, recovery, mitigation and preparation classes) give an overall picture of how all dm actions are executed. with better awareness and understanding of the whole dm processes, many benefits can be derived because ensuring success in initial dm phase will lead to success in the subsequent phases. the paper also contributes to semi-formal knowledge management practices by developing a rigorous modelling hierarchy managed and enforced by dmkr and underpinned by dmm. as the ability to offer modelling guideline to many domain users, dmm helps various users to make quicker decisions using earlier created semantic models. it facilitates three levels of modeling as in the mof framework: the m -level for a metamodel, the m -level for a model, and the m -level for a user model. instance models are derived from dmm based on conformance relationships and dm knowledge is structured in the dmkr system using model fragment units. since there are three mof modelling levels are used in this research to describe the dm knowledge, three corresponding types of model fragment (m -fragment, m -fragmnet and m -fragment) are also used. further future research will support knowledge evolution in dmkr and add further reasoning capacities to support dm first responders: the paper presents the construction of formal semantics of relating dm models at different metamodelling levels. this can form in the future a fundamental part of an automated decision support in many disaster scenarios. by populating knowledge of multiple disaster cases for the same unit fragment, a reasoning process can be easily applied. this would be of significance to dmm, as the use of the reasoning technique through metamodel for is not intensively being yet investigated. furthermore, such automated reasoning capabilities can also pave the way for the construction of a real-time dm system using models created from the metamodel for a quick response. since the dmm describes data that should exist in the response-phase models, real-time calculation can be made when fresh disaster data that is being “pushed” from the field by first responders. this can for example include locations of responders, information on the status of response activities, data mapping damage, and sensor data. results from the calculation of the data can be pushed back into the field immediately e.g. responder’s team send field reports, imagery and map updates and requests the current status information on resource deployments. references victorian bushfires royal commission ( ). systemic issues – aerial firefighting submissions of counsel assisting aßmann, u., s. zschaler and g. wagner ( ). ontologies, meta-models, and the model-driven paradigm. ontologies for software engineering and software technology, springer berlin heidelberg: - . beydoun, g. and a. hoffmann ( ). simultaneous modelling and knowledge acquisition using nrdr. th pacific rim conference on artificial intelligence (pricai ), singapore, springer-verlag. brottier, e., f. fleurey, j. steel, b. baudry and y. l. traon ( ). metamodel-based test generation for model transformations: an algorithm and a tool. th international symposium on software reliability engineering (issre ), raleigh, north california, usa. colin, a., k. thomas and hne ( ). "model-driven development: a metamodeling foundation." ieee softw. ( ): - . cook, s. ( ) "domain-specific modeling and model driven architecture." mda journal: a bpt column. cordner, s. m., n. woodford and r. bassed ( ). "forensic aspects of the victorian bushfires disaster." forensic science international ( - ): - . e-objects open source metamodel. ( ). "metamodel database compliancy in open source metamodel." retrieved march . emergency management australia (ema) ( ). disaster preparedness manual for commonwealth agencies. australian emergency manual series, national archives of australia. euro-atlantic disaster response coordination centre (eadrcc) ( ). forest fires in albania, bulgaria and the former yugoslav republic of macedonia. eadrcc executive summary. firesmith, d. ( ). method engineering using opfro. th annual european software engineering process group conference (eurosepg). johnston, s. ( ). "independent community representatives advocate." the australian retrieved - - , from http://www.johnston-independent.com/. kleijnen, j. p. c. and r. g. sargent ( ). "a methodology for fitting and validating metamodels in simulation." european journal of operational research : - . lammel, r. ( ). grammar adaptation. proceedings of the international symposium of formal methods europe on formal methods for increasing software productivity (fme' ), springer-verlag: - . larcombe, p. j., s. e. moloney and p. a. schmidt ( ). pandemic (h n ) : a clinical spectrum in the general paediatric population. archieves of disease in childhood. queensland, gold coast hospital. : - . lauras, m., truptil s., and bénaben, f. ( ) towards a better management of complex emergencies through crisis management meta-modelling, disasters ( ), – levendovszky, t., b. rumpe, b. schatz and j. sprinkle ( ). model evolution and management. mbeerts. berlin heidelberg, springer-verlag. : - . m. picka ( ). "metamodelling and development of information system." agriculture economics ( ): - . machin, b. and m. hentze ( ). comments on the present and future of wildland fire suppression decision- making processes. living on the edge: economic, institutional and management perspectives on wildfire hazard in the urban interface, advances in the economics of environmental resource. elsevier ltd. : - . navi mumbai municipal corporation ( ). fire hazards response and mitigation plan mumbai, india. north country council ( ). berlin wildfire hazard mitigation plan. bethlehem, new hampshire, usa. omg ( ). meta object facility (mof) specification, object management group. omg ( ). mda guide version . . . http://www.johnston-independent.com/ http://onlinelibrary.wiley.com/doi/ . /disa. /abstract http://onlinelibrary.wiley.com/doi/ . /disa. /abstract othman, s. h., g. beydoun ( ). model-driven disaster management. information and management. ( ), - . othman, s. h., g. beydoun and v. sugumaran ( ). "development and validation of a disaster management metamodel (dmm)." information processing & management ( ): - . panesar-walawege, r. k., t. s. knutsen, m. sabetzadeh and l. briand ( ). cresco: construction of evidence repositories for managing standards compliance. proceedings of the th international conference on advances in conceptual modeling: recent developments and new directions. brussels, belgium, springer-verlag: - . parliament of victoria ( ). victorian bushfires royal commission. victoria. i,ii,ii and iv: . paterson, r. g. ( ). wildfire hazard mitigation as safe smart growth. living on the edge: economic, institutional and management perspectives on wildfire hazard in the urban interface, advances in the economics of environmental resource. elsevier ltd. : - . powercor australia ltd ( ). bushfire mitigation strategy plan - : electricity safety (bushfire mitigation) regulations , document no: -m . f. audley: . saltor, f., m. garcía-solaco and p. gargallo ( ). diversity with cooperation in database schemata: semantic relativism. in proc. icis conference. shanzhen, y. and l. yuntao ( ). an intelligent information sharing method of heterogeneous geographic data based on unified metamodel and entity matching. geoinformatics, th international conference on. shiva, s. g. and l. a. shala ( ). software reuse: research and practice. information technology, . itng ' . fourth international conference on. stahl, t., m. voelter and k. czarnecki ( ). model-driven software engineering, technology, engineering, management, john-wiley & sons ltd. stephens, s. l. and b. m. collins ( ). fire policy in the urban wildland interface (uwi) in the united states: what are the issues and possible solutions? living on the edge: economic, institutional and management perspectives on wildfire hazard in the urban interface, advances in the economics of environmental resource. elsevier ltd. : - . trabelsi, c., r. b. atitallah, s. meftali, j.-l. dekeyser and a. jemai ( ). "a model-driven approach for hybrid power estimation in embedded systems design." eurasip journal on embedded systems (article id ): tran, n., g. beydoun and g. low ( ). design of a peer-to-peer information sharing mas using mobmas (ontology-centric agent oriented methodology). in advances in information systems development, w. wojtkowski, w.g. wojtkowski, j. zupancic, g. magyar and g. knapp (eds.). springer us, - . troy, a. and r. g. kennedy ( ). finding solutions to the urban wildland fire problem in a changing world. living on the edge: economic, institutional and management perspectives on wildfire hazard in the urban interface, advances in the economics of environmental resource. elsevier ltd. : - . united energy ( ). bushfire mitigation plan - , document no: ue pl . t. fisher: . vara, j. m., b. vela, v. a. bollati and e. marcos ( ). supporting model-driven development of object - relational database schemas: a case study. proceedings of the nd international conference on theory and practice of model transformations. zurich, switzerland, springer-verlag: - . viswanathan, g. and m. schneider ( ). bigcube: a metamodel for managing multidimensional data. international conference on software engineering and data engineering (sede ). i. rahal and r. zalila- wenkstern. usa, isca: - . vytautas stuikys, r. d., aleksandras targamadze ( ). "a model-driven view to meta-program development process." information technology and control ( ). wachsmuth, g. ( ). metamodel adaptation and model co-adaptation. proceedings of st european conference on object-oriented programming (ecoop), berlin, germany, springer-verlag. xu, d., c. wijesooriya, x.-g. wang, g. beydoun ( ), outbound logistics exception monitoring: a multi- perspective ontologies’ approach with intelligent agents, expert systems with applications, ( ), pp. – . lund university po box lund + - an architecture for expert system based feedback control Årzén, karl-erik document version: publisher's pdf, also known as version of record link to publication citation for published version (apa): Årzén, k-e. ( ). an architecture for expert system based feedback control. (technical reports tfrt- ). department of automatic control, lund institute of technology (lth). total number of authors: general rights unless other specific re-use rights are stated the following general rights apply: copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • users may download and print one copy of any publication from the public portal for the purpose of private study or research. • you may not further distribute the material or use it for any profit-making activity or commercial gain • you may freely distribute the url identifying the publication in the public portal read more about creative commons licenses: https://creativecommons.org/licenses/ take down policy if you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. https://portal.research.lu.se/portal/en/publications/an-architecture-for-expert-system-based-feedback-control(f b cc - fb- e -b -d dffd cf ).html coden: iutfd /(tfrr- ee) i l- i ( es ) an architecture for expert system based feedback control karl-erik Årzén department of automatic control lund institute of technology september department of automatic control lund fnstitute of technology p.o. box - lund sweden documcnt na¡rc¡c report datc of itøuc september l documcnt Àiumbc¡ coden: tutfd /(tfrt- ee)/ - l( e ) author(s) karl-erik Årzén supcrviso¡ sponcoring org¡anísation stu titlc dnd subtìtle an architecture for expert system based feedback control. absttæt it is a recognized problem that many industrial control loops are badly tuned or run in manual mode. two expert system approaches have been suggested for this problem. fuzzy, rule-based control replace control algorithms by linguistic rules which model the operators manual control strategy. knowledge-based control extends the range of conventional controllers by encoding general control knowledge and heuristics concerning tuning and adaptation in a supervisory expert system. an architecture for knowledge-based control is described where two concurrent processes are used for the knowledge-based system and the numerical algorithms. a modular, blackboard-based approach is used. this allows the decomposition of the problem into subtasks which are implemented as separate knowledge sourcee that can be rule-based with different inference strategies or procedural. the framework can be compared with a real-time operating system and has similar real-time primitives. the system has been implemented on a vax tl/ and used with good experiences. key words real-time expert systems, feedback control classífic¿tíon systcrm and./or indcx tcrms (íf any) s up plc ment ary b iblio graphic al ínfo tmat io n issn and kcy title isbn languagc english numbc¡ oÍ pages r.ccípìcnt's notes s ccuríty classìfrcat íon the report may bc o¡de¡ed from thc department of automatíc cont¡ol or bortowcd úårough thc llnivercíty librury , box , - o lund, swcdcn, tclcx: lubbís lund. paper to be presented at ai in real-time control, the ifac vüorkshop on - september, wales. an architecture for expert systern based feedback control kart-erik Ä.rzén department of automatic control lund institute of technology box - lund, swcden abstract. it is a recognized problem that many industrial control loops are badly tuned or ru¡l in manual modc. two cxpcrt systcm approachcs havc bccn suggcstccl ftrr tlris problem. fuzzy, rule-based control replace control algorithms by linguistic rules rv¡ic[ model the operators m¿nual control strategy. knowledge-based control extends the rangc of conventional controllers by encoding general control knowledge and heuristics concerning tuning and adaptation in a supervisory expert system. an architecture for knorvledge- based control is described where two concurrent processes are used for the knowledge-basecl system and the numerical algorithms. a modular, blackboard-based approach is used. tiris allows the decomposition of the problem into subtasks which are implemented as separate knowledge sources that can be rule-based with different inference strategies or procedural. the framework can be compared with a real-time operating system and has similar real- time primitives. the system has beer- implemented on a vax i / and used rvith goocl experiences. keyuords: real-time expert systems, feedback control . introduction there is currently a significant interest in expert sys- tem techniques in the process control community, applications of many different types have been pro- posed, implemented and a few also fielded. this pa- per considers the use ofexpert system, or knowledge- based system, techniques in the closed control loop. it is a recognized problem that many industrial con- trol loops are badly tuned or run in rnanual ¡node. this decreases the quality of the end product ancl thus increases cost. the manual control task also adds to thc already high cognitivc burclc¡r that pro- cess operators are exposed to in modern control sys- tems. the reasons for the poor control arc many. onc could be that the control loop is badly tuned from the beginning. another could be that the operating conditions have changed since the initialization of the controller. this could, c.g. bc duc to operation at dillerent operating points or time-varying dynamics. the conventional solution to the problern of poorly tuned control loops is to use adaptive controllcrs. adaptive controllers, e.g., (, .ström and wittenmarrk, ), are currently beginning to be used in i¡rdustrial practice. there are, however, problems. even though an explicit self-tuning regulator periodically updates the coefñcie¡rts of a process model thcre still a¡e rrÌany parameters that must be set explicitly. examples are model orders and time scales. such information can be diflìcult to provide and process operators typicall¡' lack the intuitive understanding that they have with conventional pid controllers- two expert system approaches have been suggested for the described problem. both i¡volve using the expert system as a part ol the feedback loop. in the u'ell-known fuzzy or rule-based approach, e.g., (tong, ), the attempt is to model the manuai control strategy of the process operator. it is exprcssed as qualitative, iinguistic rules lor horv to choose the con- trol signal in different situations, the rules replace conventional control algorithms. the intended appli- cations are control ofcomplcx proccsscs such as, c.g., cement kilns, for which either appropriate models do not exist or are inadequate. the sccond approach, from now on rcfcrrc

), a, b is the con- stant of precision control l ( < a < ; b > ). from fitness function f in eq. ( ), we know that the fitness of classifier in xcs evaluates precision of classifier in the same ac- tion set [a], and has an invert function relationship with pre- dicted error e. table ishare msci japan index (ewj) from / / to / / . symbol date open high low close volume ewj / / . . , , w.-c. tsai, a.-p. chen / expert systems with applications ( ) – . . knowledge integration model this xcs model focuses on the genetic algorithm (ga) in fig. . genetic algorithm is used to eliminate unsuitable classifiers in ac- table ishare ftse/xinhua china index (fxi) from / / to / / . symbol date open high low close volume fxi / / . . . . , fxi / / . . . , fxi / / . . . . , fxi / / . . . . , fxi .. . .. . .. . .. . .. . .. . ewj / / . . . , ewj / / . . . . , , ewj / / . . . . , ewj .. . .. . .. . .. . .. . .. . table sample data list. etfs region extend-traded funds name symbol inception date asia pacific ishares msci australia index ewa / / north american ishares msci canada index ewc / / european ishares msci sweden index ewd / / european ishares msci germany index ewg / / asia pacific ishares msci hong kong index ewh / / european ishares msci italy index ewi / / asia pacific ishares msci japan index ewj / / european ishares msci belgium index ewk / / european ishares msci switzerland index ewl / / asia pacific ishares msci malaysia index ewm / / european ishares msci netherlands index ewn / / european ishares msci austria index ewo / / european ishares msci spain index ewp / / european ishares msci france index ewq / / asia pacific ishares msci singapore index ews / / asia pacific ishares msci taiwan index ewt / / european ishares msci united kingdom index ewu / / north american ishares msci mexico index eww / / asia pacific ishares msci south korea index ewy / fi south american ishares msci brazil index ewz / / south american ishares msci south africa index eza / / asia pacific xinhua china index fund fxi / / north american ishares s&p index ivv / / north american ishares dow jones us industrial iyj / / table data coded. eits region symbol date open hi asia pacific ewa / / . asia pacific ewh / / . asian pac& ewj / / . asia pacific ewm / / . asia pacific ews / / . asia pacific ewt / / . asia pacific ewy / / asia pacific fxi / / . european ewd / / . european ewg / / . european ewi / / . european ewk / / . european ewl / / . european ewn / / . european ewo / / . european ewp / / . european ewq / / . european ewu / / . north american ewc / / . north american eww / / . north american ivv / / . north american iyj / / south american ewz / / . south american eza / / . tion set [a] rather than the whole population. in doing so, genetic algorithm starts when action set [a] have not been executed by ge- netic algorithm for an average time value. when genetic algorithm executes, two classifiers and crossover at a v probability might be randomly selected. also, it will mutate at probability. . experiment . . data this paper targets ishares msci country-specific etfs from the ishares web database (http://www.ishares.com). the data in- clude daily opening price, close price, maximum price, minimum price and trade volume over the period from jan to june , resulting in monthly observations shown in tables – . as previously mentioned, the reason for choosing ishares msci country-specific etfs is to achieve international diversification. table indicates the basic information of these etfs including the region, symbol, name, and inception date. the inception date for most of the etfs is march , and the latest inception date, for ishares msci-xinhua china (fxi), falls on october . thus, our sample period covers all etfs historical data. all of these etfs belong to barclays global investors group, known as ishares. we use ishares msci country funds measured by the msci individual country index. these etfs include eight ishares from asia pacific countries (australia, hong kong, japan, malaysia, singapore, south korea, japan, and china), ishares from european countries (austria, belgium, france, germany, italy, netherlands, spain, sweden, switzerland, and uk), four ishares gh low close volume coded . . . , – . . . , . . . , , . . . , . . . , . . . , , . . . , . . . , . . . , . . . , . . . , . . . , . . . , . . . , . . . , ––– . . . , . . . , . . . , . . . , . . . , – . . . , . . . , . . . , , . . . , http://www.ishares.com w.-c. tsai, a.-p. chen / expert systems with applications ( ) – form north american countries (s&p , dow jones, canada and mexico) and two ishares from countries in southern hemisphere (brazil and south africa). . . data coded and portfolio optimizer in the experiment, we codify the daily information of the cho- sen countries’ etf into a five-code string. the daily information in- cludes opening price (‘‘open”), maximum price (‘‘high”), minimum price (‘‘low”), closing price (‘‘close”), and trade volume (‘‘volume”). table daily global asset allocation portfolio. extend-traded fund name ticker date ishares dow jones us industrial iyj ishares goldman sachs technology index igm o ishares msci australia index ewa ishares msci austria index ewo ishares msci belgium index ewk ishares msci brazil (free) index ewz ishares msci canada index ewc ishares msci e afe index fund efa ishares msci emu index ezu ishares msci france index ewq ishares msci germany index ewg ishares msci horn kong index ewh ishares msci italy index ewi ishares msci japan index ewj ishares msci malaysia (free) index ewm ishares msci mexico (free) index eww ishares msci netherlands index ewn ishares msci singapore (free) index ews ishares msci south africa index eza ishares msci south korea index ewy ishares msci spain index ewp ishares msci sweden index ewd ishares msci switzerland index ewl ishares msci taiwan index ewt ishares msci united kingdom index ewu ishares russell index iwb ishares s&p index ivv xinhua china index fund fxi nasdaq trust shares qqq table traditional sharpe ratio asset allocation. etfs region extend-traded funds name symbol asia pacific ishares msci australia index ewa asia pacific ishares msci hong kong index ewh asia pacific ishares msci japan hides ewj asia pacific ishares msci malaysia index ewm asia pacific ishares msci singapore index ews asia pacific ishares msci taiwan index ewt asia pacific ishares msci south korea index ewy asia pacific ishares ftse/xinhua china index fxi european ishares msci sweden index ewd european ishares msci germany index ewg european ishares msci italy index ewi european ishares msci belgium index ewk european ishares msci switzerland index ewl european ishares msci netherlands index ewn european ishares msci austria index ewo european ishares msci spain index ewp european ishares msci france index ewq european ishares msci united kingdom index ewu north american ishares msci canada index ewc north american ishares msci mexico index eww north american ishares s&p index ivv north american ishares dow jones us industrial iyj south american ishares msci brazil index ewz south american ishares msci south africa index eza us t-bill the daily information for the chosen etf on / / and its coded strings are shown in table . it is worth noting that the coded string is configured to represent the differences in the daily infor- mation, and we further determine whether the differences fall into predetermined ranges or not before assigning the state code ( , , or –) to them. it is also worth noting that the ranges could be open- ended. the difference in daily information could be the differences of different day’s averaged price moving and averaged trade vol- ume (brock, lakonishok, & lebaron, ). in order to simplify the experiment, the experiment uses for positive and for coded as-is (%) to-be (%) / / / / / / – / / –– / / / / / / / / / / – / / / / / / / / / / / / / / – / / / / / , / / / / / / / / / / / / / / – / / / / / / annual return (%) annual std. deviation (%) sharpe ratio (%) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . � . . � . . . . . . . . . . . fig. . accumulated portfolio. w.-c. tsai, a.-p. chen / expert systems with applications ( ) – negative in terms of the difference in the daily information. and we further use to associate with ‘‘buy” and to associate with ‘‘sell.” in other words, the system will adjust the weight in the global ishares. the daily rules discovery is shown in table , the column of ‘‘as-is” is indicative of information at day t � while the column of ‘‘to-be” is reserved for day t (today). . . traditionally sharpe ratio the traditional portfolio model used the ‘‘sharpe ratio” to eval- uate the optimal asset allocation. hence we compare the monthly sharpe ratio global asset allocation with the xcs model global asset allocation in table . this sharpe ratio measures are used to test the etf’s performance (sharpe, ) . . international global asset allocation in this paper, we implement the integration knowledge model with xcs expert system to study the global markets that include the us, china, taiwan, japan, south africa, and malaysia. in fig. , the ishares track the indices of international capital markets before any global asset allocation strategies tracking those indices could be implemented for our international global asset allocations. the global allocation element of etfs contributes to the global risk diversification and generates sufficient gains in a transparent and low cost manner that is not easily achievable by global index funds. . . experiment result the results of the experiments are summarized in fig. . fig. shows the profit accumulation result. the average of the cumu- lated profit is better than the traditional sharpe ratio, and the high- est profit is about , units after days, which is about . % per day. therefore, it is a good performance when the index ended up lower at the end of the -day period than the start thereof. our xcs model is about . % better than the asset alloca- tion strategy on the basis of sharpe ratio. fig. . international markets analysis from xcs exports system. . conclusion as we know, the country-specific etfs offer the benefits of international portfolio diversification at a lower cost with a lower tracking error in a more tax-efficient way than passive open or closed-end county funds. this paper focuses on the soft computing algorithm, xcs, and compares with the traditional asset allocation model, according to sharpe ratio. the statistical shows that dy- namic artificial intelligence model is better than the non-efficiently monthly sharpe ratio model. additionally, using a limited numbers of factors from the real international market this paper has shown some promise using extended classifier trading mechanism in country-specific etfs. the xcs experts system consists of wilson’s xcs technique, which provides a good online learning system for our model. in the fast changing security market, genetic algorithm, rule base, neural network etc. do not satisfy our needs. rather, xcs’s online learning is generally perceived as a more suitable op- tion. xcs can give trader or investor a real-time advice to make rel- atively more accurate trading decisions in the international markets. in the future, although the experiment has shown good results, the model proposed by the current paper may still have some rooms to improve by having the inputted factors changed. especially, this work does not include the commodities etf. in addition, this study includes no short etf either. hence, the study could be further developed after having the above-mentioned inputted factors considered. the next step would be to verify xcs in different products such as commodities etfs, and actively man- aged etf. references allen, f., & karjalainen, r. ( ). using genetic algorithms to find technical trading rules. journal of financial economics, , – . amin, g. s., & kat, h. m. ( ). hedge fund performance – : do the ‘‘money machines” really add value? journal of financial, , – . andrea, g.-s. ( ). holland classifier systems. acm sigapl apl quote quad. proceedings of the international conference on applied programming languages, ( ). w.-c. tsai, a.-p. chen / expert systems with applications ( ) – bhargava, a. ( ). stochastic specification and the international gdp series. econometrics journal, royal economic society, ( ), . brock, w., lakonishok, j., & lebaron, b. ( ). simple technical trading rules and the stochastic properties of stock returns. the journal of financial, xlvii( ), – . butz, m. v., & wilson, w. ( ). an algorithmic description of xcs. technical report , illinois genetic algorithms laboratory. cumby, r., & glen, j. ( ). evaluating the performance of international mutual funds. the journal of finance, ( ), – . enu, c. s., kolodny, r., & resnick, b. g. ( ). u.s. based international mutual funds: a performance evaluation. journal of portfolio management, , – . fuhr, d. ( ). exchange- traded funds: a primer. journal of asset management, , – . holland, j. h. ( ). adaptation in natural and artificial systems. mit press. holland, j. h. ( ). escaping brittleness: the possibilities of general purpose learning algorithms applied to parallel rule-based systems. machine learning – an artificial intelligence approach, , – . holland, j. h., holyoak, k. j., nisbett, r. e., & thagard, p. r. ( ). induction: processes of inference, learning, and discovery. mit press. karpoff, j. m. ( ). the relationship between price changes and trading volume. journal of financial and quantitative analysis, , – . kashima, h. ( ). risk-sensitive learning via minimization of empirical conditional value-at-risk. ieice transactions on information and sciences, e - d(december), – . kovacs, t. ( ). a learning classifier systems bibliography. school of computer science, university of birmingham. lanzi, p.-l., stolzmann, w., & wilson, s. w. (eds.). ( ). learning classifier systems: from foundations to applications. lecture notes in artificial intelligence, lnai (vol. ). springer-verlag. leigh, w., modani, n., purvis, r., & robert, t. ( ). stock market trading rule discovery using technical charting heuristics. expert system with application, , – . miffre, j. ( ). country-specific etfs: an efficient approach to global asset allocation. journal of asset management, ( ). redman, a. l., gullett, n. s., & manakyan, h. ( ). the performance of global and international mutual funds. journal of financial and strategic decisions, , – . sakai, h., & masuyama, s. ( ). cause information extraction from financial articles concerning business performance. ieice transactions on information and systems, e -d(april), – . sharpe, w. f. ( ). asset allocation: management style and performance measurement. journal of portfolio management. shukla, r., & singh, s. ( ). a performance evaluation of global equity mutual funds: evidence from to . global finance journal, ( ), – . autumn–winter. trippi, r. r., & desieno, d. ( ). trading equity index futures with a neural network. journal of portfolio management, – . wachter, j. a., & warusawitharana, m. ( ). predictable returns and asset allocation: should a skeptical investor time the market? journal of econometrics, , – . yuan, y., & zhuang, h. ( ). a genetic algorithm for generating fuzzy classification rules. fuzzy sets and systems, , – . strategy of global asset allocation using extended classifier system introduction literature review international etfs global asset allocation sharpe ratio risk free rate xcs knowledge integration system architecture transaction data-encoding model knowledge extraction model knowledge integration model experiment data data coded and portfolio optimizer traditionally sharpe ratio international global asset allocation experiment result conclusion references sjnl - -do .tex uva-dare is a service provided by the library of the university of amsterdam (https://dare.uva.nl) uva-dare (digital academic repository) a grid-based hiv expert system sloot, p.m.a.; boukhanovsky, a.v.; keulen, w.; tirado ramos, a.; boucher, c.a.b. doi . /s - - - publication date published in journal of clinical monitoring and computing link to publication citation for published version (apa): sloot, p. m. a., boukhanovsky, a. v., keulen, w., tirado ramos, a., & boucher, c. a. b. ( ). a grid-based hiv expert system. journal of clinical monitoring and computing, ( - ), - . https://doi.org/ . /s - - - general rights it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), other than for strictly personal, individual use, unless the work is under an open content license (like creative commons). disclaimer/complaints regulations if you believe that digital publication of certain material infringes any of your rights or (privacy) interests, please let the library know, stating your reasons. in case of a legitimate complaint, the library will make the material inaccessible and/or remove it from the website. please ask the library: https://uba.uva.nl/en/contact, or a letter to: library of the university of amsterdam, secretariat, singel , wp amsterdam, the netherlands. you will be contacted as soon as possible. download date: apr https://doi.org/ . /s - - - https://dare.uva.nl/personal/pure/en/publications/a-gridbased-hiv-expert-system(f b de- fec- b - - fe b eb).html https://doi.org/ . /s - - - un co rr ec te d pr oo f techbooks journal: jocm ms code: ch pips no: do disk - - : pages: journal of clinical monitoring and computing ( ) xxx: – c© springer a grid-based hiv expert system peter m.a. sloot, alexander v. boukhanovsky, wilco keulen, alfredo tirado-ramos, and charles a. boucher from the section computational science, university of amsterdam, kruislaan , sj amsterdam, the netherlands, institute for high performance computing and information sys- tems, bering st, , st. petersburg, russia, virology educa- tion, utrecht, the netherlands, university medical center, university of utrecht, ga utrecht, the netherlands. received—, and in revised form—. accepted for publication—. based on “a grid-based hiv expert system”, by p.m.a. sloot, a.v. boukhanovsky, w. keulen, and c.a. boucher, which appeared in the ieee/acm international symposium on cluster computing and the grid, cardiff, uk, may - , . c© ieee. address correspondence to peter m.a. sloot, section computa- tional science, university of amsterdam, kruislaan , sj amsterdam, the netherlands e-mail: sloot@science.uva.nl sloot p ma, boukhanovsky av, keulen w, tirado-ramos a, boucher ca. a grid-based hiv expert system. j clin monit ; xxx: – abstract. objectives. this paper addresses grid-based in- tegration and access of distributed data from infectious dis- ease patient databases, literature on in-vitro and in-vivo phar- maceutical data, mutation databases, clinical trials, simulations and medical expert knowledge. methods. multivariate analyses combined with rule-based fuzzy logic are applied to the inte- grated data to provide ranking of patient-specific drugs. in addi- tion, cellular automata-based simulations are used to predict the drug behaviour over time. access to and integration of data is done through existing internet servers and emerging grid-based frameworks like globus. data presentation is done by standalone pc based software, web-access and pda roaming wap access. the experiments were carried out on the das, a dutch grid testbed. results. the output of the problem-solving environ- ment (pse) consists of a prediction of the drug sensitivity of the virus, generated by comparing the viral genotype to a relational database which contains a large number of phenotype-genotype pairs. conclusions. artificial intelligence and grid technology is effectively used to abstract knowledge from the data and pro- vide the physicians with adaptive interactive advice on treatment applied to drug resistant hiv. an important aspect of our research is to use a variety of statistical and numerical methods to iden- tify relationships between hiv genetic sequences and antiviral resistance to investigate consistency of results. key words. grid, hiv, pse, expert system, artificial intelligence, bio-statistics. . introduction . . motivation forty two million people worldwide have been infected with hiv and million have died, over the last years. figure shows the pan-epidemic extent of hiv infections. effective antiretroviral therapy has lead to sustained hiv viral suppression and immunological recovery in patients who have been infected with the virus. the incidence of aids has declined in the western world with the intro- duction of effective antiretroviral therapy, though questions on “when to start treatment? what to start with? how to monitor patients?” remain heavily debated. adherence to antiretroviral treatment remains the cornerstone of effec- tive treatment, and failure to adhere is the strongest pre- dictor of virological failure. long-term therapy can lead to metabolic complications. other treatment options are now available, with the recent introduction to clinical practice of fusion inhibitors, second-generation non-nucleoside re- verse transcriptase inhibitors, and nucleotide reverse tran- scriptase inhibitors. the sheer complexity of the disease, author's proofs un co rr ec te d pr oo f journal of clinical monitoring and computing vol xxx no xxx fig. . worldwide spread of hiv infections, history and near future per- spective. the distribution of the data, the required automatic updates to the knowledgebase and the efficient use and integration of advanced statistical and numerical techniques necessary to assist the physician motivated us to explore the novel possibilities supported by grid technology. in this position paper we describe ongoing research in our laboratories (utrecht, st. petersburg and amsterdam) addressing the development of a grid based medical deci- sion support system. the goal of the research is to investi- gate novel computational methods and techniques that sup- port the development of a user friendly integrated support system for physicians. we use emerging grid-technology to combine data discovery, data mining, statistical analyses, numerical simulation and data presentation [ ]. the paper is organized as follows. chapter describes the background of hiv research and a prototypical rule- based approach to data analyses. in chapter we give an overview of the two computational techniques we study to understand the temporal variability of hiv populations through stochastical modeling and the evolution of hiv infection and the onset of aids through cellular automata (ca) modeling. chapter describes a first approach to advanced data presentation through roaming devices such as personal digital assistants (pda’s). . . background . . . clinical aspects of hiv the clinical management of patients infected with human immunodeficiency virus (hiv) is based on studies on the pathogenesis of the disease and the results of trials evaluat- ing the effects of anti-hivdrugs. retrospective analysis of large cohorts has identified laboratory markers for disease progression, such as the amount of virus (hiv-rna) and the number of t helper cells (cd + cells) in blood. in ad- dition the results of prospective drug trials have generated data on effectiveness of individual drugs and drug combi- nations and the effect of drug resistant viruses on therapy outcome. currently clinicians are limited in the practical use of this information because in most cases they are only provided with statistical relationships between individual parameters and disease or therapy outcome. large data sets have not been analyzed and made available in such a way that it allows a clinician to use the available data in more clinical settings. the availability of large databases and the development of innovative data mining approaches create the opportunity to develop systems which allow the prac- ticing clinician to determine the risk profile for disease development, or the change or success for a given regimen for his individual patients. such a system will determine the rate of success for different drug regimens by taking into account the effect and interaction of all relevant laboratory and clinical parameters and by comparing the results for similar patients available in the database. currently there are fifteen drugs licensed for treatment of individuals infected with hiv. these drugs belong to two classes, one inhibiting the viral enzyme reverse transcrip- tase and another inhibiting the viral protease. these drugs are used in combination with therapy to maximally inhibit viral replication and decrease hiv-rna to below levels of detection levels (currently defined as below copies per ml) in blood. treatment with drug combinations is suc- cessful in inhibiting viral replication to undetectable levels in only % of the cases. in the remaining % of cases viruses can be detected with a reduced sensitivity to one or more drugs from the patients’ regimen. the molecular base for resistance has been, and still is, focus of extensive research. over amino acid positions in the viral enzyme reverse transcriptase (rt) and positions in the protease enzyme can undergo changes when exposed to selective drug pressure in vitro or in vivo. for some drugs, at cer- tain positions, a change towards a specific new amino acid is seen. at other positions several alternative amino acids may appear and cause (variable) levels of resistance to one or more drugs. in theory, therefore, an infinite number of combinations of amino acid changes could appear and cause resistance in vivo. preliminary clinical observations however show that specific amino acid changes at a limited number of positions and a limited number of combina- tions prevail. in addition to changing drug sensitivity some amino acid changes may also influence the replication po- tential of hiv. amino acids selected initially during a failing regimen cause resistance to the drugs the patient is taking, but at the same time may decrease the capacity of the virus to replicate. changes appearing later do not function to further increase resistance but merely function to restore author's proofs un co rr ec te d pr oo f sloot et al.: grid-based hiv expert system the capacity of the virus to replicate (“viral fitness”). sev- eral clinical studies have been performed recently to evalu- ate the clinical benefit of resistance-guided therapy. these studies show that a better virological response is obtained in patients who are failing their therapy, when their new regimen is chosen on the basis of their resistant profile. in three out of the four studies from last year the results showed that if new regimens were selected on the basis of the mutations (viral resistance genotype) the results were better as compared to standard care approaches. currently, the basis for clinical interpretation of the viral genotype is based on data sets relating mutations to changes in drug sen- sitivity, and/or data sets directly relating mutations present in the virus to clinical responses to specific regimens. ini- tially, experts compared the observed mutations to lists of published sequences taken from the literature, and based on this comparison would select a regimen. . . . prototype support system recently, first generation bioinformatics software pro- grams have been developed to support clinicians. examples of such systems are the virtual phenotype developed by virco nv, and a first generation decision support system (retrogram tm) developed by virology networks bv in collaboration with parts of our research team. the out- put of these programs consists of a prediction of the drug sensitivity of the virus generated by comparing the viral genotype to a relational database containing a large num- ber of phenotype-genotype pairs. the retrogram decision software interprets the genotype of a patient by using rules developed by experts on the basis of the literature, taking into account the relationship of the genotype and phe- notype. in addition, it is based on (limited) available data from clinical studies and on the relationship between the presence of genotype directly to clinical outcome. it is im- portant to realise however that these systems focus on bio- logical relationships and are limited to the role of resistance. the next step will be to use clinical databases and inves- tigate the relationship between the viral resistance profile (mutational profile and/or phenotypic data) and therapy outcome measures such as amount of virus (hiv-rna) and cd + cells. a summary of the flow of data is shown in figure . . . . data collection large high quality clinical and patient databases are used to explore the relationships described above and to de- velop a first prototype matching system. the athena co- hort is a large dutch observational clinical cohort study fig. . from molecule to man: hierarchical data flow model for infectious diseases. aiming at the surveillance of antiretroviral treatment sup- ported by the government. the cohort consists of patients from whom data are centrally collected through a decentralized data entry system. within the cohort pa- tients are studied intensively, whose phenotypic and geno- typic data, drug levels and cd + and hiv-rna patterns are collected. phenotype, genotype, viral fitness and drug levels as cd + and hiv-rna patterns will be collected from two large international trials (sponsored by roche pharmaceuticals), evaluating the effect of a new fusion in- hibitor drug (t ), and representing patients. the third database will be from the international multi-center great study, sponsored by virology networks bv. within this study the value of the retrogram decision support program is evaluated and similar parameters as described above will be collected. within this study patients will be enrolled. the viradapt study showed that the virological response was better in the patient group in which genotype and rule- based interpretation was used as compared to the standard of care arm [ ]. on the basis of these results, a more elabo- rate decision support software system (retrogram version . ) was built in collaboration with virology networks b.v. this system ranks the efficacy of the antiretroviral drugs within each class. the ranking is based on expert interpretation of two types of data. the software system estimates the drug sensitivity for the fifteen drugs by in- terpreting the genotype of a patient by using mutational algorithms. these mutational algorithms are developed by a group of experts on the basis of the scientific literature, taking into account the published data relating genotype to phenotype. in addition, the ranking is based on data from clinical studies on the relationship between the presence of particular mutations and clinical or virological outcome. the athena cohort is a large dutch observational clini- cal cohort study aiming at the surveillance of antiretroviral author's proofs un co rr ec te d pr oo f journal of clinical monitoring and computing vol xxx no xxx treatment supported by the dutch government. the co- hort consists of patients from whom clinical, viro- logical, immunological and data on drug side effects are centrally collected through a decentralised data entry sys- tem. within this cohort patients are studied intensively, phenotypic and genotypic data, drug levels and cd + and hiv-rna patterns are collected. from two large interna- tional trials (sponsored by roche pharmaceuticals) eval- uating the effect of a new fusion inhibitor drug (t ), representing patients from whom also phenotype, genotype, viral fitness, drug levels as cd + and hiv-rna patterns will be collected. the third database will be from the international multi-center great study sponsored by virology networks bv, within this study the value of the retrogram decision support program is evaluated and sim- ilar parameters a described above will be collected, within this study patients will be enrolled. another dataset will come from the italian musa study, in this trial data will be collected from patients followed over a year. entry point to the trial is failing a fist or second regimen, subse- quently patients will be genotyped and a new regimen will be selected on the basis of retrogram . or the virtual phenotype from virco (belgium). throughout the duration of the project we will collect additional datasets. these datasets may serve to further re- fine our models and first version software and may also be use to perform validation studies. . . . data analysis the primary goal of the data analysis is to identify pat- terns of mutations (or naturally occurring polymorphisms) associated with resistance to antiviral drugs and to predict the degree of in-vitro or in-vivo sensitivity to available drugs from an hiv genetic sequence. the statistical challenges in doing such analyses arise from the high dimensional- ity of these data. a variety of approaches have been de- veloped to handle this type of data, including clustering, recursive partitioning, and neural informatics. neural in- formatics is used for synthesis of heuristic models received by methods of knowledge engineering, and results of the formal multivariate statistical analysis in uniform systems. clustering methods have been used to group sequences that are “near” each other according to some measure of genetic distance [ ]. once clusters have been identified, recursive partitioning can be used to determine the im- portant predictors of drug resistance, as measured by in- vitro assays or by patient response to antiviral drugs. prin- ciple component analyses can help to identify what are the most important sources of variability in the hiv genome. an important aspect of our research is to use a variety of methods to identify relationships between hiv genetic se- quences and antiviral resistance to validate the consistency of results. the molecular sequences of the viral enzymes reverse transcriptase and protease are the micro parameters in the model. in theory an infinite number of combinations of mutations could appear and cause (variable) changes in viral drug sensitivity and viral replication capacity (see also ta- ble ). clinical datasets however show that specific amino acid changes at a limited numbers of positions in a lim- ited number of combinations prevail. hiv-rna and cd are the primary parameters determining disease outcome. hiv-rna, the amount of hiv-rna genomic copies per ml plasma, has been validated as being highly predictive of clinical outcome. hiv-rna and cd + cell numbers are now the standard endpoint in clinical trials for approval of new antiretroviral drugs. a patient’s hiv-rna may range between a few hundred to millions of rna copies per ml plasma. the cd + cell numbers in peripheral blood range typically between zero and thousand. whereas the predictive clinical value of both parameters has been deter- mined initially in untreated individuals, they have also been shown to be of predictive value also for patients under an- tiretroviral therapy. recently observations have been pub- lished indicating that in some patients under highly active antiretroviral therapy (haart) a disconnect may occur between the response in hiv-rna and in cd counts. typically, in these patients a rise in hiv-rna as conse- quence of incomplete inhibition of viral replication under therapy is not paralleled by a continuous decrease in cd counts. this disconnect has been explained by a decrease table . parameters for the data analyses. here the hierarchical ap- proach shown in figure is extended to detail the content of the parameters micro parameter protease mutations reverse transcriptas mutations primary parameter hiv-rna cd drug resistance macro parameter meta parameter: virological viral fitness meta parameter: clinical weight opportunistic infections and tumors survival intervention parameter drug dosage bio-availability of drug/drug level author's proofs un co rr ec te d pr oo f sloot et al.: grid-based hiv expert system in the viral replicative capacity (‘viral fitness’) which leads to a decrease in capacity to lower cd counts. the patient’s weight and secondary opportunistic infec- tions and/or malignancies are parameters that determine disease outcome and survival time. currently there are fif- teen drugs licensed for treatment of individuals infected with hiv: more than ten inhibitors have been developed which inhibit the reverse transcriptase process. these in- hibitors can be classified in two sub-categories that dif- fer in the way they inhibit the rt-enzyme, nucleoside (analogue) rt-inhibitors (nrti) and the non-nucleoside rt-inhibitors (nnrti). these compounds inhibit the protease enzyme, which acts much later on in the hiv replication cycle than reverse transcriptase. the protease is responsible for cleaving a long poly- protein into smaller functional proteins. the overall ex- posure to antiretroviral drugs has been shown to be an important factor for the degree of success for a given ther- apy. the overall exposure can be captured by parameters as dosage and bio-availability which will codetermine the drug level within an individual patient. given the relation- ships between exposure and antiviral efficacy, variability in drug levels (which may be due to differences in patient adherence to their regimens) will contribute to virologi- cal and immunological outcome. individuals with relatively low exposure are more likely to experience virological fail- ure than those with a high exposure. . methods and materials . . modeling the dynamics and temporal variability of hiv- populations in addition to rule based and parameter based decision sup- port we developed statistical models and cellular automata based models to study the dynamics of the hiv popula- tions. these numerical models run on grid-resources. the output is integrated with the medical support system and accessible to the end-user. in this paragraph we briefly outline the two computational methods. details are be- yond the scope of this paper; we refer to the references provided. . . . a cellular automata model to study the evolution of hiv infection and the onset of aids a cellular automata model to study the evolution of hiv infection and the onset of aids is developed. the model takes into account the global features of the immune re- sponse to any pathogen, the fast mutation rate of the hiv, and a fair amount of spatial localization, which may occur in the lymph nodes. the dynamics of the cellular automata requires high throughput computing, which is provided by the resource management of the grid. in this section, we employ non-uniform cellular automata (ca’s) to simulate drug treatment of hiv infection, in which each compu- tational domain may contain different ca rules, in con- trast to normal uniform ca models. ordinary (or par- tial) differential equation models are insufficient to de- scribe the two extreme time scales involved in hiv in- fection (days and decades), as well as the implicit spatial heterogeneity. zorzenon dos santos et al. [ ] reported a cellular automata approach to simulate three-phase pat- terns of human immunodeficiency virus (hiv) infection consisting of primary response, clinical latency and onset of acquired immunodeficiency syndrome. we developed a non-uniform ca model to study the dynamics of drug therapy of hiv infection, which simulates four-phases (acute, chronic, drug treatment responds and onset of aids). our results indicate that both simulations (with and without treatments) evolve to the same steady state. three different drug therapies (mono-therapy, combined drug therapy and haart) can also be simulated in our model. our model for prediction of the temporal behaviour of the immune system to drug therapy qualitatively corresponds to clinical data. pseudo code a: hi model (adapted from zorzenon dos santos r. m., phys. rev. let. ). h = healthy cell, a and a are infected cells at different time steps. assume: {h, a (t), a (t+ τ ), d}; time-step = week; simulation of lymph-node; moore neighbourhood and square lattices used rule : (a) if it has at least one infected-a neighbor, it becomes infected-a (b) if it has no infected-a neighbor but does have at least r ( < r < ) infected-a neighbors, it becomes infected-a (c) otherwise it stays healthy rule : an infected-a cell becomes infected-a after τ time steps rule : infected-a cells become dead cells rule : (a) dead cells can be replaced by healthy cells with probability prepl in the next step. (b) each new healthy cell introduced may be replaced by an infected-a with probability p infec author's proofs un co rr ec te d pr oo f journal of clinical monitoring and computing vol xxx no xxx this ca (pseudo-code a) mimics in a simple way the dynamical properties of a hiv infection; next we intro- duce drug therapy into the model by modelling a response function presp and changing only rule . pseudo code b: advanced hi model, taking into account drug therapy effects. rule : (a) if there is one a neighbor after the starting of drug therapy, n( ≤ n ≤ ) neighbor healthy cells become infected-a in the next time steps with probability presp. otherwise, all of eight neighbors become infected-a . n represents effectiveness of drugs. n = : no replication; n = : less effective for the drug. presp (t − ts ) represents certain response function of drug effects over the time steps (t). the ts is the starting of treatment. the main success of the presented ca model is the ad- equate modeling of the four-phases of hiv infection with different time scales into one model. moreover, we could also integrate all of the three different therapy procedures. the simulations show a qualitative correspondence to clin- ical data. during the phase of drug therapy response, tem- poral fluctuations for n > were observed, this is due to the relative simple form of the response distribution func- tion (pdis)applied to the drug effectiveness parameter n at each time-step. the simulation results indicate that, in contrast to ode/pde, our model supports a more flexible approach to mimic different therapies through the use of mapping the parameter space of pdis to clinical data. there- fore there is ample room to incorporate biologically more relevant response functions into the model. the data inte- gration required for the ca, the parametric computation and the data presentation are supported by the grid. . . . multivariate stochastic modeling the modeling of human immunodeficiency virus (hiv- ) genotype datasets has a goal to identify patterns of mutations (or naturally occurring polymorphisms) as- sociated with resistance to antiviral drugs and to predict the degree of in-vitro or in-vivo sensitivity to available drugs from an hiv- genetic sequence. the statistical challenges in doing such analyses arise from the high dimensionality of these data. direct application of the well-known genetic approaches [ ] to analysis of hiv- genotype results in a lot of problems. principal difference is in the fact that, in hiv dna analysis, the main scope of interests is the so-called relevant mutations – a set of mutations, associated with the drug resistance. these mutations might exist in different positions over the amino-acid chains. moreover, the sheer complexity of the disease and data require the development of the reliable statistical technique for its analysis and mod- eling. a multivariate stochastic model for describing the dynamics of complex non-numerical ensembles, such as observed in the (hiv) genome, has been developed in [ ]. this model was based on principle component analyses for numerated variables. generally speaking, the interpretation of numerated variables in terms of relevant mutations is not clear. below we develop this model directly for the ensem- ble of relevant mutations in the rt and protease parts of the hiv- genome. each element of the ensemble is pre- sented as the cortege �k = {ξ j }n kj = , k = , m with the variable dimension n k -the total number of the mutations in the gene. each value ξk is a literal index and corresponds the position and new value of the amino acid (e.g., v, i, etc.). it allows to associate each mutation with the cat- egorical random variable i ∈ . . . k , where k is the total number of possible mutations. each sub sample of genomes with a fixed number of mutations n = const may be con- sidered as the realizations of a categorical random vector. the representation above is based on the proximity to the “wild-type” virus and takes into account only the relevant mutations in a genome. it allows for significant compression of the dna representation and simplifies the interpretation of the results. principle of the modeling approach. the joint variability of dif- ferent mutations in the hiv- genomes is a complicated phenomenon. the dimension of the probabilistic charac- teristics is high, and its analytical investigations and inter- pretation are hard. hence, for the studying of hiv- pop- ulations we use a computational statistical approach that allows to numerically generate an ensemble with the same probabilistic properties by means of a monte-carlo pro- cedure. this is a well-known powerful method to study complex system variability. the idea of the stochastic modeling is shown in the figure . it is based on the evolutionary hypothesis, consid- ering the group with n + mutations as subgroup of group with n mutations in a previous step. for each gene the tran- sit from n to n + mutation groups is driven by a stochastic operator d(n+ ), which defines the mutations on the n + step, when the mutations on the previous n steps are known. the initial step of the stochastic procedure begins from the whole ensemble of wild-type viruses. the number of the genomes that has been mutated at each step of the stochas- tic procedure is in accordance with mn = ρn m, where ρn are the probabilities of the occurrence of genotypes with n mutations in a total population of m genes. author's proofs un co rr ec te d pr oo f sloot et al.: grid-based hiv expert system fig. . temporal behaviour of the cd count, with modeled brownian movement for lymphocytes [ ]. fig. . as in figure , with additionally modeled mono therapy in week [ ]. fig. . principle of the modeling. the stochastic operator d may be considered as a “black box”. it is formalized in terms of the conditional probabil- ities of the occurrence of mutation ξi , if the mutation ξ j arise in the previous step of the generation. for genotypes with mutations only the values di j are the conditional probabilities of the pairs. in this case the matrix {di j } is the transition markov probability matrix, containing the conditional probabilities for simple markov chains with the number of these states corresponding to quantity of the relevant mutations. in more complicate cases, where n > , the probability matrix {di j } consists of the con- ditional probabilities to meet mutation ξ j in certain gene, when the mutation ξi is present. this approach allows us to reduce the complicated sta- tistical description of the dataset to a rather simple model, using only three probabilistic distributions as the initial pa- rameters of the model: distribution of number n of the mutations ρn ; • distribution p ( )ξ for the relevant mutations in the group n = ; • transient probability matrix d. all these parameters might be identified on the sample datasets of the hiv- population. identification of the model. for the identification of parameters of the model, a large database of hiv-infected patients, col- lected over several years in usa, is used [ ]. these databases contain genotypes of patients examined from au- gust , to may , . we observed different mutations in the rt genome, including mixed muta- tions, and different mutations in the protease genome, including mixed mutations. distribution ρn of number of mutations. the practice of hiv treatment however, has shown that the variability of the number of mutations n is high, due to the complexity of the drug combinations that has been applied. the sample estimate of distribution ρn of the number of mutations in protease is shown in the figure . it is seen, that the distri- butions have a clear first peak (n = ), and a shelf (or second peak), corresponding to n = ÷ . therefore we expect that there are two groups of genomes in the database, cor- responding to the low and high number of mutations. the possible interpretation of the discovered bi-modal distri- bution is that we have two groups of patients. one group is the “new” patients who had one or two treatments, thus their genotype contains relative small numbers of muta- tions. the second group is the “old” patients, which have a long treatment history, or new patients, infected through treated hiv- patients [ ]. distributions of the relevant mutations pξ . distribution ρn al- lows describe the variability of the groups of the “new” and “old” patients, only. for a more detailed study of the virus mutations driving by the certain drugs combinations, the probabilities of occurrence of the relevant mutations ξ should be considered. they are estimated by the sample author's proofs un co rr ec te d pr oo f journal of clinical monitoring and computing vol xxx no xxx fig. . statistical description for distribution of mutations in protease. frequencies: pξ = {number of genes with mutation ξ } m . ( ) here m is the total number of genomes in the dataset. equation ( ) describes the marginal impact of each muta- tion in the total population, without any information about number and occurrences of other mutations. the prob- abilities of the most significant relevant mutations ξk (in decreasing order of its probability) are shown in figure . the marginal estimates of pξ over the total dataset show only general impacts of the mutations. for a detailed analysis of its behavior we also consider the occurrences p (n) ξ of mutations in the groups of genotypes with exactly n mutations. these values were computed also by means of equation ( ), where m def= mn = ρn m – the number of genes with n mutations in a database. the sample estimates of these occurrences are also shown in the figure . it is clearly seen that the inputs of some mutations are rather dif- ferent for different n, both for the protease and rt parts of the genome. e.g., for rt, for n = , the mutations v and n have the main input. the distribution p ( ) ξ is the limit distribution from the procedure shown in figure . from figure we also observe that the total sum ∑ k pξk > %, excluding case n = . this demonstrates that the analysis of the marginal mutations is not enough for general statistical description of all dna ensemble vari- ability, because some positions of dna may be statistically dependent [ ], especially in relation to viral fitness. hence, the joint characteristics of its variability must be taking into account. transient probability matrix d.the conditional probability of the occurrence of mutation ξi , if the mutation ξ j arises from the previous steps of the generation, is estimated by: di j = {number of genotypes with mutations ξi and ξ j simult&aneously} {number of genotypes with mutation ξi } . ( ) the dimensionality of the related matrix, obtained from equation ( ), may be rather high. in order to decrease the dimensionality we consider the algebraic technique of orthogonal expansion, applied to transient probability ma- trices [ ]. d = �� / �. ( ) where � are the eigenvectors of matrix ddt , and �-of matrix dt d. it allows considering the coefficients a k = √ λk as the principal components (pc) [ ], and represents the probability ( ) as a series: di j = ∑ k √ λk φi k ψ j k . ( ) the values λk shows the part of the probability, explained by k-th pc. the sum of the first k-th coefficients λk may be interpreted as a measure of convergence of the series ( ). in table the values of the first λk for the rt and protease parts of the hiv- genome are shown. these data were obtained for the total database. it can be seen that the series ( ) converges rather fast in both cases: e.g. for the rt part only the first term of the series explain more % of conditional probability (the first five terms explain %). let us consider the normalized bases φ̃i k = λ . k φi k , ψ̃ j k = λ . k ψ j k . it allows to present the terms in equa- tion ( ) as the p i jk = φ̃i k ψ̃ j k and interpreted these values as the independent factor loadings, driving the changes of the conditional probability di j over all the mutations ξi , ξ j in the database. for example, in the figure the estimates table . normalized (%) values of the expansion coefficients λk in equation ( ) # of pc part of the genome rt . . . . . . . protease . . . . . . . author's proofs un co rr ec te d pr oo f sloot et al.: grid-based hiv expert system fig. . orthogonal basic functions of expansion ( ) for transient probability matrix. of the first basic functions are shown for rt and protease parts of the genotype (the input of multiplication of func- tions are in the table ). it is clearly seen, that the first term p i j = φ̃i ψ̃ j reflects the total occurrence of the mu- tations in a genotype (see figure ): for the mutations with the maximal occurrences the input to conditional proba- bilities of its pairs is also high. model validation. the simulation model is based on the ρn , p ( ) ξ , d distributions of the mutations only. no infor- mation of more complicate mechanisms (distributions of pairs, triples, etc.) has been used for this identification. the main goal of the verification is the possibility to reproduce these features of the ensemble through the de- pendencies formalizing the matrix d. we compared the total occurrences of all mutations in genotypes, estimated on the initial and simulated samples, see also figure (solid line). it is seen, that the results of the simulation and sample are rather close. the error of the simulation increases proportionally to absolute value of the occurrences. nevertheless, for some cases the error of the simulation is larger then the boundary of the confidence interval. this systematic error may be explained by possible variations in matrix d for groups of the “old” and “new” patients. application to forecast of hiv- evolution in time. the evolu- tion of total world populations of hiv- and the associ- ated changing of the related drug resistance levels should be taken into account. the stochastic models, used to de- scribe the hiv- genotype ensemble in terms of parame- ters and shown in the figure , can be used for the analysis of its temporal variability during the observation period (viii. –v. ). the temporal variability of the data may be considered in terms of the samples of the seasons ( -months periods). the volumes of seasonal samples are from till genotypes; that is enough for obtain- ing the stable estimations. only the hypothesis of linear trends is considered: ξ (t ) = a t + b + δ(t ), where a is the most interesting parameter—value of the trend, b is the shift parameter, and δ is the white noise. in the table the integral parameters of trends of the various parameters of the hiv- population (mean value of the parameter, value of the trend, determination coefficient r and the sample value of f-criterion) are shown. trends of single mutations occurrence pξ . the database allowed us to investigate trends in codon frequency in the period of till . results for protease and rt are shown in table . the majority of the mutations in the genotype have a negative trend, only i in protease has significant positive trend. trends of bi-modal distribution for number of mutations in geno- types ρn . for the decreasing of the data dimensionality and the statistical discrimination of two groups in the dataset we consider the model of the mixture of two bernoulli distributions: ρn = p g ckm q k ( − q )m −k + ( − p g )ckm q k ( − q )m −k ( ) where p g is an input of the first group of mutations (and p g is an input of the second group, m , m -are maximal numbers of mutations in groups and q , q -are probabil- ities to find each one (arbitrary) mutation in the groups. the use of bernoulli distribution logic (based on the rep- etition of the independent events) is more close to the description of the mutation process, then the poisson dis- tribution, generally applying to description of rare events. temporal variability of the parameters ( p , q , q , m , m )t of the ρn approximation by equation ( ) are shown in table . in both cases only the parameter p g (weight of the left part for group of m mutations) has a clear author's proofs un co rr ec te d pr oo f journal of clinical monitoring and computing vol xxx no xxx table . trend analysis of the parameters of the hiv- genotype population (f is compared with fisher’s test f( , , %) = . ) occurrence of mutations, % pg , %, coefficients √ λk , equation ( ) parameter i m i v equation ( ) k = k = k = protease part mean . . . . . . . a ( /month) . − . − . . . . . . r . . . . . . . . f . . . . . . . . rt part l y n n k = k = k = mean . . . . . . . a ( /month) − . − . − . − . . . . . r . . . . . . . . f . . . . . . . . significant positive trend. for protease value p g increased from % in summer, to % in summer (with average increment a = . % per month). taking into account trends for separate mutations we observed a “degradation” of genotypes: the number of patients with simple genotypes (small number of mutations) is growing but a number of patients with big count of mutations is decreased. trends of transient probabilities d. the analysis of the trends of parameters for distribution ( ) shows that the input of the first group of mutations with low number n is increased. hence, it may be a consequence of the temporal variations of the interdependencies between different mutations, gov- erned by the developing of the drug therapy. for the anal- ysis of these hypothesis, let us consider the trends for the matrix d, equation ( ). taking into account the expan- sions ( , ), we may reduce the complicate problem for joint trend analysis for components di j to the procedure of trend analysis for independent time series – components of expansions ( ). from the table it can be seen, that all the components have a clear positive trends. taking into account the shape of first bases functions, see figure , it is clear, that generally the joint probabilities di j of the mu- tations is increased also; moreover, the power of increasing corresponds to the total occurrences of the mutation in the ensemble. the discrimination of the groups of “old” and “new” patients in terms of bi-modal distribution ( ) allow to fore- cast the growth of the total number of hiv-infected people in time: n(t ) = nnew patients (t ) + nold patients (εt ), ε � . ( ) here ε – is the slow time parameter, which shows the rapid increasing of the new patients group in comparison with the old patients. the part of “new” patients of the sample is p g (old patients−( − p g )) from ( ). hence, the growth curve is: n(t ) = nold patients ( ) [ + p g (t ) − p g (t ) ] , ( ) where p g (t ) = p + a g t -is the linear trend with the pa- rameters from table , and n old patients ( ) is the initial value of “old” (treated) patients on the beginning of the forecast. in figure the “crucial” forecast of the hiv- popula- tion growth are shown. it is based on the fact that altogether fig. . qualitative forecast of hiv- population grows. – mean value ( ), – % confidence interval. author's proofs un co rr ec te d pr oo f sloot et al.: grid-based hiv expert system million people worldwide have been infected with hiv at the beginning of xxi century, and million have died over the last years. moreover, not taken into account is the arising of new drugs and different prophylactic and social preventive activities for restriction of hiv- infec- tion. really, this result is qualitative only; for quantita- tive conclusions the more sophisticated research should be done. . results . . data presentation: roaming pda access . . . user scenario retrogramtm (www.retrogram.com) is a unique hiv- genotype expert based interpretation software program, which weighs the effect of specified genotype changes on clinical drug activity. it accepts a list of substitutions to the protease and reverse transcriptase genes with respect to the nl - reference strain. this is accomplished by running a “simulation”, which applies some hundred rules relat- ing substitutions on the hiv genome to knowledge of effects on drug response. the latter comes from over hun- dreds of references from the clinical literature. the rules are checked against the reported substitutions, and each drug is evaluated for its suitability. in a later stage we added web- access where a web interface is used to submit the input and take out the output. we want to make the simulations wireless-accessible. developing a wireless internet version from scratch will not be cost-efficient and causes maintain- ability problems. for example, the rules mentioned above are often changed and these changes have to be reflected in both versions. furthermore, for privacy and security rea- sons the developer is not granted access to the source code of the “simulation”. thus, it is much more convenient to have wireless access to the web-based interface. in this case the “simulation” take places in a unique server and privacy and security are guaranteed. a typical user scenario is de- scribed below and the associated graphical representation of the retrogram web access is given in figure . after the user has successfully logged in, the patient detail page is displayed (figure ). the form, taking place in this page is used to enter the personal data of the patient. two fields are required in the form, patient id and data of sample. according to the information taken from the laboratory the user enters the laboratory test results (i.e. protease or rt substitutions) for the patient in the laboratory informa- tion page. next a script invoked on the server does the following: fig. . web-based retrogram use case sequence. script : server validation script validate inputs: validate protease or rt substitutions if they conform to certain rules. a single substitution should be represented by an integer (for position in the gene) and a letter (for the amino acid). the position in the gene is in the rage from to for protease position and from to for rt position. the amino acid code is one of the following codes: a c d e f g h i k l m n p q r s t u v w y. submit the inputs to the “simulation” program and take back the drugs ranking result. show the drugs ranking result in the ‘hiv therapy decision support’ screen: after applying certain rules on the laboratory test result return to the final drugs ranking or drug’s level of suitability indication as follows: a (green): this drug can be used b (yellow): consider use if no class a drug available c (amber): consider use if no class a or b drug available d (red): consider use if no class a, b or c drug available u (grey): unranked, insufficient data available in the ‘hiv therapy decision support’ screen, clicking on any drug name in the ranking lists will display a list of avail- able references from the scientific literature supporting the particular ranking for that drug. in the ‘hiv therapy deci- sion support’ screen, clicking on the ‘interpret substitution’ button will show classification of the patient’s substitutions into relevant, natural or additional. author's proofs un co rr ec te d pr oo f journal of clinical monitoring and computing vol xxx no xxx fig. . web retrogram: user enters patient substitutions (left), drug ranking results (right). . . . roaming, wireless access in the designing phase of wireless versions of the application the constraints of the mobile devices should be considered. at the same time we have tried to maintain the same level of usability and readability as in the original web version. this is accomplished by maintaining the same structure as that in the web but with some modifications. for example, the patient detail form has many fields and putting them in one screen would cause problems in the usability of the program (it’s supposed that the mobile device has a resolution comparable to a normal pda, i.e., something around × pixels). thus we use three screens for patient detail data. the patient detail web page has required fields. we put them in the first screen after the ‘login’ screen. in this way, if the user is not interested in entering optional data, she can directly go to the laboratory information. proxy method implementation. a proxy method is imple- mented for accessing the web-based software from mobile devices. the proxy server takes places between the remote server (the retrogram server) and the mobile device. a mininavigator script developed in the proxy is responsible for the following: • take the patient data from the mobile user (i.e. patient detail, laboratory information) • create an http communication with the remote server, • submit data to the remote server. these data are basically the input for the retrogram ‘simulation’. • take the result from the remote server (html code generated from retrogram.asp script), • parse html code and retrieve only relevant informa- tion (i.e. drug ranking, error messages, drug references etc.). it uses this relevant information to build wireless pages (i.e. wml page in case of wap or web-clipping page). • send the wireless pages to the mobile device. the proxy is implemented using php: hypertext pre- processor as a server-site scripting language [ – ] running on the apache web server [ ]. two versions are developed using the proxy method: wap version and web clipping. if a user wants to enter the ‘patient details’ fields, he has to move from one screen to the other and come back again. the fields already filled in the previous screens should not be lost. thus maintaining the client’s state is necessary. in the wap case we simply use cookies but in web clipping cookies are supported only in palm os . version or higher. for this reason the “hidden field” method is used this is another method used for maintaining state in the internet. the following figures are the user interfaces that have been captured. they track the user’s path through the running of the application, as shown in figures (a) and (b), where the user enters the patient’s details and accesses ranking results. j me implementation. the same user interface is applied in the j me implementation. there are two main differences between the j me implementation and the proxy one: . j me enables the device to communicate directly to the retrogram server without an intermediate proxy . in j me the client’s interface is contained within the device. in the proxy method, every time the interface should be changed, the proxy is responsible for gener- ating a new page. the following illustrates the necessary steps one should take in order to fetch an html page generated from a author's proofs un co rr ec te d pr oo f sloot et al.: grid-based hiv expert system fig. . (a) user corrects the input and submit again (left), drug ranking re- sults (right). (b) users clicks to the drug ‘indinavir’ (left), references supporting this ranking (right). script in the remote host. specifically this is an example illustrating how the user can login to a script in the ret- rogram server and extract the cookie from the header re- sponse: . open an http connection . open an input stream . make an http post request . extract the cookie from the header response . close the connection in the j me implementation of retrogram the entire client’s interface takes places in the device. the connec- tion to the server is established in the following cases: user login, with connection with the server is necessary in order to validate the user and/or password. the user submits the fig. . j me method; user enters patient’s substitutions (left), drug ranking results (right). username and password, and the application judges them for their correctness by scanning the html response from the retrogram server. the user submits the patient’s lab- oratory information data. the application should connect to the server in order to submit the data, take the result (html format) and extract the drugs ranking. next the user looks for the references that suggest a certain drug ranking. the database with all the references exists in the retrogram server, therefore the connection is necessary. the application submits to a retrogram script the cookie and the name of the drug. the drug references are given back from the server in html format. the application should clean up the html tags and show the references as plain text. finally the user looks for classification of the patient’s substitutions. this classification is part of the ret- rogram ‘simulation’ and thus the connection to the server is still necessary. in figure we illustrate the process of taking the drugs ranking using the j me method. currently we have the j me version in use for different users to study the usability and extendibility. more details on the implementation can be found in reference [ ]. . . virtual laboratory infrastructure . . . a virtual organization for retrogram-centered workflow grid technology is a major cornerstone of today’s com- putational science and engineering, with its basic unit of grid organization called the virtual organization (vo). a vo is a set of grid entities, such as individuals, appli- cations, services or resources, which are related to each other by some level of trust. in the most basic example, author's proofs un co rr ec te d pr oo f journal of clinical monitoring and computing vol xxx no xxx fig. . a retrogram-centered workflow. service providers would only allow access to the mem- bers of the same vo. we are currently building a dis- tributed grid-based overall decision support infrastructure to support the retrogram-centered workflow shown in figure . this vo will offer a grid virtual laboratory that will assist users in the interpretation the genotype of a patient by using rules developed by experts on the basis of the lit- erature, taking into account the relationship between the genotype and phenotype. the workflow is based on highly distributed available data from clinical studies and on the relationship between the presence of genotype and the clin- ical outcome. in order to cover the fast temporal and spatial scales required to infer information from a molecular (ge- nomic) level up to patient medical data multi-scale methods are applied, where simulation, statistical analysis and data mining are combined and used to enhance the rule-based decision. in this scenario, information sources are widely distributed, and the data processing requirements are highly variable, both in the type of resources required and the pro- cessing demands. experiment design, integration of infor- mation from various sources, as well as transparent schedul- ing and execution of experiments will be supported by this support system based on distributed grid middleware. the das testbed (netherlands) will initially provide the addi- tional computational power for our compute intensive jobs. we will reuse grid middleware from successful european projects such as crossgrid (www.crossgrid.org) and vl-e (www.vl-e.nl) to provide basic grid services of data man- agement, resource management, and information services on top of globus. for transparent use of this infrastructure we will build a presentation layer that will provide a user- friendly interface to both medical doctors and scientists. . discussion . . conclusions and future work in this paper we discussed an integrative approach to bio- medicine at large and to infectious diseases in particular. we showed how in the understanding of processes ‘from molecule to man’ grid technology can play a crucial role. in order to cover the fast time and spatial scales required to infer information from a molecular (genomic) level up to author's proofs un co rr ec te d pr oo f sloot et al.: grid-based hiv expert system patient medical data, we need to apply multi-scale meth- ods where simulation, statistical analysis, data-mining is combined in an efficient way. moreover the required in- tegrative approach asks for distributed data collection (e.g. hiv mutation databases, patient data, literature reports etc.) and a virtual organization (physicians, hospital administra- tion, computational resources etc.). also the access to and use of large-scale computation (both high performance as well as distributed) is essential since many of the compu- tations involved require near real-time response and are to complex to run on a personal computer or pda. fi- nally data presentation is crucial in order to lower the barrier of actual usage by the physicians, here the grid technology (server-client approach) can play an important role. although many of the aspects discussed in this paper have proven to work in concept, the complete integration of the systems and the evaluation of day-to-day use is still under development [ ]. in addition each of the underlying methods (rule-based, statistical and ca based models) remain topics of further studies. we will set up a use-base with the system described running under various european grid testbeds. the first testbed we will use is the so-called das , and eventually the crossgrid testbed, which supports specific features for interactive computa- tion, an essential ingredient for a medical decision support system. the authors gratefully acknowledge fan chen and ferdinand alimadhi for assistance in implementing the ca models and the roaming pda access. the dutch virtual laboratory on e- science project supported parts of the research presented here: http://www.vl-e.nl. glossary grid: distributed architecture for solving computational problems by making use of the resources from the mem- bers of a virtual organization, treating them as a virtual cluster. ca: cellular automata, a discrete model studied in com- putational theory and mathematics, which consists of regular grid of cells, each in one of a finite number of states. decision support system: computer-based system that helps in the process of decision-making. web interface: user interfaces for information available via the web. proxy: computer service which allows clients to make in- direct network connections to other services. http: hyper text transfer protocol, a request/response protocol for transferring information on the web. html: hyper text markup language, a markup language designed for the creation of web pages. wml: wireless markup language, a markup language used in mobile phones. j me: java platform micro edition, a collection of java interfaces for embedded consumer appliances such as cellular phones. das : distributed asci super computer , a wide-area distributed computer connecting dutch universities. references . zhao z, belleman rg, van albada gd, sloot pma. ag-ive: an agent-based solution to constructing interactive simula- tion systems, in series lecture notes in computer science, april ; : – . . durant j, clevenbergh p, halfon p, delguidice p, porsin s, simonet p, montagne n, dohin e, schapiro jm, boucher c, dellamonica p. improving hiv therapy with drug resis- tance genotyping: the viradapt study. lancet ; : – . . sevin ad, degruttola, nijhuis m, schapiro jm, foulkes as, para mf, boucher cab. methods for investigation of the re- lationship between drug-susceptibility phenotype and human immunodeficiency virus type genotype with applications to aids clinical trials groupw . the journal of infectious diseases ; : – . . the genotype database is obtained from a large service testing laboratory from the us. it contains the resistance profiles of the protease and reverse transcriptase genes of the hiv- virus obtained from plasma samples of hiv- infected patients. no clinical background information on medication or drug history is available. . mathematical methods for dna sequences. in. waterman ms, eds. crc press inc., boca raton, florida, . . kiryukhin i, saskov k, boukhanovsky av, keulen w, boucher, ca, sloot pma. stochastic modeling of temporal variability of hiv- population. in: sloot pma, abrahamson d, bogdanov av, dongarra jj, zomaya ay, gorbachev ye, eds. compu- tational science – iccs , melbourne, australia and st. petersburg, russia, proceedings part i, in series lecture notes in computer science, vol. , pp. – . springer verlag, june . isbn - - - . . zorzenon dos santos rm, coutinho s. dynamics of hiv infec- tion: a cellular automata approach. phys rev lett ; ( ): – – . . sloot pma, chen f, boucher ca. cellular automata model of drug therapy for hiv infection. in: bandini s, chopard b, tomassini m, eds. th international conference on cellu- lar automata for research and industry, acri , geneva, switzerland, october – , . proceedings, in series lecture notes in computer science, vol. , pp. – . october . . php: hypertext preprocessor: http://www.php.net. author's proofs un co rr ec te d pr oo f journal of clinical monitoring and computing vol xxx no xxx . the resource for php developers: http://www.phpbuilder.com . zend technologies – php tools for the development, pro- tection and scalability of php applications – php for linux, unix and apache, encoder, accelerator studio, debugger: http://www.zend.com. . the apache software foundation: http://www.apache.org. . alimadhi f. mobile internet: wireless access to web- based interfaces of legacy simulations, msc thesis, uni- versity of amsterdam, the netherlands, september : http://www.science.uva.nl/research/pscs/papers/master.html. . cross-grid: grid technology of interactive distributed com- putation: http://www.eu-crossgrid.org/. . little sj, holte s, routy jp, daar es, markowitz m, collier ac, koup ra, mellors jw, connick e, conway b, kilby m, wang l, whitcomb jm, hellmann ns, richman dd. antiretroviral- drug resistance among patients recently infected with hiv. n engl j med ; ; ( ): – . . karlin s. a first course in stochastic processes. academic press. ny-london, . . sloot pma, boucher ca, kiryukhin i, saskov k, boukhanovsky av. a grid-based problem-solving envi- ronment for biomedicine. in: nørager s, ed. proceedings of the first european healthgrid conference, january, th- th, , pp. – . commission of the european commu- nities, information society directorate-general, brussels, belgium, . author's proofs un co rr ec te d pr oo f query q . au: pls. provide dates. author's proofs wp-p m- .ebi.ac.uk params is empty sys_ exception wp-p m- .ebi.ac.uk no params is empty exception params is empty / / - : : if (typeof jquery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/ . . /js/jig.min.js"][/script]'.replace(/\[/g,string.fromcharcode( )).replace(/\]/g,string.fromcharcode( ))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} page not available reason: the web page address (url) that you used may be incorrect. message id: (wp-p m- .ebi.ac.uk) time: / / : : if you need further help, please send an email to pmc. include the information from the box above in your message. otherwise, click on one of the following links to continue using pmc: search the complete pmc archive. browse the contents of a specific journal in pmc. find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/med/ an algebraic approach to rule based expert systems available on line at www.rac.es/racsam revista de la real academia de ciencias computational sciences exactas, fisicas y naturales. survey serie a: matematicas madrid (españa / spain) racsam ( ), , – . doi: . /racsam. . an algebraic approach to rule based expert systems eugenio roanes-lozano, luis m. laita, antonio hernando and eugenio roanes-macı́as abstract this article presents a survey of the authors’ research on knowledge extraction and verification of rule based expert systems (rbes) using algebraic inference engines and based on gröbner bases theory. a shell, including a graphic user interface and inference engines for different logics (both classic and modal multi-valued) as well as in different computer algebra systems, is also presented here. the shell distinguishes three levels: at the lower level, we provide the computer algebra system code of the algebraic inference engines; at the intermediate level, the rbes developer has to detail the rules and integrity constraints of a certain rbes; and, finally, at the upper level, the final user deals with a simple gui, where he can perform knowledge extraction or verify the rbes, after choosing the logic and inputing a consistent set of facts. we believe that this shell can be really useful for teaching and quick rbes design. una aproximación algebraica a los sistemas expertos basados en reglas resumen. este artı́culo presenta una panorámica de la lı́nea de investigación de los autores en ex- tracción de conocimiento y verificación de sistemas expertos basados en reglas (rbes) usando motores de inferencia algebraicos y basada en la teorı́a de bases de gröbner. se presenta también una shell, que incluye una interfaz gráfica de usuario y motores de inferencia para distintas lógicas (tanto clásicas como modales multivaluadas) y en distintos sistemas de cómputo algebraico. la shell distingue tres niveles: en el más bajo proporcionamos el código del motor de inferencia para el sistema de cómputo algebraico elegido; en el intermedio el desarrollador del rbes tiene que detallar las reglas y las restricciones de integridad de un cierto rbes; y, finalmente, en el nivel superior, el usuario final trata con una sencilla interfaz gráfica de usuario, en la que puede llevar a cabo extracción de conocimiento o verificar el rbes, después de elegir la lógica y de introducir un conjunto consistente de hechos. creemos que esta shell puede ser realmente útil para la enseñanza y para el rápido diseño de rbes. submitted by david rı́os insua. received: august , . accepted: october , keywords: rule based expert systems, logic and symbolic computing, groebner bases mathematics subject classifications: b , p , b c© real academia de ciencias, españa http://www.rac.es/racsam e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as introduction this article presents a survey of the authors’ research on knowledge extraction and verification of rule based expert systems (rbes) using algebraic inference engines. these inference engines are based on the use of gröbner bases. the leader of the research team is luis laita, and the first works on the topic are dated in the early ’s, although he had already been working in the field using other techniques. the present paper is self-contained, in such way that introductory well-known sections , , may be skipped by an acquainted reader, while the other sections convey a view of new results and approaches by the authors. the article begins with a brief introduction to lattices, lattice orders, boolean algebras and boolean rings (section ). it is followed by the construction of the polynomial model for boolean logic due to these authors (sections and ). then a brief introduction to modal multi-valued logics can be found (section ). section introduces the main result of this theory, relating to be a tautological consequence with a polynomial ideal membership. a brief introduction to rbes is included afterwards (section ). section details the adaptation of the algebraic model for logic to rbes. in section it is shown how two levels of inconsistency can be distinguished in the multi-valued case (including an interpretation in algebraic geometry). section introduces how (in rbes whose underlying logic is classic boolean logic), given a set of facts, it is possible to find out which new fact can make a certain goal be inferred, just computing a gröbner basis. implementations in the computer algebra systems (cas) cocoa and maple are included in section . finally, a gui for keeping the cas hidden from the final user of the rbes is presented in section . in view that quite a number of topics are treated, many proofs are omitted for the sake of brevity (in fact, some of the proofs are very long, for instance, the one connecting the concept of being tautological consequence in multi-valued logic with an ideal membership in a polynomial residue class ring). anyway, references where the reader can find details on the topics treated can be found along the paper. an introduction to boolean algebras and boolean rings an elementary introduction to boolean algebras may be found in [ ]. the proofs omitted in this section can be found in [ , ] (these references also extend both sections and ). for a deeper study of boolean algebras see, for instance, [ , , , , ]. . lattices and lattice orders definition a set l where two binary operations t and u are defined is said to be a lattice if and only if both operations are commutative and associative and the absorption laws holds, i.e., if and only if for all elements a, b, c ∈ l: at b = bta au b = bua at (bt c) = (at b) t c au (bu c) = (au b) u c at (bua) = a au (bta) = a proposition if (l,t,u) is a lattice, both operations verify idempotency, i.e., for every element a ∈ l: ata = a aua = a surprisingly, idempotency, although redundant, is often required as a fourth condition for a structure to be a lattice. definition a non-strict partial order defined over the set l is said to be a lattice order if and only if every pair of elements of l has a unique infimum and a unique supremum. an algebraic approach to rule based expert systems proposition from the two lattice operations, t and u , a lattice order v can be defined: for all elements a, b ∈ l: a v b ⇔ at b = b (⇔ au b = a) conversely, from a lattice order v , the two operations of a lattice can be defined: for all elements a, b ∈ l: at b = supv(a,b) au b = infv(a,b) example let us consider the lattice of “subsets of e”: (p(e),∪,∩). from (p(e),∪,∩), (p(e),⊆) is obtained. conversely, from (p(e),⊆), (p(e),∪,∩) is obtained. . boolean algebras definition a lattice (l,t,u) is said to be distributive if and only if both operations are distributive w.r.t. the other operation, i.e., if and only if for all elements a, b, c ∈ l: at (bu c) = (at b) u (at c) au (bt c) = (au b) t (au c) definition a lattice is said to be bounded if and only if it has a greatest element (top), g, and least element (bottom), l. definition a bounded lattice (l,t,u) is said to be complemented if and only if for every element a ∈ l, there is an element a′ ∈ l such that: ata′ = g aua′ = l the top and bottom of a lattice are traditionally denoted and . however this seems not to be an adequate notation, as the top and bottom of (l,t,u) and those of (l,u,t) are exchanged. definition a boolean algebra is a distributive and complemented lattice. example (p(e),∪,∩, ′) is a boolean algebra: • ∪ is distributive w.r.t. ∩ and viceversa. • the least and the greatest of the lattice (for ⊆) are ∅ and e, respectively. moreover, the usual complement of a set, ′ , works as expected: ∀a ∈p(e) : a∪a′ = e ; a∩a′ = ∅ let us underline that being distributive is completely independent from being complemented, as shown below. example the lattice of the linear subspaces (also denoted “vector subspaces”) of r , the sum of linear subspaces and intersection is a complemented, but not distributive, lattice: • the complement of a line is any other (different) line, and the complement of the whole plane is { }, • if r, s, t are different lines, r + (s ∩t) = (r + s) ∩ (r + t)). example the lattice (n, lcd, gcd) (where lcd and gcd stand for “least common multiple” and “greatest common divisor”, respectively) is distributive, but it is not complemented: and are the least and greatest of the lattice (respectively), but, for instance, has no complement. example the lattice of the convex subsets of the plane, the convex union ∗ ∪ (i.e., the least convex set that contains the usual union) and intersection is a lattice, but it is neither distributive nor complemented: • for instance, if a, b, c are three disjoint and aligned squares of equal size and a is the one in the middle a∩ (b ∗ ∪ c) = a, meanwhile (a∩b) ∗ ∪ (a∩c) = ∅ ∗ ∪∅ = ∅, • there is no convex subset of the plane that behaves as the complement of a line. e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as . boolean rings definition a boolean ring is a ring with unit such that the second operation is idempotent. let us include below some well-known results that will be useful when obtaining the polynomial model for boolean logic. let us denote by (l, ∆,u) a general boolean ring, by its neutral element, by ′ the symbol for the opposite and by its unit. proposition in any boolean ring (l, ∆,u) any element is its own opposite. proof. for any elements a, b ∈ l: (a ∆ b) ∆ = a ∆ b = (a ∆ b) u (a ∆ b) = (aua) ∆(au b) ∆(bua) ∆(bu b) = a ∆((au b) ∆(bua)) ∆ b = (a ∆ b) ∆((au b) ∆(bua)) but (l, ∆) is a group, so the opposite is unique, and consequently: = (au b) ∆(bua). therefore, in the particular case that a = b, from the idempotency of u, we would obtain: = a ∆ a. � proposition any boolean ring (l, ∆,u) is a commutative ring. proof. it follows from the equality = (au b) ∆(bua) in the previous proof and the facts that (l, ∆) is a group (and therefore the opposite is unique) and that any element is its own opposite. � proposition a boolean ring (l, ∆,u) can be defined from a boolean algebra (l,t,u, ′): for all elements a, b ∈ l: a b = (au b′) t (a′ u b) conversely, a boolean algebra (l,t,u, ′) can be defined from a boolean ring (l, ∆,u): for all elements a, b ∈ l: at b = (a b) (au b) and if and are the neutral elements of ∆ and u (respectively), then is the least, is the greatest, and: a′ = a ∆ example (p(e),∪,∩, ′) is a boolean algebra (least: ∅; greatest: e). if we define the symmetric difference of two elements of p(e), a and b, as: a b = (a∩b′) ∪ (a′ ∩b) we have that (p(e), ,∩) is a commutative ring with unit: all properties are a straightforward conse- quence of the same property of the boolean algebra or can be obtained from a simple symbolic manipulation (like the commutativity and associativity of ). moreover: • ∅ is the neutral element for (and the absorbent element of ∩ ), • e is the neutral element for ∩ , • ∀a ∈p(e), a a = ∅, so any element in p(e) is its own opposite. example (p(e), ,∩) is a boolean ring. if we define in this ring the following operation: atb = (a b) (a∩b) the operation ∪ is obtained. if we define a′ = a ∆ e the usual complement of a set is obtained. an algebraic approach to rule based expert systems a polynomial model for propositional logic we can obtain a polynomial model for the boolean algebra associated to (propositional) boolean logic, as suggested in figure . propositional boolean algebra oo boolean algebra isomorphism // oo �� polynomial boolean algebra oo �� propositional boolean ring oo ring isomorphism // polynomial (boolean) ring figure . isomorphisms between polynomial and propositional boolean algebras and boolean rings. in logic, we usually work in the structure in the upper left corner of the diagram of figure . meanwhile, in commutative algebra we normally work in the structure in the lower right corner of this diagram. but we can move from the propositional boolean algebra to the polynomial ring following the two alternative paths in this diagram. the advantages of working in a polynomial ring are that: • there is a powerful tool that solves the ideal membership problem: gröbner bases, • there are many implementations of buchberger’s algorithm for computing gröbner bases, i.e., per- forming effective calculations in algebra. in fact, most cas, like maple, mathematica, axiom, mu- pad, reduce, derive, cocoa, singular, risa/asir,. . . include implementations of this algorithm. as a consequence, we shall be able to implement very easily an algebraic tool to perform effective compu- tations in propositional boolean logic. moreover, it will be possible to extend this tool to perform effective computations in modal multi-valued logics (like Łukasiewicz, kleene and bochvar). finally, it will be pos- sible to use the same computational tool to perform verification and knowledge extraction in rbes based on both classic boolean and modal multi-valued logics. . building the taylor-made polynomial (boolean) ring (a, +, ·) let us build a polynomial boolean ring (a, +, · ) so that: • every element is its own opposite, i.e., for all a ∈a, a + a = ·a = , so it would be reasonable to consider a as a ring over a field with characteristic , • idempotency holds for the second operation, so it would be reasonable to consider the residue class ring over the polynomial ideal 〈p − p,q − q, . . . ,r − r〉, where p, q, . . . , r are the polynomial variables. therefore, we shall consider: a = z [p,q, . . . ,r]/〈p −p,q −q, . . . ,r −r〉 proposition the polynomial product in a is idempotent. proof. it is enough to take into account that any polynomial a ∈a can be written: a = δ + δp ·p + δq ·q + · · · + δr ·r + δpq ·p ·q + · · · + δpqr ·p ·q ·r + · · · where all the δ belong to z , and compute a . � it is well-known that an algebraic structure like (a, +, · ) is a ring (denoted residue class ring) and, therefore, we have the following: e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as corollary the ring (a, +, · ) is a boolean ring. proposition all elements in (a, +, ·) are zero divisors (in fact an analogous result holds in any boolean ring). proof. a · ( + a) = a + a = a + a = ·a = . � . the polynomial boolean algebra (a, +̃, · , +) corresponding to the po- lynomial (boolean) ring (a, +, ·) let us define a (polynomial) boolean algebra (a,t, · , ′) from the (polynomial) boolean ring (a, +̃, ·), as shown in proposition . for all a, b ∈a: at b = (a + b) + (a · b) a′ = a + the first operation will be hereinafter denoted +̃ and + will denote the addition of . is the least and is the greatest of the boolean algebra (a, +̃, · , +): for all a, b ∈a, a+̃ = a + + a · = a a · = a+̃ = a + + a · = a · = a unlike what happens in a ring, it is not usual to establish priorities for the operations in a boolean algebra (as the structure is completely “symmetrical”). nevertheless, we shall consider, hereinafter, that · is prioritary w.r.t. +̃. proposition the lattice order defined in (a, +̃, · , +) as suggested in proposition : ∀a,b ∈ a, a ≤ b ⇔ a · b = a is, precisely, “is a multiple” (a is a multiple of b ⇔∃k ∈a : a = b ·k). proof. ⇒) ∀a,b ∈a: a ≤ b ⇔ a · b = a ⇒ a is a multiple of b ⇐) ∀a,b ∈a: a is a multiple of b ⇔∃k ∈ a : a = b·k ⇒ a·b = (b·k)·b = b ·k = b·k = a ⇔ a ≤ b. � proposition for all a, b ∈a: ( ) a · b = a ( ) a+̃b = b ( ) ( + a)+̃b = ( ) a · ( + b) = are equivalent (an analogous result holds in any boolean algebra). the boolean algebra isomorphism ϕ let ∨, ∧, ¬, → denote the logic disjunction, conjunction, negation and implication, respectively. let (c,∨,∧,¬,→) be the boolean algebra of the propositions that can be constructed using a finite number of propositional variables p , q, . . . , r. let us denote tautology by and contradiction by . let us consider the boolean algebra (a, +̃, · , +, “is a multiple”), where a is the residue class ring a = z [p,q, . . . ,r]/〈p −p,q −q, . . . ,r −r〉 an algebraic approach to rule based expert systems we define: ϕ: (c,∨,∧,¬,→) −→ (a, +̃, · , +, “is a multiple”) in the following way. for propositional variables p −→ p q −→ q . . . . . . r −→ r and for any a, b ∈c a∨b −→ a+̃b ¬a −→ + a therefore, as an immediate consequence of the de morgan laws: a∧b −→ a · b moreover, as the (lattice) order can be obtained from the operations of the lattice, ϕ preserves the ordering. proposition ϕ is well defined. proof. for all propositions a, b: a ↔ b ⇒ a → b and b → a ⇒ ⇒ ϕ(a) is a multiple of ϕ(b) and ϕ(b) is a multiple of ϕ(a) ⇔ ⇔ a is a multiple of b and b is a multiple of a ⇔ a = b. � corollary ϕ is an order-preserving homomorphism, and ϕ( ) = ,ϕ( ) = . remark in order to save space, we shall sometimes write b|a (“b divides a”) instead of “a is a multiple of b”. proposition ϕ is surjective. proof. the elements of the (boolean) ring a (they are the same as those of the boolean algebra a) are a linear algebraic combination of the polinomial variables, and: ϕ(p ∧q) = p ·q ϕ((¬p ∧q) ∨ (p ∧¬q)) = p + q. � proposition ϕ is injective. proof. let us suppose that for two propositions, a and b, we have: ϕ(a) = a, ϕ(b) = b and a = b. as a = b, we have: a|b and b|a. but ϕ preserves the ordering (i.e., → and “is a multiple” do correspond), so: b|a ⇒ a → b; a|b ⇒ b → a and therefore a ↔ b. � e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as remark to be precise, we really consider c/ ↔ and a/ = (lindenbaum algebras) in order for the relations → and ≤ to be anti-symmetric. example let p and q be the propositional variables of c. then c has elements and (c,→) is the transitive closure of the diagram in figure . it corresponds in ϕ with (a, “is a multiple”), where a = z [p,q]/〈p −p,q −q〉 and “is a multiple” is the transitive closure of the diagram in figure . � � �� @ @ @r �� � pppq p ∧¬q ¬p ∧¬q p ∧ q ¬p ∧ q �� �� ���: - � � � � � ��* xxxxxxxz -xxxxxxxz -xxxxxxxz a a a au b b b b bbn � �� � � � � � � ��� (p ∨¬q) ∧ (¬p ∨ q) ¬q p q ¬p (p ∧¬q) ∨ (¬p ∧ q) �� �� ���: - �� �� ���: xxxxxxxz �� �� ���: - - h h h h h hhj a a a au b b b b b bn � �� � � � � � � � �� p ∨¬q ¬p ∨¬q p ∨ q ¬p ∨ q � � ��� �� pppq @ @ @r figure . (c,→) when there are two propositional variables. � � �� @ @ @r �� � pppq p · ( + q) ( +p)·( +q) p · q ( + p) · q �� ��� - � � � �� pppppq - pppppq - pppppq j j j ĵ a a a a a au � � ��* � � � � � � �� + q p q + p + (p + q) p + q �� ��� - �� ��� pppppq �� ��� - - q q q qqs @ @ @ @r j j j j j ĵ � � � � � � � �� � � ��* p+̃( + q) ( + p)+̃( + q) p+̃q ( + p)+̃q � � ��� �� pppq @ @ @r figure . (a, “is a multiple”) when there are two polynomial variables. . ideals of a boolean algebra and ideals of a ring definition let (r, +, ·) be a commutative ring. a subset i ⊆ r is said to be an ideal of r if and only if: ) ∀i, i′ ∈ i : i + i′ ∈ i ) ∀i ∈ i,∀a ∈r : a · i ∈ i (i.e. i is a subring of a such that the product of an element of the ring by an element of the subring always belongs to the subring). an algebraic approach to rule based expert systems definition let (r, +, ·) be a commutative ring. let s = ∅, s ⊆ r. the ideal generated by s = {p , . . . ,pm}, denoted 〈p , . . . ,pm〉, is the smallest ideal containing s. if s = {q}, the ideal generated by q, 〈q〉, is said to be a principal ideal, and results to be: 〈q〉 = {a ∈a : q|a} despite the concept of ideal is usually associated with rings, ideals are also defined in boolean algebras. in such case, the definition is different (based on the lattice order), and only ideals generated by a single element are considered. definition in the propositional boolean algebra (c,∨,∧,¬,→), the (principal) ideal generated by q is defined as eq = {x ∈c : x → q} similarly, in the polynomial boolean algebra (a, +̃, · , +, “is a multiple”), the (principal) ideal gener- ated by q ∈a is: eq = {a ∈a : a is a multiple of q} proposition as ϕ preserves the ordering, the ideals of the boolean algebra (c,∨,∧,¬,→) do corre- spond in ϕ with the ideals of the boolean algebra (a, +̃, · , +, “is a multiple”). proposition the ideals of the boolean algebra (a, +̃, · , “is a multiple”) are obviously ideals of the ring (a, +, · , “is a multiple”). theorem (a, +, ·) is a ring where all ideals are principal ones. proof. let s , s , . . . , sn ∈a. we shall prove that 〈s ,s , . . . ,sn〉 = 〈s +̃s +̃ · · ·+̃sn〉 in two steps: i) +̃ is an internal operation in the ideal (because + and · also are internal operations) and therefore: s +̃s +̃ · · ·+̃sn ∈ 〈s ,s , . . . ,sn〉⇒ 〈s +̃s +̃ · · ·+̃sn〉⊆ 〈s ,s , . . . ,sn〉 ii) that s +̃s +̃ · · ·+̃sn|si; i = , . . . , n can be easily proven. but, then: si ∈ 〈s +̃s +̃ · · ·+̃sn〉; i = , . . . , n and, consequently, the minimum ideal that contains {s ,s , . . . ,sn}, is contained in 〈s +̃s +̃ · · ·+̃sn〉, i.e.: 〈s ,s , . . . ,sn〉⊆ 〈s +̃s +̃ . . . +̃sn〉. � remark let us observe that (a, +, ·) is not a “principal ideals domain” because it is not an integrity domain (see proposition ). corollary as a consequence of proposition , proposition and theorem , the ideals of the polyno- mial boolean algebra (a, +̃, · , “is a multiple”) are precisely the ideals of the polynomial ring (a, +, ·). introductory note about multi-valued propositional logics a simple introduction to the topic can be found in [ ]. e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as . kleene’s and Łukasiewicz’s three-valued logic kleene’s and Łukasiewicz’s multi-valued logics are two of the most commonly used multi-valued logics. we shall begin introducing these three-valued logics, in which a third truth value is considered. in kleene’s logic this third valued is “undecided”. if an assertion is assigned the truth value “undecided” that means that at present, we can’t assign a truth-value to this conjecture (but the conjecture is either true or false). for instance, if we are reasoning in a medical diagnostic rbes, until we receive the result of a test, its truth value is “undecided”, but it will become either true or false in the future (when we receive the results). kleene’s a → b is equivalent to ¬a∨b. two modal unary connectives, necessary (�) and possible (♦) are usually considered in multi-valued logics. kleene’s three valued logic truth-tables are detailed in tables and (note that represents “false”, represents “undecided” and represents “true”). it would also be possible to assign numbers to truth values in a different way. for instance, in kleene’s three valued logic, it is also common to consider that represents “false”, that represents “undecided” and that represents “true”. in such case the corresponding truth tables would obviously change. ¬ � ♦ table . truth-tables of kleene’s three-valued logic unary connectives. ∧ ∨ → ↔ table . truth-tables of kleene’s three-valued logic binary connectives. in Łukasiewicz’s logic the third truth value considered is “indeterminate”. if a statement is considered to be indeterminate, it is neither true not false, but indeterminate in a metaphysical sense. all connectives have the same truth tables as in kleene’s logic, except → and ↔ (see table ). → ↔ table . truth-tables of the Łukasewicz’s three-valued logic binary connectives whose truth-tables are different from kleene’s ones. remark it is also possible to consider more than three truth values, so that the confidence in the truth- fulness of statements can be graded. moreover, there are other multivalued logics aside from kleene’s and Łukasewicz’s. an algebraic approach to rule based expert systems . tautological consequence definition a propositional formula a is a tautological consequence of the propositional formulae a , a , . . . , am, denoted {a ,a , . . . ,am} |= a , if and only if whenever the formulae a ,. . . ,am hold (i.e., take the truth value “true”), then a also holds (i.e., takes the truth value “true”). remark let us observe that: • in boolean logic, there are formulae that can only take the truth value “false”, such as p ∧¬p , • in multi-valued logic, there are also formulae that can only take the truth value “false”, such as �p ∧�¬p . • in multi-valued logic, there are formulae that, although can take other truth values different from “false”, they can never take the truth value “true”, such as p ∧¬p (observe that not only �p ∧ �¬p |= p ∧¬p , but also p ∧¬p |= �p ∧�¬p ). . the isomorphism ϕ in the multi-valued case let (c,∨,∧,¬,→) be a p-valued logic, where p is a prime number (we require p to be a prime in order for zp to be a field) and c is the set of propositions that can be constructed using the propositional variables x , x , . . . , xn. if in the residue class ring a = zp[x ,x , . . . ,xn]/〈x p −x ,x p −x , . . . ,x p m −xm〉 adequate polynomial translations of the logical connectives are defined (so that they behave as suggested by the truth-tables), the natural correspondence ϕ, defined as in the boolean case (section ), is also an isomorphism. the main theorem . a brief note about gröbner bases in the early ’s, both heisuke hironaka and bruno buchberger independently proved that, for each poly- nomial ideal, a basis completely identifying it always exists. they denoted their bases as standard bases and gröbner bases (gb) [ , ], respectively. the latter’s great advantage was that it provided a constructive method (buchberger’s algorithm). some gb are particularly important: we call them reduced gröbner bases. we say that a gröbner basis is reduced if and only if the leading coefficient of all its polynomials is and we cannot simplify any of its polynomials by adding a linear algebraic combination of the rest of the polynomials in the basis. the input to buchberger’s algorithm is a polynomial set, a term order (for instance, “total degree” or “pure lexicographical”), and a variable order (for instance, x > y > z) and its output is the ideal’s reduced gb with respect to the specified term and variable orders. the key point is that, once the term order and the variable order are fixed, such a reduced gb completely characterizes the ideal: any ideal has a unique reduced gb. as a consequence we have two key results: i) two sets of polynomials generate the same ideal if and only if their reduced gb are the same, ii) { } is the only reduced gb for the ideal that is equal to the whole ring. an elementary introduction to the use of gb in algebraic system solving can be found in [ , ]. for detailed introductions see, for instance, [ , , , , ]. there are interesting applications in other fields like transportation engineering [ , ]. e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as . connecting “to be a tautological consequence” with an ideal member- ship the main theorem is the following: theorem a logical formula, a , is a tautological consequence of the set of formulae {a ,a , . . . ,am} in a certain p-valued modal logic (where p is a prime number) if and only if the polynomial translation of ¬a in a = zp[x ,x , . . . ,xn]/〈x p − x ,x p − x , . . . ,x p m − xm〉, ϕ(¬a ), belongs to the ideal 〈ϕ(¬a ),ϕ(¬a ), . . . ,ϕ(¬am)〉 of a. the key point is that this theorem can take advantage of the two important results stated at the end of the previous section. the proof of this theorem is really long and can be found in [ , ]. some introductory notes about rbes a rbes consists of an “input”, an “output”, a “knowledge base” and an “inference engine”. a graphic user interface is frequently provided. the input of a rbes is concerned with the information related to the environment of the rbes. since the environment may change, this information is also subject to change as times goes by. this information is described by means of a finite number of different atomic propositions: such propositions are termed as input atomic propositions. input atomic propositions determine what we will call potential facts (see definition ). the output of a rbes is concerned with the information inferred by the rbes. it is described by means of a finite number of atomic propositions which we shall call output atomic propositions. the knowledge base (kb) of the rbes is concerned with the information contained in the system, which is used along with the input of the rbes to infer the output of the system. in a rbes based on propositional boolean logic, this knowledge-base will be mostly represented by a finite set of rules. remark we shall use hereinafter the following notation: z may mean: • no symbol at all or “negation” (denoted ¬), if the underlying logic is classic boolean, • no symbol at all, “negation”, “possibility” (denoted ♦), “necessity” (denoted �), or a combination of these symbols, like “necessarily-not” (denoted �¬), equivalent to ¬♦, if the underlying logic is a modal multi-valued one. . rules definition the kb consists of a certain number of logical formulae of the form: ∧ki= zx[i] −→∨ l j= zx[j] known as production rules (and usually refereed to as, simply, rules). for example, formula x[ ] ∧x[ ] ∧¬x[ ] ∧♦x[ ] ∧¬x[ ] −→ x[ ] ∨�¬x[ ] is read as: “if x[ ] and x[ ] and not x[ ] and possibly x[ ] and not x[ ] hold, then x[ ] or necessarily-not x[ ] holds”. let us underline that there is no constraint in the structure of the rules: the ocurrences of “∧” in the consequent and of “∨” in the antecedent of a rule can be avoided (if such symbols ocurred in a rule, the rule could be split into rules without these symbols). an algebraic approach to rule based expert systems . potential facts and facts. integrity constraints definition a formula a is a potential fact if and only if a ≡ zx[i], where x[i] is an input atomic proposition. definition that a given set of facts (i.e., a subset of the set of potential facts that is stated as true) is consistent means that, for each propositional variable appearing in the potential facts of the system, x[k], at most one expression of the form zx[k], is chosen. that a given set of facts is maximal means that, for each propositional variable appearing in the potential facts of the system, x[k], exactly one expression of the form zx[k], is chosen. definition an integrity constraint (ic) is a piece of knowledge added by the experts and expressing that some potential facts cannot hold at the same time. for instance an ic could be: man ∧ pregnant. the negation of the ics (sometimes denoted “nics”) have to be added to the kb. . the inference engine. rbes inconsistency the inference engine (ie) is an automated tool (in our case, a program in the cas), that verifies consistency and draws tautological consequences from the information contained in the kb. definition that a formula a be obtained by forward firing of the formulae in the union, g, of a con- sistent set of facts and the set of all production rules and the set of the negation of the integrity constraints of a rbes means that g |= a. definition a rbes is said to be inconsistent if there is a consistent set of facts verifying that con- tradiction can be obtained by forward firing of these facts and the rules and the negation of the integrity constraints in the kb. otherwise the rbes is said to be consistent. knowledge extraction in rbes through the use of an alge- braic inference engine this extension of the algebraic method to rbes was already mentioned in [ ] and detailed in [ ]. . a polynomial model for the propositional boolean algebra associated to rbes the boolean algebra associated to a rbes whose underlying logic is classic boolean logic is a structure (c∗,∨,∧,¬,→) where → is the relation obtained applying the rules of logical deduction to the implications of c and the rules, facts and the negations of the integrity constraints of the rbes (consequently, → is not the usual implication), and c∗ is the set of equivalence classes defined by this enlarged equivalence relation ↔ in c. if the rules rule , . . . , rulev , the facts fact , . . . , fact m and the integrity constraints ic , . . . , ic u of a rbes are added as true to the boolean algebra c of the previous section, the structure obtained is isomorphic to the image of a in the natural surjective homomorphism ψ : a−→a/j where j is the ideal 〈ϕ(¬rule ), . . . ,ϕ(¬rulev),ϕ(¬fact ), . . . ,ϕ(¬fact m),ϕ(¬ic ), . . . ,ϕ(¬ic u)〉 e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as . the main theorem adapted to knowledge extraction in rbes the main theorem detailed above can be adapted to treat the logic underlying a rbes: theorem a formula a can be obtained by forward firing of the formulae in the union of a consistent set of facts and the set of all production rules and the negation of the integrity constraints of a rbes, if and only if, the polynomial translation of the negation of a belongs to the ideal j of a = zp[x ,x , . . . ,xn]/〈x p − x ,x p −x , . . . ,x p n −xn〉, generated by the polynomial translations of the negations of the given potential facts, of the negations of all the production rules, and the integrity constraints of the rbes. remark observe that, denoting i = 〈xp − x ,x p − x , . . . ,x p m − xn〉, we could alternatively check that a belongs to the ideal i + j of zp[x ,x , . . . ,xn]. we usually work this way as most cas are not prepared to perform computations in a residue class ring. we have extensively used this approach in rbes design and verification. [ ] is an interesting example in medicine (regarding techniques for coronary surgery), where we could find a situation that could lead to an inconsistency and was not taken into account by the panel of experts. rbes strong and weak inconsistency. algebraic interpreta- tion an introduction to verification can be found in [ ]. this section summarizes [ ]. we shall make the following distinction in the multi-valued case: definition we shall say that there is strong inconsistency if and only if the conjunction of the facts in a certain consistent set of facts, the rules and the negations of integrity constraints of the rbes, can only take the truth value “false”. definition we shall say that there is weak inconsistency if and only if all formulae can be obtained by forward firing of the formulae in the union of a certain consistent set of facts, the set of all production rules and the set of the negations of integrity constraints of the rbes. proposition independently of the number of truth values (p) of the underlying logic: strong inconsistency ⇒ weak inconsistency. proof. if there is strong inconsistency, we have a consistent set of facts {fact , . . . , fact m} such that: fact ∧ . . .∧ fact m ∧ rule ∧ . . .∧ rulev ∧ nic ∧ . . .∧ nic u can only take the truth values false. but {fact , . . . ,fact m, rule , . . . , rulev, nic , . . . , nic u} |= fact ∧ . . .∧ fact m ∧ rule ∧ . . .∧ rulev ∧ nic ∧ . . .∧ nic u so a contradiction can be obtained by forward firing of the formulae in {fact , . . . , fact m, rule , . . . , rulev, ic , . . . , ic u}, and, therefore, we have weak inconsistency. � an algebraic approach to rule based expert systems proposition if the underlying logic is boolean: weak inconsistency ⇒ strong inconsistency. proof. if there is weak inconsistency, we have a consistent set of facts {fact , . . . , fact m} such that any formula follows of the formulae in {fact , . . . , fact m, rule , . . . , rulev, nic , . . . , nic u} by forward firing. in particular, {fact , . . . , fact m, rule , . . . , rulev, nic , . . . , nic u} |= contradiction but we are considering that the underlying logic is boolean, so: fact ∧ . . .∧ fact m ∧ rule ∧ . . .∧ rulev ∧ nic ∧ . . .∧ nic u → contradiction and therefore fact ∧ . . .∧ fact m ∧ rule ∧ . . .∧ rulev ∧ nic ∧ . . .∧ nic u can only take the truth value “false”. � remark the previous result does not hold if the underlying logic is a multi-valued one. example let us consider a small rbes whose kb consists of the rules: r :x[ ] → x[ ] r :x[ ] →¬x[ ] and whose underlying logic is kleene’s three-valued logic. the potential facts are x[ ] and x[ ]. if we draw consequences from the maximal set of potential facts {x[ ],x[ ]}, we obtain x[ ] ∧¬x[ ]. as this formula can only take the truth values “undecided” or “false”, any formula can be obtained by forward firing of it, and, consequently, by forward firing of the formulae in {x[ ],x[ ],r ,r }. therefore we have weak inconsistency. meanwhile, we cannot obtain a formula that is always “false” by forward firing of the conjunction of the facts in any consistent set of facts and r ∧r . therefore, we do not have strong inconsistency. corollary there is weak inconsistency if and only if ∈ 〈f¬(ϕ(rules)),f¬(ϕ(facts)),f¬(ϕ(nics))〉. but, what is the meaning of strong inconsistency in the polynomial model? definition the algebraic variety corresponding to the polynomial ideal i is the set of points of the respective affine space which satisfy all polynomials in the ideal i. we shall denote it by v (i). lemma for any positive prime integer p > , we have: i) there are non-zero ideals in zp[x ,x , . . . ,xm] whose algebraic variety is the whole affine space. ii) the only ideal in a = zp[x ,x , . . . ,xm]/i; i = 〈x p −x ,x p −x , . . . ,x p m −xm〉 whose variety is the whole affine space is the zero ideal. proof. i) for example x · (x − ) = x − x ∈ z [x] is not the null polynomial but its value is always in { , }. therefore 〈x + x〉 verifies condition i). the same occurs with x · (x− ) · (x− ) ∈ z [x]. its value in { , , } is always . this construction can be obviously generalized to zp[x]. e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as ii) case i: univariate polynomials. if , , . . . , p− are roots of a polynomial in a = zp[x]/〈xp −x〉, then the polynomial is either or it can be factored as x · (x− ) · · · · · (x− (p− )) · · · · therefore, if this polynomial is different from , its total degree is greater or equal than p. but in a all polynomials are simplified to polynomials of degree lesser than p, so this polynomial simplifies to . case ii: multi-variate polynomials. it can be proved by applying the previous case to the intersections of the hypersurface by hyperplanes parallel to the “vertical” axis. � theorem there is strong inconsistency if and only if v (〈ϕ(rules ∧ facts ∧ nics)〉) is the whole (respective) affine space. proof. rules ∧ facts ∧ nics can only take the truth value “false” ⇔ ⇔ polynomial ϕ(rules ∧ facts ∧ nics) can only take the value ⇔ ⇔ v (〈ϕ(rules ∧ facts ∧ nics)〉) is the whole (respective) affine space. � but we can express the previous result in terms of ideals instead of algebraic varieties, what makes it easily decidable using gröbner bases: corollary there is strong inconsistency if and only if 〈ϕ(rules ∧ facts ∧ nics)〉 is the zero ideal of a = zp[x ,x , . . . ,xm]/i. backward reasoning in rule based expert systems whose underlying logic is classic boolean logic details on the topic of this section, our most recent result, can be found in [ ]. based on the classical works of automatic discovery of geometric theorems [ , ] (a topic already advanced in [ ]), we have now adapted this idea to a specific task in rbes whose underlying logic is classic boolean logic: given a set of facts and a goal, finding a new fact such that if it also held, this goal would be inferred. the main result proven in [ ] (theorem below) shows how it is enough to compute one single reduced gb in order to detect them: theorem let a , . . . , ak, b ∈c∗ be formulae such that {a , . . . ,ak} |= b. let c be a potential fact. let g be the reduced gb of the ideal 〈ϕ(¬a ), . . . ,ϕ(¬ak),ϕ(b)〉 + i. we have that: {a , . . . ,ak,c} |= b ⇔ ϕ(c) ∈ g observe that the membership relation in the right hand side of the equivalence refers to a membership to the set of polynomials in the reduced gb, not to the ideal generated by that basis. implementation in a computer algebra system as most cas include an implementation of gb and of “normal forms” (reductions modulo an ideal), to develop an inference engine in a cas following the method detailed above (based on checking polynomial ideal memberships) is straightforward and surprisingly brief. an algebraic approach to rule based expert systems . polynomial translation of the logical connectives for instance, the polynomial expressions corresponding to the basic logical formulae in Łukasiewicz’s three-valued logic (if is assigned to “true”, to “undetermined” and to “false”) are detailed afterwards. • ¬m is translated into the polynomial: −m • ♦m is translated into the polynomial: m • �m is translated into the polynomial: m + m • m ∨n is translated into the polynomial: m n + m n + mn + mn + m + n • m ∧n is translated into the polynomial: m n + m n + mn + mn • m → n is translated into the polynomial: m n + m n + mn + mn + m + • m ↔ n is translated into the polynomial: m n + m n + mn + mn + m + n + (note that the coefficients of these polynomials belong to z ). the polynomials corresponding to the different logics can be obtained by solving an algebraic system that is obtained from the truth tables. . cocoa implementation these polynomial translations can be input almost directly in cocoa . [ , ], which provides a “normal form” command, nf(pol,i), that returns the reduction of polynomial pol modulo ideal i. we have to define the polynomial ring and ask cocoa to use it: use z/( )[x[ .. ]]; and then we can define the ideal memory.i, generated by the expressions x i −xi (denoting it this way, it is a global variable, and it can be an input to nf): memory.i:=ideal([x[k_]ˆ -x[k_] | k_ in .. ]); and introduce the polynomial expressions for the three-valued connectives of Łukasiewicz modal logic (the operators are prefix ones) : neg(m):=nf( -m,memory.i); pos(m):=nf( *mˆ ,memory.i); nec(m):=nf(mˆ + *m,memory.i); or(m,n):=nf(mˆ *nˆ +mˆ *n+m*nˆ + *m*n+m+n,memory.i); and(m,n):=nf( *mˆ *nˆ + *mˆ *n+ *m*nˆ +m*n,memory.i); imp(m,n):=nf( *mˆ *nˆ + *mˆ *n+ *m*nˆ +m*n+ *m+ ,memory.i); iff(m,n):=nf(mˆ *nˆ +mˆ *n+m*nˆ + *m*n+ *m+ *n+ ,memory.i); for instance, the cocoa implementation of the polynomial expressions of a -valued modal logic can be found in [ ]. whether a is a consequence of a set of formulae or not, can be checked with cocoa just typing: nf(neg(a[ ]),memory.i+j); (where j is the polynomial ideal generated by the polynomial expressions of the negation of the formulae in the given set). if the output of the command is , the answer is “yes”, otherwise the answer is “no”. in cocoa . these operators would be defined using define...enddefine instead of directly :=. e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as it can also be directly checked using the boolean command isin, that tests ideal memberships, by typing: neg(a[ ]) isin memory.i+j; whether the rbes is inconsistent or not can also be checked using gröbner bases by typing in cocoa: gbasis(memory.i+j); if the output is , the rbes is inconsistent; otherwise (the output is normally a large set of polynomials), that set of facts doesn’t lead to an inconsistency. using isin makes it even simpler, as it directly returns “true” or “false”, just typing: isin memory.i+j; . maple implementation let us implement the model in maple. we have developed different implementations in maple, but the one presented afterwards is the most modern and fastest [ ]. in maple (version ≥ ) it is possible to define a polynomial ring over a finite field using maple’s ore algebra. in this way, the normalform command is computed in such ring. we load the gröbner and ore algebra packages first, and then define the list of variables, the polynomial ring and the order that will be used by the gb-related commands: with(groebner): with(ore_algebra): sv:=x[ ],x[ ],x[ ]: a:=poly_algebra(sv,characteristic= ): orde:=monomialorder(a,’plex’(sv)): and the ideal i (denoted ii, as i is a reserved word in maple), using map in order to save time: fu:=v->vˆ -v: ii:=map(fu,[sv]); ii := [x[ ] - x[ ], x[ ] - x[ ], x[ ] - x[ ]] we can now define the functions that associate to the logical connectives their polynomial expressions, as done in cocoa above. neg :=(m::algebraic) -> normalform( -m,ii,orde): nec :=(m::algebraic) -> normalform(expand(mˆ + *m),ii,orde): pos :=(m::algebraic) -> normalform(expand( *mˆ ),ii,orde): ‘&and‘:=(m::algebraic,n::algebraic) -> normalform(expand( *mˆ *nˆ + *mˆ *n+ *m*nˆ +m*n), ii,orde): ‘&or‘ :=(m::algebraic,n::algebraic) -> normalform(expand(mˆ *nˆ +mˆ *n+m*nˆ + *m*n+m+n), ii,orde): ‘&imp‘ :=(m::algebraic,n::algebraic) -> normalform(expand( *mˆ *nˆ + *mˆ *n+ *m*nˆ +m*n+ *m+ ), ii,orde): ‘&iff‘ :=(m::algebraic,n::algebraic) -> normalform(expand(mˆ *nˆ +mˆ *n+m*nˆ + *m*n+ *m+ *n+ ), ii,orde): an algebraic approach to rule based expert systems description of the shell a gui was presented at flins’ conference [ ], and a complete shell is now provided. when working with the shell we could distinguish three levels. at the lower level, we provide the cas code for performing knowledge extraction and verification in different logics (i.e., we provide the algebraic inference engines). for instance, to work with the cas cocoa . in boolean logic with the propositional variables x , . . . , x , the code that contains the algebraic model of the ie should be included in the file named code cocoa.txt. it would look like: use z/( )[x[ .. ]]; memory.i:=ideal([x[k_]ˆ -x[k_] | k_ in .. ]); neg(m):=nf( +m,memory.i); or(m,n):=nf(m+n+m*n,memory.i); ... (similar to the code shown in section . ). some lines regarding i/o between the gui and the cas follow. at the intermediate level, the rbes developer details the concrete potential facts, rules and integrity constraints of a certain rbes (the code must be stored in file rbes.txt). example for instance, the production rules and ic of a tiny rbes whose underlying logic is classic boolean and their cocoa translations can be (note that in a real rbes there would be dozens of rules): r : x[ ] → x[ ] r :=imp(x[ ],x[ ]); r : x[ ] ∧¬x[ ] → x[ ] r :=imp(and(x[ ],neg(x[ ])),x[ ]); r : x[ ] ∧x[ ] → x[ ] r :=imp(and(x[ ],x[ ]),x[ ]); nic : ¬(x[ ] ∧x[ ]) nic :=neg(and(x[ ],x[ ])); at the upper level, the final user does not have to deal directly with the cas but with a simple gui. both the number of truth values of the logic (e.g., ) and the desired logic (e.g., kleene’s) are to be chosen beforehand. then consistency can be checked and knowledge extraction performed. to do so, a list formed by a proposition and a (consistent) list of facts has to be introduced. the gui calls the cas and checks firstly if they lead to inconsistency. afterwards, it checks if the given proposition follows from the given set of facts. example let us consider the tiny rbes of example . if the list of given facts is [x[ ],¬x[ ]], an inconsistency is not obtained. moreover, x[ ] ∧x[ ] follows from them (see figure ). this gui has been implemented in visual-basic and runs on computers using the most common windows-based operating system. porting it to other operating systems is under consideration now. it can call (at least) the mathematical systems: cocoa, maple, maxima, singular, risa-asir and octave. conclusions we believe that this algebraic approach is flexible and powerful (as several logics can be easily and effec- tively simulated) and that the whole shell now put together can be really useful for teaching and for quick rbes designing. this approach has the advantage over that of kapur-narendran [ ] and hsiang [ ] for boolean logic (or alonso et al. [ ] and chazarain et al. [ ] in the multi-valued case) of providing an algebraic structure that is isomorphic to the propositional algebra, instead of only performing effective com- putations in logic using gröbner bases. this is the key for going one step further and obtaining an algebraic structure isomorphic to a rbes based on one of these logics, i.e., from a = zp[x,y, . . . ,z]/〈xp −x,yp −y, . . . ,zp −z〉 e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as figure . screenshot of a gui for rbes consistency checking and knowledge extraction. to a/j where j is the ideal generated by the negation of the facts, rules and integrity constraint, so that knowledge extraction and verification of rbes can be performed. acknowledgement. this work was partially supported by the research projects tin - (government of spain) and ucm - (ucm - bsch gr. / , research group aceia, spain). we would also like to thank the anonymous referees for their most valuable comments and sugges- tions. references [ ] adams, w. w. and loustaunau, p., ( ). an introduction to gröbner bases. graduate studies in mathe- matics , ams. [ ] akritas, a. g., ( ). elements of computer algebra with applications, wiley - interscience. [ ] alonso, j. a. and briales, e., ( ). lógicas polivalentes y bases de gröbner, en m. vide ed., procs. of the v congress on natural languages and formal languages, barcelona univ. press, – . an algebraic approach to rule based expert systems [ ] buchberger, b., ( ). applications of gröbner bases in non-linear computational geometry. in j. r. rice (editor), mathematical aspects of scientific software, ima volumes in math. and its applications, , springer- verlag. [ ] buchberger, b., ( ). bruno buchberger’s phd thesis : an algorithm for finding the basis elemen- tals of the residue class ring of a zero dimensional polynomial ideal, j. symbolic comput., , - , – . doi: . /j.jsc. . . [ ] capani, a. and niesi, g., ( ). cocoa user’s manual (v. . b), depto. de matemáticas, universidad de genova. [ ] chazarain, j.; riscos, a.; alonso, j. a. and briales, e., ( ). many-valued logic and gröbner bases with applications to modal logic, j. symbolic comput., , – . [ ] chou, s.-c., ( ). mechanical geometry theorem proving. kluwer academic publishers. [ ] cox, d.; little, j. and o’shea, d., ( ). ideals,varieties, and algorithms, springer-verlag. [ ] halmos, p. r., ( ). lectures on boolean algebras, springer-verlag. [ ] halmos, p. and givant, s., ( ). logic as algebra, the mathematical asociation of america. [ ] hermes, h., ( ). la teorı́a de retı́culos y su aplicación a la lógica matemática, conf. mat. vi., csic- madrid. [ ] hsiang, j., ( ). refutational theorem proving using term-rewriting systems, artificial intelligence, , – . doi: . / - ( ) - [ ] kapur, d. and mundy, j. l., ( ). wu’s method and its application to perspective viewing. in d. kapur and j. l. mundy, eds., geometric reasoning. mit press, cambridge ma, – . [ ] kapur, d. and narendran, p., ( ). an equational approach to theorem proving in first-order predicate calculus, crd , general electric corporate research and development report (schenectady, ny, march , rev dec ). also in a. k. joshi, ed., proceedings of ijcai- . morgan kaufmann, – . [ ] kreuzer, m. and robbiano, l., ( ). computational commutative algebra. springer-verlag. [ ] laita, l. m. and de ledesma, l., ( ). knowledge-based systems verification. in kent, j. g. williams (editors), encyclopedia of computer science and technology. marcel dekker, – . [ ] laita, l. m.; de ledesma, l.; roanes-lozano, e. and roanes-macías, e., ( ). an interpretation of the propositional boolean algebra as a k-algebra. effective calculus. in j. campbell, j. calmet (editors), proceedings of aismc- , springer-verlag lncs , – . [ ] laita, l. m.; roanes-lozano, e.; de ledesma, l. and alonso, j. a., ( ). a computer approach to verification and deduction in many-valued knowledge systems, soft comput., , , – . doi: . /s [ ] laita, l. m.;roanes-lozano, e.; maojo, v.; de ledesma, l. and laita, l., ( ). an expert system for managing medical appropriateness criteria based on computer algebra techniques, comput. math. appl., , , , – . doi: . /s - ( ) - [ ] mendelson, e., ( ). theory and problems of boolean algebras and switching circuits, schaum’s, macgraw-hill. [ ] monk, d., ( ). handbook of boolean algebras, north-holland. [ ] recio, t. and vélez, m. p., ( ). automatic discovery of theorems in elementary geometry, j. automat. reason., , – . [ ] roanes-lozano, e.; hernando, a.; laita, l. m. and roanes-macías, e., ( ). a shell for rule- based expert systems development using groebner bases-based inference. in d. ruan, j. montero, j. lu, l. martı́nez, p. d’hondt, e. e. kerre (editors), computational intelligence in decision and control. proceedings of the th international flins conference, world scientific proceedings series on computer engineering and information science , – . doi: . / http://dx.doi.org/ . /j.jsc. . . http://dx.doi.org/ . / - ( ) - http://dx.doi.org/ . /s http://dx.doi.org/ . /s - ( ) - http://dx.doi.org/ . / _ e. roanes-lozano, l. m. laita, a. hernando and e. roanes-macı́as [ ] roanes-lozano, e.; hernando, a.; laita, l. m. and roanes-macías, e., a groebner bases- based approach to backward reasoning in rule based expert systems, ann. math. artif. intell., to appear. doi: . /s - - - [ ] roanes-lozano, e.; laita, l. m. and roanes-macías, e., ( ). maple v in a.i., the boolean alge- bra associated to a kbs, can nieuwsbrief, , – . [ ] roanes-lozano, e.; laita, l. m. and roanes-macías, e., ( ). a polynomial model for many- valued logics with a touch of algebraic geometry and computer algebra, math. comput. simulation, , – . [ ] roanes-lozano, e.; laita, l. m. and roanes-macías, e., ( ). cálculos efectivos en lógica proposi- cional booleana interpretada como un anillo de clases residuales (polinomial) sobre z , boletı́n de la sociedad “puig adam” de profesores de matemáticas, , – . [ ] roanes-lozano, e.; laita, l. m. and roanes-macías, e., ( ). a groebner bases based many- valued modal logic implementation in maple. in s. autexier et al. (editors), aisc / calculemus / mkm , lnai springer-verlag, – . doi: . / - - - - [ ] roanes-lozano, e.; muga, r.; laita, l. m. and roanes-macías, e., ( ). a terminal area topology-independent gb-based conflict detection system for a-smgcs, racsam rev. r acad. cienc. exactas, fı́s. nat. ser. a mat., , - , – . [ ] roanes-lozano, e.; roanes-macías, e. and laita, l. m., ( ). geometric interpretation of strong inconsistency in knowledge based systems. in v. g. ganzha, e. w. mayr, e. v. vorozhtsov (editors), computer algebra in scientific computing. proceedings of casc’ ). springer-verlag, – . [ ] roanes-lozano, e.; roanes-macías, e. and laita, l. m., ( ). railway interlocking systems and groebner bases, math. comput. simulation, , , – . doi: . /s - ( ) - [ ] roanes-lozano, e.; roanes-macías, e. and laita, l. m., ( ). the geometry of algebraic sys- tems and their exact solving using groebner bases, computing in science and engineering, , , – . doi: . /mcise. . [ ] roanes-lozano, e.; roanes-macías, e. and laita, l. m., ( ) some applications of gröbner bases, computing in science and engineering, , , – . doi: . /mcise. . [ ] robbiano, l. et al., ( ). cocoa . online help, (electronic file companying cocoa v. . ). [ ] stone, m., ( ). the theory of representations for boolean algebras, trans. amer. math. soc., , , – . doi: . / [ ] turner, r., ( ). logics for artificial intelligence, ellis horwood. [ ] winkler, f., ( ). polynomial algorithms in computer algebra. springer-verlag. eugenio roanes-lozano luis m. laita depto. de álgebra depto. cc. de la computación e i.a. facultad de educación facultad de informática universidad complutense de madrid universidad politécnica de madrid c/ rector royo villanova s/n campus de montegancedo -madrid boadilla del monte, -madrid spain spain eroanes@mat.ucm.es laita@fi.upm.es antonio hernando eugenio roanes-macı́as depto. de sistemas inteligentes aplicados depto. de álgebra escuela universitaria de informática facultad de educación universidad politécnica de madrid universidad complutense de madrid carretera de valencia km c/ rector royo villanova s/n -madrid -madrid spain spain ahernando@eui.upm.es roanes@mat.ucm.es http://dx.doi.org/ . /s - - - http://dx.doi.org/ . / - - - - _ http://dx.doi.org/ . /s - ( ) - http://dx.doi.org/ . /mcise. . http://dx.doi.org/ . /mcise. . http://dx.doi.org/ . / introduction an introduction to boolean algebras and boolean rings lattices and lattice orders boolean algebras boolean rings a polynomial model for propositional logic building the taylor-made polynomial (boolean) ring the polynomial boolean algebra corresponding to the polynomial (boolean) ring the boolean algebra isomorphism varphi ideals of a boolean algebra and ideals of a ring introductory note about multi-valued propositional logics kleene's and Łukasiewicz's three-valued logic tautological consequence the isomorphism varphi in the multi-valued case the main theorem a brief note about gröbner bases connecting ``to be a tautological consequence'' with an ideal membership some introductory notes about rbes rules potential facts and facts. integrity constraints the inference engine. rbes inconsistency knowledge extraction in rbes through the use of an algebraic inference engine a polynomial model for the propositional boolean algebra associated to rbes the main theorem adapted to knowledge extraction in rbes rbes strong and weak inconsistency. algebraic interpretation backward reasoning in rule based expert systems whose underlying logic is classic boolean logic implementation in a computer algebra system polynomial translation of the logical connectives cocoa implementation maple implementation description of the shell conclusions doi: . /j.eswa. . . a multiagent approach for diagnostic expert systems via the internet khaled shaalan a,*, mona el-badry b , ahmed rafea c a department of computer science, faculty of computers and information, cairo university, tharwat street, orman, giza , egypt b central laboratory for agricultural expert systems (claes), p.o. box , dokki, giza, egypt c department of computer science, american university, , sharia kasr el-aini, p.o. box , cairo, egypt abstract in recent years there has been considerable interest in the possibility of building complex problem solving systems as groups of co- operating experts. this has led us to develop a multiagent expert systems capable to run on servers that can support a large group of users (clients) who communicate with the system over the network. the system provides an architecture to coordinate the behavior of several specific agent types. two types of agents are involved. one type works on the server computer and the other type works on the client computers. the society of agents in our system consists of expert systems agents (diagnosis agents, and a treatment agent) working on the server side, each of which contains an autonomous knowledge-based system. typically, agents will have expertise in distinct but related domains. the whole system is capable of solving problems, which require the cumulative expertise of the agent community. besides to the user interface agent who employs an intelligent data collector, so-called communication model in kads, working on the client sides. we took the advantage of a successful pre-existing expert systems—developed at claes (central laboratory for agricultural expert systems, egypt)—for constructing an architecture of a community of cooperating agents. this paper describes our experience with decomposing the diagnosis expert systems into a multi-agent system. experiments on a set of test cases from real agricultural expert systems were preformed. the expert systems agents are implemented in knowledge representation object language (krol) and java languages using kads knowledge engineering methodology on the www platform. q elsevier ltd. all rights reserved. keywords: expert systems; agents; www . introduction as computer hardware and software become increas- ingly powerful, so do applications, which used to be considered beyond the scope of automation. to cope with increased demands, software systems are becoming correspondingly larger and more complex (genesereth & ketchpel, ; jennings & varga, ; jennings & wooldridge, ). in recent years, there has been considerable interest in the possibility of building complex problem solving systems as groups of co-operating experts. distributed artificial intelligence (dai) is the study of how such systems might be built. historically, empirical research in dai has focused on three main areas: blackboard architecture (engelmore & morgan, ; nwana, ), systems based on negotiation (smith, ), and multi-agent planning systems (durfee and lesser, ). more recently, co-operating expert systems have emerged as a research area of some importance (nwana & ndumu, ). agent-based computing represents an exciting new synthesis both for artificial intelligence (ai) and, more generally, computer science. it has the potential to significantly improve the theory and the practice of modeling, designing, and implementing computer systems (jennings, ). since the s, software agents and multi- agent systems have grown into what is now one of the most active areas of research and development activity in computing generally (jennings & wooldridge, ). there are many reasons for the current intensity of interest, but certainly one of the most important is that the concept of an agent as an autonomous system, capable of interacting with other agents in order to satisfy its design objectives, is a natural one for software designers (wooldridge & ciancar- ini, ). just as we can understand many systems as being composed of essentially passive objects, which have state, - /$ - see front matter q elsevier ltd. all rights reserved. doi: . /j.eswa. . . expert systems with applications ( ) – www.elsevier.com/locate/eswa * corresponding author. tel.: þ - - ; fax: þ - - - . e-mail addresses: shaalan@mail.claes.sci.eg (k. shaalan); mona@ mail.claes.sci.eg (m. el-badry); rafea@aucegypt.edu (a. rafea). http://www.elsevier.com/locate/eswa and upon which we can perform operations, so we can understand many others as being made up of interacting, semi-autonomous agents. there are several benefits of using the intelligent agent paradigm for software systems. agents can provide a high level of abstraction for dealing with intelligent systems and distributed systems. the agent abstraction is a natural extension of object-oriented technology, encapsulating the agents knowledge within an active process and providing a standard interface for intercommunications. this leads to another benefit of agent-based systems, interoperability (genesereth & ketchpel, ). with a common agent communication language, such as kqml (finin, fritzon, mckay, & mcentire, ; finin, labrou, & mayfield, ; labrou & finin, ), programs written as agents can communicate and cooperates with other such programs. in other words, agent-based system architecture provides a consistent interface for intelligent systems to interact with. finally, agent systems often provide a high-level ‘human- like’ interface to take the gui revolution one step further (moore et al., ). multi-agent systems are the best way to characterize or design distributed computing systems (huhns & stephens, ). multi-agent systems are commonly intended as computational systems, where several (semi-) autonomous entities interact or work together to perform some tasks. there are several motivations for having multiple agent systems (nwana, ). they include: † to solve problems that are too large for a centralized single agent to do due to resource limitations or the sheer risk of having one centralized system; † to allow for interconnecting and interoperation of multiple existing legacy systems such as expert systems and decision support systems; † to provide solutions to inherently distributed problems such as distributed sensor networks (durfee and rosenschein, ) or air-traffic control; † to enhance modularity (which reduces complexity), speed (due to parallelism), reliability (due to redun- dancy), flexibility (i.e. new tasks are composed more easily from the more modular organization) and reusability at the knowledge level (hence shareability of resources). the rationale for interconnecting computational agents and expert system is to enable them to cooperate in solving problems, to share expertise, to work in parallel on common problems, to be developed and implemented modularly, to be fault tolerant through redundancy, to represent multiple viewpoints and the knowledge of multiple experts, and to be reusable (huhns & stephens, ). nevertheless, with the advent of the internet, many researchers have been taking a closer look at distributed software systems. recently, a large share of this research has focused on intelligent distributed systems, which have come to be known as multi-agent systems. as a result, new development methodologies specifically designed for multi-agent systems have been introduced (iglesias, garijo, & gonzalez, ) and several tools are now available for building multiagent systems (raphael & deloach, ). at the central laboratory for agricultural expert system (claes), at the agriculture research center of ministry of agriculture and land reclamation in egypt, a number of successful agricultural expert systems were developed and deployed to a large number of users (rafea, , ; rafea, el-azhari, & hassan, ; rafea, el-azhari, ibrahim, soliman, & mahmoud, ; rafea, hassan, & hazman, ; rafea & mahmoud, ; rafea & salah, ; rafea, warkentin, & ruth, , ). these applications are traditional standalone systems. in these expert systems, the knowledge acquired from multiple domain experts that belongs to different disciplines are implemented as an integrated system. the aim of this research is to harness the potential of the agent technology for constructing a community of cooperating agents capable of diagnosing disorders in the agriculture domain. by implementing expert systems as multi-agents that perform their tasks remotely, the expertise can be published on the web. expert systems running on the internet can support a large group of users who communicate with the system over the network. this paper is organized as follows. section gives an overall structure of the proposed architecture of our system. section describes the user interface agent. section introduces the coordination agent. section presents the selected expert systems agents. section shows the ability of the system in producing diagnosis of disorders based on a set of input observations from real-world cases in the multi- agent environment. section concludes the paper. . overview of the proposed system our proposed system considers the two most common synergetic expert system applications—diagnosis and treat- ment. from the architectural point of view, the system provides a framework to coordinate the behavior of several specific agent types. the society of agents in our system consists of expert system agents (diagnosis agents, and a treatment agent) working on the server side, the diagnosis knowledge is distributed among several agents, each agent is an autonomous expert for a certain domain knowledge. typically, agents will have expertise in distinct but related domains. the whole system is capable of solving problems, which require the cumulative expertise of the agent community. in addition to the user interface agent who employs an intelligent data collector working on the client sides. those agents communicate by passing messages to the coordination agent. the clients communicate with the server through socket connection via the internet. at the client side the user interface agent send requests for k. shaalan et al. / expert systems with applications ( ) – diagnosis/treatment services to the server. answers are sent back to clients. we have applied our system in the domain knowledge of diagnosing disorders of the grapes crop, see fig. . a diagnosis agent is responsible for producing diagnosis of disorders based on a set of input observations. the treatment agent, who is responsible for finding a suitable recommendation to treat a certain disorder taking into consideration the possible causes of this disorder, provides a set of operations to be performed in order to treat the disorder and remove the causes. the treatment agent can only be consulted after a successful diagnosis session. in the early stage of this research, we have designed the system with only two expert agents, diagnosis and treatment. but later on, we discovered that the diagnosis agent combines different specialization that are well defined in the agricultural domain. consequently, we improved our architecture by furtherly decomposing the diagnosis agent. in the agent community decomposition concerns the partitioning of the problem domain into agents. the key problem lies in deciding how the domain is to be partitioned. this decision can only be reached by the extensive consultation with the domain experts. in our case, we figured out that the grape diagnosis consists of six agents, namely: fungal agent, insect agent, nematode agent, snail agent, mites agent, and nutrition deficiency agent. the new architecture concerning diagnosis has the following advan- tages: † it tries to simulate what may happen in real world in a very fine-grained manner. each agent corresponds to a domain expert specialization. † partitioning of domain knowledge makes it easy to maintain knowledge, and to improve performance. in real life, the domain expert who diagnoses a certain disorder can also give the treatment method of this disorder. the treatment methods suggested by different domain experts may conflict with each other such that a decision is needed to decide which one is applied first, and so forth. so, we proposed in our architecture to separate the treatment process from the diagnosis process, and collectively kept into the treatment agent in order to resolve conflicts that may arise in the treatment application. expert systems agents include a bilingual kqml translator module. the message transferred through the system takes the kqml format. expert system agents are implemented in krol (shaalan, rafea, & rafea, ), while the user interface agent is implemented in java. kqml message needs to be translated to/from prolog. so we have designed a module which performs this bilingual translation. similarly, another module for translating user interface messages is designed, which translates kqml to java and the vice versa. . the user interface agent the user interface agent provides access to the expert systems agents. for portability, this program is a java applet. when a user browses the www page of the system, this applet is downloaded. this is responsible for providing means through which the user could initiate problem- solving sessions, by activating the communication model, as well as handling the presentation of the expert system results. the communication model acts as an intelligent data collector. the user interface agent includes a java-kqml bilingual translator. . . the communication model common-kads project has addressed many issues related to the development of expert system. one of the most important issues is known as the communication fig. . the society of agents. k. shaalan et al. / expert systems with applications ( ) – model (hoog et al., ). one of the best advantages of the communication model is that it completely separates the user interface model from the knowledge model. however, dividing an integrated expert system into a two-component architecture—the problem solving component and an inter- face or a front-end component—is not usually a straightfor- ward task since the control of the interface is usually managed by the problem-solving component itself. in what follows we show the importance of the communication model in our system. for instance, in many expert systems, the questions to be asked are determined by answers to previously asked questions. in this case several solutions are possible. the first and the simplest of these, is devising an application specific interface where the user is presented with all possible inputs. needless to say, this approach will not meet the needs of any reasonably large application, and in addition, will confuse the user. another approach entails keeping control embedded within the problem solving server application. in that case, the interface will be used to present the user with an input request upon receiving such a request from the server. while this approach might seem reasonable, it suffers from major limitations. first, it relies on heavy communication between the server application and the client front-end, so the user may have to wait for prolonged periods of time depending on the network traffic and bandwidth. in case of synchronous communication, it can engage the server in one connection for an indefinite amount of time, making it impossible for other users to make use of that same server. although time-out operations could be implemented to avoid indefinite postponement, the application server will still not be fully utilized. in case of asynchronous communication the server will have to maintain extra knowledge such that data inputs could be mapped to application clients. this would be necessary in order to maintain data values that are consistent with its clients. if however, the interface component employed a communication model that had just enough knowledge about which inputs it should ask about and in which cases, then the client could use this knowledge to collect all needed inputs in an intelligent fashion. next, send them in one batch to the server for processing. in this case the connection between the client and the server will only be open for the period of sending the inputs, processing them at the server and receiving the output. for most practical applications, this period is usually reasonably short. meantime, if other clients need to service a request, the request will be placed in a wait queue where the waiting time will be short enough to make that wait transparent. however, care must be taken in selection of the queue size. otherwise, if the number of the clients for the application grows, then wait times might also grow to unsatisfactory figures. the implementation of the communication model is inspired by the hierarchical classification model (hc). hc is a problem solving method identified by chandrasekran ( ) for solving diagnostic types of problems as part of his generic task approach to expert system development. in hc, knowledge is represented as hypotheses hierarchically organized in a tree structure such that a general hypothesis is always above more specific ones in the tree. using a control strategy known as ‘establish and refine’, hypothese are explored top down. if a hypothesis at the top level succeeds (establishes), its immediate descendants are required to be established themselves one by one. this process of attempting to establish the descendants is referred to as ‘refining’ the parent hypothesis. if, on the other hand, a hypothesis fails, then it is said to be ruled out and so are all the hypothese beneath it in the tree (el-beltagy et al., ). in our model, knowledge components for which input is desired, are also organized in a hierarchical fashion. . the coordination agent this component acts as the interface between the user interface agent and other agents. it is responsible for establishing the communication link with the desired expert system agent based on the user request, through the socket connection. in the proposed agent society, tcp/ip is chosen as the low-level transport mechanism for agent-to-agent communication. in our implementation, a general-purpose client sockets class esclient is defined that inherits from socket class of the java.net package. each time the user interface agent needs to communicate with any of the other agents, it sends to the coordination agent a message. the coordination agent in turns creates an instance of this class giving it both the host names of the target agent, and the port number on which the agent’s server socket resides. host names and port numbers are passed in the first place to the coordination agent as parameters from the applet’s initiating html page. . the expert system agents each expert system agent consists of two basic components, a mediator and the expert system itself. as indicated by its name, a mediator is the module responsible for sending and receiving messages through the network connection. this communication can be established through a server socket. all agent communication takes place through the coordination agent. the expert system consists of three sub-components: domain knowledge: it contains the static knowledge of the domain, concepts, properties, and two types of relations. a concept is identified through its name. concepts can have properties that are defined through their names and descriptions of the values that the property can take. the relations may be relations k. shaalan et al. / expert systems with applications ( ) – between concepts or relations between property expressions. inference knowledge: describes what inferences can be made on the basis of the knowledge in the domain knowledge. it describes what inference can made, but not how or when they are made. in kads the selected knowledge engineering methodology, the following terms are used to denote the various aspects of a primitive inference: inference step, and role. an inference step performs an act that operates on some input data and has the capability of producing a new piece of information as its output. the elements on which inference steps operate or produce are called roles. a role can be an input role or an output role. it acts as a placeholder for domain objects. domain objects can be linked to more than one role. task knowledge: actually describes the steps of execut- ing the knowledge sources in the inference knowledge. in a previous work (shaalan et al., ), we described the implementation of expert systems in krol. in krol, webkrol is the library that handles all the network necessary function. . . the diagnosis agent as mentioned above there are six diagnosis agents. the user interface agent is responsible for loading the diagnosis communication model in the client side through a www browser. the diagnosis session proceeds as follows, see fig. . . the diagnosis communication model asks the user about all the observations and collects his responses. . the user interface agent calls the java-kqml translator module to convert the collected observation into a kqml message format. . the user interface agent sends this message to the coordination agent. . the coordination agent initiates the socket connection with the server and broadcast this message to all diagnosis agents at the server site. . at the server side, for each agent, after receiving the diagnosis data, the prolog-kqml translator translates the kqml message into a prolog term. this message is directly interpreted and executed by the diagnosis expert system. fig. . observation and disorder flow in the diagnosis session. k. shaalan et al. / expert systems with applications ( ) – . the results are converted again to a kqml format and sent back to the coordination agent. . the coordination agent collects all the answers and sends it to the user interface agent. . the user interface in turn formulates the output and displays it. . . the treatment agent treatment agent session can be conducted after a successful diagnosis session. the treatment sessions that correspond to this behavior is described as follows. after the diagnosis session reaches a successful disorders (i.e. at least one of the diagnosis agent find disorders). this information is sent to the coordination agent that transfers it to the user interface agent. if the user requests a treatment for these disorders, the user interface agent calls the treatment communication model and start collecting data related to treatment. both the treatment input data and the diagnosis result (the confirmed disorders) will transfer to the coordination agent. the coordination agent in turn opens the connection with the treatment agent and mediates this information to it. the treatment agent in turn tries to get a treatment for these disorders and sends the results back to the coordination agent that will transfers it to the user interface agent that will display its contents through the www browser, see fig. . . an example of utilization and testing to bring out the advantages of employing multi-agent technology and expert systems, it was necessary to look at some real test cases from their conception to realization. we fig. . disorder and treatment flow from a diagnosis session to a treatment session. k. shaalan et al. / expert systems with applications ( ) – have generated test cases that trigger mostly knowledge of the six diagnosis agents. for each case a brief description of its objective is given as well as its related input/output screens in diagnosis session and in treatment session. the function of coordination agent is shown through screens at run time. it is recommended that the end user of our system use a java enabled browser (e.g. internet explorer). throughout this section we show screens taken from running our system on the ms internet explorer. when the user connects to the server site through any available web browser, the front page shown in fig. appears. this page initiates the user interface agent and carries the necessary information about the address and the port numbers of all expert system agents. we have adopted a testing methodology based on manually generated test cases taken from the diagnosis and treatment documents of grapes crop. the objective is to cover all the agents, verify the proposed system by considering the number of expert systems agents that succeed in diagnosis: -agent succeed, -one agent succeed, and many agent succeed. the methodology also trace down the function of the coordination agent in each case. the main steps in our methodology are: . walk through the diagnosis design document and manually generate random test cases. it is worth noting that the generated test cases are by no means exhaustive. the test case is generated by walking backward through the inference chain, starting from a target disorder(s) and randomly selecting values that conclude the selection reached at any given point in the inference chain. . for each disorder obtained from step , walk through the treatment design document and manually generate random test cases. this time the test cases are generated by walking through the inference chain to get the treatment operation(s). . merge two or more test cases in order to test the behavior of the multi-agent environment. this will try to simulate real life situations, when it may happen that one or more disorders appear at the same time (this step is used for testing the possibilities of many-agent succeeded). . run the generated test cases on the system. . compare the results with the target disorder. for time and space constraints, we will show a test case that tries to diagnose and treat disorders in case of three diagnosis agents respond. . . verify three diagnosis agents: nutrition deficiency, mite, and snail objective. the objective of test case is to conduct an experiment that demonstrates the capabilities of the system to diagnose and treat disorders in case of three diagnosis agents respond. in this case we choose disorders that belong to nutrition deficiency, mite, and snail expert agents. this test case covers the diagnosis and the treatment for the disorders due to iron deficiency, zinc deficiency, manganese deficiency, mite, and land snail disorders. the treatment of iron deficiency will be on date / / , the material used is iron_chelate. the treatment of zinc deficiency will be on date / / , the material used is zinc_sulphate. the treatment of manganese deficiency will be on date / / , the material used is manganese_sulphate. the treatment of mite will be on date / / , the material used is vertamic. the treatment of land snail will be on date / / , the material used is iron_sulphate. . . . diagnosis session diagnosis input: growth stage: vegetative leaves color: yellow leaves status: small, drop color of spots: brown position of leaves color: whole leaf type of infection: new leaves range of infection: adjacent spots depth of water: medium snail exist: yes fig. . an example of input screens in the diagnosis session.fig. . front page screen. k. shaalan et al. / expert systems with applications ( ) – an example of input screens is shown in fig. . after all the required symptoms are collected, the user interface automatically sends them to the coordination agent. which in turn will establish the connection and broadcast the collected data to all the diagnosis agents. the answers from those agents are sent back to the user interface agent. diagnosis output: the output from the diagnosis agents of this case is shown in fig. , which indicates that the three agents that are involved in the test case have found that the input symptoms match a certain disorder in their knowledge base. the user interface asks the user whether or not it may call the treatment agent. assume we continue with the treatment agent. . . . treatment session the user interface agent will call the communication model of treatment. then the treatment communication model start asking about some information concerning treatment, such as the infection range, and the material used. after that the coordination agent starts the connection with the server and calls the treatment agent. treatment input: zinc used material: zinc sulphate. treatment output: the result from treatment agent again returns back to the user interface agent through the coordination agent. the recommended treatment includes the date of the treat operation, the material used, the material quantity, the method if any, and the treatment advice, see fig. . the right screen, is the detailed screen of the disorder shown on its left. . . . coordination agent fig. shows a trace of the flow of control of the output from coordination agent to the diagnosis agents, and the input from each agent at run time. each agent located in a server with server name listen to a port number. the input shows that there are three agents answer with reply (at port , , ) and all the other three agents answer with sorry. the coordination agent will then send to the treatment agent. . conclusions in this paper we have described our experience with decomposing the diagnosis expert system into a multi-agent system. our approach consists of a number of cooperating agents capable of solving the diagnosis and treatment problem in the agricultural domain. it has been applied to the grapes crop for providing services to a large group of users over the internet. the system can be geared towards any other related system or application. the widespread use of the internet and www provides an opportunity for developing es that fit into the agent technology widely available. fig. . output in the diagnosis session. fig. . an example of output screens in the treatment session. k. shaalan et al. / expert systems with applications ( ) – the new architecture concerning diagnosis has the following advantages: † it tries to simulate what may happen in real world in a very fine-grained manner. each agent corresponds to a domain expert specialization. † partitioning of domain knowledge makes it easy to maintain, easy to understand, and improve performance. † the distribution of responsibilities among agents, achieves modularity and reduced complexity. † re-usability at the knowledge level was achieved since the proposed agents are capable of providing their services to other expert systems as well as to other applications provided that these other applications use the same ontology and understand the same communi- cation language. the expert systems agents are implemented in krol using kads knowledge engineering methodology. the user interface agent and the coordination agent are implemented in java. experiments on a set of test cases from real agricultural expert system were performed. the results observed in our experiments were satisfactory and proved that the system is robust. the system can diagnose and treat disorders and the domain ontology consists of concepts and more than rules. references chandrasekaran, b. ( ). generic tasks as building blocks for knowl- edge-based system: the diagnosis and routine design examples. the knowledge engineering review, ( ), – . durfee, e. h., & lesser, v. r ( ). using partail global plans to coordinate distributed problem solvers. in a. h. bond l. gasser, & kaufmann, (eds.), readings in distributed artificial intelligence (pp. – ). san mateo, ca. durfee, e. h., & rosenschein, j. s ( ). distributed problem solving and multi-agent systems: comparisons and examples. in proceedings of the th international dai workshop. el-beltagy, s., rafea, a., kamel, a., sticklen, j., schulthess, u., & ward, r ( ). an expert system for wheat disorders diagnosis and treatment using a hierarchical classification problem solver. second ifac/ifip/ eurageng workshop on artificial intelligence in agriculture, wagenin- gen, netherlands. new york: pergamon press. engelmore, r. s., & morgan, t. (eds.), ( ). blackboard systems. reading, ma: addison-wesley. finin t., fritzon r., mckay d., & mcentire r ( ). koml as an agent communication language. in the proceeding of the third international conference on information and knowledge management. new york: acm press. available at http://www.cs.umbc.edu/kqml/papers. finin, t., labrou, y., & mayfield, j. ( ). kqml as an agent communication language. in j. m. bradshaw (ed.), software agents (pp. – ). cambridge, ma: the mit press. genesereth, m. r., & ketchpel, s. p. ( ). software agents communi- cations of the acm, ( ), – . hoog, r. de., martil, r., wielinga, b. j., taylor, r., bright, c., & velde, v. d. w ( ). the commonkads model set. deliverable esprit project p , kads-ii/wp i-ii/rr/uva/ / . . amsterdam: uni- versiteit van amsterdam. huhns, m. n., & stephens, l. ( ). multiagent systems and societies of agents. in g. weiss (ed.), multiagent systems. a modern approach to distributed artificial intelligence (pp. – ). cambridge, ma: mit press. iglesias, c., garijo, m., & gonzalez, j. ( ). a survey of agent-oriented methodologies. in j. p. müller, m. p. singh, & a. s. rao (eds.), intelligent agents: agents theories, architectures, and languages (vol. ). lecture notes in computer science, berlin: springer. jennings, n. r. ( ). on agent-based software engineering. artificial intelligence, , – . jennings, n. r., varga, l. z., aarnts, r. p., fuchs, j., & skarek, p. ( ). transforming standalone expert systems into a community of fig. . output/input to the coordination agent. k. shaalan et al. / expert systems with applications ( ) – http://www.cs.umbc.edu/kqml/papers cooperating agents. engineering applications of artificial intelligence, ( ), – . jennings, n. r., & wooldridge, m. ( ). applying agent technology. applied artificial intelligence, ( ), – . jennings, n. r., & wooldridge, m ( ). applications of intelligent agents. chapter . agent technology: foundations, applications, and markets. berlin: springer. labrou y., & finin t ( ). a proposal for a new kqml specification. online at http://www.cs.umbc.edu/kqml/papers/. moore, r., dowding, j., bratt, h., gawron, j. m., gorfu, y., & cheyer, a ( ). commandtalk: a spoken-language interface for battlefield simulations. in proceedings of the fifth conference on applied natural language processing, washington, dc (pp. – ). association for computational linguistics. nwana, h. s. ( ). software agents: an overview. the knowledge engineering review, ( ), – . nwana, h. s., & ndumu, d. t. ( ). a perspective on software agents research. the knowledge engineering review, ( ), – . rafea, a ( ). on integrating agricultural expert systems with databases and multimedia. in proceedings of the first international conference on multiple objective decision support systems for land, water, and environmental management: concepts, approaches, and applications, honolulu, hawaii, usa. rafea, a. ( ). agriculture. in j. liebwitz (ed.), a chapter in the handbook of applied expert systems. boca raton, fl: crc press. rafea, a., el-azhari, s., & hassan e ( ). integrating multimedia with expert systems for crop production management. in proceeding of the second ifac/ifip/enrageng workshop on artificial intelligence in agriculture, netherlands. rafea, a., el-azhari, s., ibrahim, i., soliman, e., & mahmoud, m ( ). experience with the development and deployment of expert systems in agriculture. in proceeding of iaai- , montreal, canada. rafea, a., hassan, h., & hazman, m. ( ). automatic knowledge acquisition tool for irrigation and fertilization expert systems. expert system with applications (eswa): an international journal, ( ), – . rafea, a., & mahmoud, m ( ). the evaluation and impact of neper wheat expert system. fourth international workshop on artificial intelligence in agriculture ifac/cigr, budapest, hungary. rafea, a., & salah, a. ( ). guiding object-oriented design via the knowledge level architecture: the irrigated wheat testbed. mathematical and computer modeling, ( ), – . rafea, a., warkentin, m., & ruth, s ( ). an expert system for cucumber production in plastic tunnels. in proceeding of the world congress on expert systems, florida, orlando, usa (pp. – ). rafea, a., warkentin, m., & ruth, s. ( ). knowledge engineering: creating expert systems for crop production management in egypt. in c. mann, & s. ruth (eds.), expert systems in developing countries: practice and promise (pp. – ). boulder, co: westview press. raphael, m. j., & deloach, s. a ( ). a knowledge base for knowledge-based multiagent system construction. national aero- space and electronics conference (naecon), dayton, oh, october – . shaalan, k., rafea, m., & rafea, a. ( ). krol: a knowledge representation object language on top of prolog. expert system with applications (eswa): an international journal, , – . smith, r. g. ( ). the contract net protocol. ieee transactions on computers, c- ( ). wooldridge, m., & ciancarini, p. ( ). agent-oriented software engineering: the state of the art. in p. ciancarini, & m. wooldridge (eds.), agent-oriented software engineering. lecture notes in ai volume, berlin: springer. k. shaalan et al. / expert systems with applications ( ) – http://www.cs.umbc.edu/kqml/papers/ a multiagent approach for diagnostic expert systems via the internet introduction overview of the proposed system the user interface agent the communication model the coordination agent the expert system agents the diagnosis agent the treatment agent an example of utilization and testing verify three diagnosis agents: nutrition deficiency, mite, and snail conclusions references nasa technical reports server (ntrs) pii: - ( ) - computers ind. engng vol. , nos i- , pp. - , - / $ . + . printed in great britain. all rights reserved copyright © pergamon press plc integrated decision support and expert systems in a computer integrated manufacturing environment chla-heo chang dept. of industrial & systems eng. university of michigan-dearborn dearborn, michigan jimmlng t. lin faculty of adrnlnlstrstion university of ottawa ottawa, ontario abstract in a computer integrated manufacturing (cim) environment, well planning, control and operational process require both expert knowledge of the area, and powerful decision support capabilities. this paper discusses the features of decision support systems and expert systems, and their integration to support the major functions from marketing and strategic level considerations to manufacturing operational planning and process. from the hierarchical structure of information flow in a company, this paper attempts to find the best way of combining decision support systems with expert systems in enhancing the planning, control and operational functions in a cim environment. introduction application of information technology has proven significantly improve the productivities in the manufacturing environment. more and more manufacturing functions are computerized, and results are usually encouraging. among them we can think of cad (computer-aided design), cam (computer-aided manufacturing), capp (computer-aided process planning), cat (computer-aided testing), etc.. then cim (computer integrated manufacturing), an evolved concept, represents an integration of all the relevant activities. in fact, not only engineering functions are involved, the recent development believes business functions like marketing, sales, distributions, finance and administration are also relevant to the manufacturing production and planning processes. ibm's copics (communication oriented production information and control system) and mapic ii (manufacturing accounting and production information control system version ) integrate information across functional areas and manage the resources of manpower, facilities and materials effectively (see [ ]). the manufacturing production functions use business data to expedite planning, forecasting, scheduling and control. the trend is to develop a fully integrated system. application of information technology in clm decision sunoort svstems manufacturing planning and control functions were handled manually until the late sixties. practitioners finally accepted information technology in the manufacturing environment. ibm brought out an integrated concept called copics (communication oriented production information and control system), which significantly influenced the further development of computer-based production planning and control systems. one group of the early products was referred to as material requirements planning systems (mrp) which generated material requirements for the manufacturing production lines. although many successful cases were brought forth, a rather large portion of the attempts to use mrp ended in dissatisfaction and sometimes even failures. after years of experience, the major causes to mrp's failure were gradually found and corrected. first of all, provided every information known, mrp could mechanically calculate material requirements very precisely, but unfortunately most manufacturing environments were constantly undergoing dynamic changes. secondly, it would not be realistic to develop production and requirement planning without the awareness of support from manufacturing resources. that included manpower, machinery, material and capital requirements. realizing the drawback, the new version of mrp with more decison support routines was developed under a new name called manufacturing resource planning (mrp ii). compared to mrp, mrp ii puts more emphasis on matching resources with demands in order to attain the corporative plans, that which will eventually be converted to production chang and lin: decision support and expert systems in a cim environment plans. the process is often a recursive one. attempts and adjustments are required to ensure the feasibility of the outcomes. potential capacity problems will be identified. the long term capacity planning pinpoints the discrepancy between the demands and the available resources in the long run, and decides on what kind of adjustment is needed to support a realistic master production scheduling. getting to the operational level, short term time phased production scheduling is set to observe the predicted loads, job-shop status and bottlenecks. the short term capacity planning will handle the adjustments by temporarily increasing or decreasing its capacities. with those modifications, mrp ii becomes more flexible in handling the dynamic environment and the output scheduling is also more realistic (see [ ]). mrp ii and likewise systems support manufacturing management with necessary information so that the planners can make judgment and decisions. let's use the long term capacity planning as an example. this function needs input knowledge such as the forecast of demands, the current capacity status, the inventory status and the production status. the forecast of demands and master production schedules generated by the systems will indicate a rough estimated manpower and machine capacity in each level of the assembly process. a long term (rough-cut) time-phased capacity requirement will be generated and compared against the existing capacity status. if a manufacturing firm is found undercapacity, the finding may justify the needs for increasing man and machine power. alternatives such as hiring new workers, increasing work shifts, opening new production lines, installing new machines, replacing current facilities and equipments with more efficient ones, and even building a new plant would be considered possible solutions. on the other hand, when the firm is found constantly overcapacity, solutions like reducing working shifts, laying off employees, closing clown production lines and plants will be under consideration to correct the problem. decision support systems will be used to test alternatives in either case. the objective is to develop a realistic and optimum long term capacity plan. the help from the information models often sharpens up such planning and control functions and improves the overall competitive status. although the information technology is getting more powerful and sophisticated than ever, in many situations human intelligence and expertise still play a determinant role. often users' limited knowledge in their areas of responsibility would affect their level of performance. in recent years, the development of knowledge based expert systems provides a solution to permit an expert's performance to the ordinary users. although the commercialization of expert systems began only a few years ago, the progress in developing applications for business and industry has been dramatic. manufacturing systems, by no exception, also find the invaluable use of expert systems to support their operations and management. for example, it is no news .for manufacturing engineers and operators to face substantial production line shut down caused by the failure of complex equipments, or spend costly process planning for each new manufacturing part, or go through time consuming scheduling at planning and shop floor levels. all these activities need the knowledge and expertise to perform their corresponding functions. the development of expert systems gives a lot of hope in those areas. to construct a knowledge based expert system, human expertise is transferred into means that can be accessed by the system, and forms a knowledge base. knowledge and expertise often need refinement judged by the testing process, and should constantly be updated. the process will continue to keep the system's performance at the export level. inference mechanism of the system accesses the knowledge base to generate analysis and judgment. user-friendly interface allows non-technical users to operate the system without extensive training. some system provides explanations behind the generated output when questioned by the user, or suggests alternatives from which the user can choose. for more practical use, some expert systems are constructed so as to make judgment from incomplete or fuzzy information. although it is costly and time-consuming to build an expert system, there are a lot of attractive benefits. the duplication cost is very low and easy. the systems are highly mobile, eliminating geographical barriers and are available twenty-four hours a day. besides, the system can incorporate knowledge from different experts from different time and locations into a common knowledge base. however, expert systems do have limitations. for one, they are usually designed to serve specific problems only. they are not the general problem solver, and it is too costly to be one. because of that reason, only the specific complex problems are considered the targets of expert systems. process planning, for example, is a major application of expert systems in the manufacturing environment. the function converts the design information into manufacturing process information in terms of process routes, process equipments and tools, proceedings of the th annual conference on computers & industrial engineering ~ j ~ marketing ~ ~ i stratig|escompany &~requirementi - ~ & customeri~ i ~specification i v i re.anohi i business plans + der/and ~ planning" process control ~'~ f t design planning planning i.o~.~;~ m,,~ro~ material bill of i i material requirements planning i .ort-ter.i - i . ~ r i production inventory "anage"ent~-- ~m°ngi-i ~ vendor i . .=ction i i suppliee i l < / w a r e h o u s , n g i ipost-salesi - * j , n s p e c t , o . packag,ng • service • " ) w ~ v l shipping i i feedback i v f i g u r e i n f o r m a t i o n f l o w in a c i m e n v i r o n m e n t operation sequences, etc., and is considered to be the key element in integrating the computer-aided design (cad) and the computer-aided manufacturing (cam). the function requires a considerable amount of expertise, has been the area of expert system application ever since the early eighties. the maintenance of complex equipment is another example. it usually involves fault diagnostic procedures incorporating various maintenance rules and judgment decisions. the personnel expertise and knowledge play an important role in locating the problem and determining the appropriate corrective actions (see [ ] and [ ]). integration of decision support and expert systems p_~i_vjmmcea decision support system provides the flexibility and adaptability in dealing with ill-defined problems, partial information, conflicting objectives and ad hoc questions. such a system is found particularly useful to support the solutions of a wide range of semistructured and unstructured problems like strategic and tactic planning. knowledge-based expert system incorporates knowledge from experts to provide users with expert level considerations. most expert systems are developed to support specific and narrow application domains. dss and es represent two kinds of information technology tools serving for distinct purposes, yet they are not necessarily conflicting to each other. in most situations, they can actually complement each other. a combination of dss and es can yield an even better result. user may use dss regarding information acquisition and information evaluation, and at the same time use es to get intelligence for a particular domain and a suggestion for the tentative decision. the joint effort can provide users with better information to make final decisions. information flow in a clm environment to determine which kind of information technology, dss or es, should be applied to which function in a cim environment, or whether the integration of dss and es would be appropriate for a particular function, one should first study the information flow in the cim. most companies now believe the importance of "the voice of customer", and it is the customer requirements that drive the company's operations. the product life cycle in such companies will start with marketing research, where customer needs and competitors' strategies are studied. collected information will help to set up long term strategies and develop business plans. that is a typical area where traditional decision support systems demonstrate their abilities using modeling, "what-if" simulation, and quantitative analysis to help user reason and make decisions. the marketing research and business plans determine the kinds of product the company wants to produce. the specific descriptions of the product rely on the design and manufacturing considerations. it used to be an iterative process between the product design division and manufacturing division to ensure that the final design meets customer requirements and will not cause manufacturing chang and lin: decision support and expert systems in a cim environment complications. the recent simultaneous engineering concept encourages both design and manufacturing divisions participate in the design process and work together, so the design will be manufacturable and also fully functional in the way customers expect. the final design will be converted to manufacturing process plans by the process planning function. this function requires a lot of knowledge about the manufacturing environment and expertise to develop the route plans. numerous existing expert systems are able to perform this function. then the process control plans will be developed to ensure the quality during operation. on the other hand, the forecast of product demands and the information of available capacities and resources are used to generate manufacturing planning schemes. the master production schedules (mps) lay out a time-phased product demand, and the material requirements planning (mrp) determines the requirements for each level of a product's structure (bill-of-material). at shop floor, the foremen will adjust the production schedule based on the short term available capacities, resources and production orders. operations at the production lines usually emphasize on efficiency, productivity and high quality. various expert systems were developed to support the individual operation (see [ ] and [ ]). inteoratlon of dss and es following the flow of information in a cim environment, we may classify the major functions into three levels: long term strategic planning, short term process and operational planning, and production operations. marketing and customer research, setting company strategies and business plans all belong to the first category. traditional decision support systems are developed to support such functions. but expert systems can further play the role of consultants by providing expert advice on complex issues that deal with legal problems, tax problems and others. in that case, outputs from expert systems will be directed to the dss for possible solutions. dss here is still the main tool for the user, and es only plays the supporting role (see [ ] and [ ]). from the development of company strategies and business plans to the production stage, mps, mrp, capacity planning and shop floor control follow one information flow path that is served by mrp (or mrp ii) systems. product design, process and control planning follow the other information flow path, where numerous expert systems have been developed to support these individual functions. the integration of dss and es at this level will keep all the current systems with additional components to enhance the overall result. along the first information flow path, expert knowledge about the customer behavior and forecasting is required to develop viable master production schedules. to choose a capacity adjustment strategy, an expert's advice will make a significant difference. most shop floor control requires knowledge of manufacturing and plant operation procedures. therefore expert systems should be developed to assist dsss in those aforementioned functions. along the second information flow path, there are expert systems already developed to support product design function and process planning. the results of quantitative analysis provided by dsss are often used as input to those expert systems. for example, engineering analysis is required during a design process, and dsss can be useful in choosing a combination of feeding rate and operation speed, and also other parameters in developing process plans. under that situation, dss outputs will be directed to the es (see [ ]). figure . integration of dss & es in an organization at the operational level, many production and operational processes can be benefited by using expert systems achieve expert performance. most operations at this level require little decision making. for specific operations at a narrow application domain, expert system is the ideal tool. operations such as welding, material handling, and machining are all becoming the service targets for expert systems (see [ ]). hence it would make a lot of sense to adopt an integrated system if other information systems, when combined with expert systems, would generate more economic and efficient results. therefore we should use expert systems to perform the operation and use dss and other information technologies as supplements in the system. r=.o.ugl.u_~jg~ decision support systems provide the flexibility and adaptability to support the solution of a wide range of semistructured and unstructured problems like strategic and long proceedings of the th annual conference on computers & industrial engineering term planning. expert systems incorporate [ ] knowledge from experts to provide users with expert level considerations in specific and narrow application domains. in a computer integrated manufacturing environment, well planning, control and operation require both expert knowledge of the area, and decision support capabilities. this paper discusses the [ ] features of decision support systems and expert systems, and their integration to support the major functions from marketing and strategic level considerations to manufacturing operational planning and process. from the hierarchical structure of information flow in a [ ] company, this paper attempts to find the best way of combining decision support systems with expert systems in enhancing the planning, control and operational functions in a cim environment. [ ] [ ] [ ] [ ] [ ] be.eebekge~ chang, c., "expert systems for manufacturing operations and management," proceedings of north amedcan ibscug conference, oxford, pp. - , . chang, c., "quality function deployment (qfd) processes in an integrated quality information system," int. j. of computers and industrial eng., vol. , nos. - , pp. - , . chang, c., "the structure of quality information system in a computer integrated manufacturing environment," int. j. of computers and industrial engineering, vol. , nos. . , pp. - , . chang, c. and t. steiner "computer aided manufacturing planning and control expert model," lasted applied simulation and modeling conference proceedings, vancouver, pp. - . . [ ] [ ] [ ] [ ] [ ] chang, c. and l. tsui "integrated manufacturing planning & control system using mrp i ," proceedings of the ismm int. symposium of computer applications in design, simulation and analysis, reno, pp. - , . chang, c. and k. yeung "manufacturing process planning expert systems," proceedings of third int. conference of expert systems: theory & applications, los angeles, pp. - , . groover, m. and e. zimmers, jr. cad/cam: computer-aided design and manufacturing, prentice hall, englewood cliffs, . heragu, s. and a. kusiak "analysis of expert systems in manufacturing design," ieee transactions on systems, man and cybernetics, vol. smc- , no. , pp. - , . laudon k. and j. laudon management information systems, macmillan pub., new york, . long, l, management information systems, prentice hall, englewood cliffs, . miska, k. "the new mavens of manufacturing," manufacturing engineering, oct. pp. - , rouse, w. "intelligent decision support for advanced manufacturing systems," manufacturing review, vol. , no. , pp. - , . sharit, j., r. eberts and g. salvendy "a proposed theoretical framework for design of decision support systems in computer-integrated manufacturing systems: a cognitive engineering approach," int. j. of production research, vo. , no. , pp. - , . pii: s - ( ) - computers and chemical engineering ( ) – use of expert systems for the synthesis of downstream protein processes m. elena lienqueo *, juan a. asenjo department of chemical engineering, centre for biochemical engineering and biotechnology, uni ersity of chile, beaucheff , santiago, chile abstract this paper describes recent developments in rational process design and their application in biotechnology for large-scale downstream processes. for implementing an expert system, it is necessary to divide the separation process in two parts: recovery and purification, because the information and available heuristic knowledge are different in each part. in the case of the recovery process, the implementation of an expert system, called prot – ex, only use heuristic rules from literature and human experts. the sequences suggested were tested with a real example and it gave a satisfactory result. for the purification process two criteria have been defined for selecting the optimal sequences, the ssc criterion and the purity criterion. both criteria were implemented in an expert system, prot – ex – purification. this expert system was validated with experimental examples, in two cases and, the sequences suggested were adequate and valid. however, the sequences suggested by purity criterion have lesser steps than the sequences suggested by ssc criterion and the purity algorithm spends lesser time and computational resources than ssc algorithm, then the use of purity criterion is more recommendable for selection of purification sequences. finally, we consider the proposed sequences found by the expert system a very good starting point for industrial process selection, a clear case of ‘expert amplification’. © elsevier science ltd. all rights reserved. keywords: protein recovery; protein purification; downstream processing; bioseparation; expert system www.elsevier.com/locate/compchemeng . introduction in the modern biotechnology industry, after success- ful fermentation or enzyme reactions the desired product must be separated and purified. the necessary steps to obtain these are commonly known as down- stream processing (dsp), which can account for up to % of the total cost, excluding the cost of the pur- chased raw materials (lee, ). dsp is usually com- posed of a sequence of recovery and purification operations, whose final aim is to obtain the required protein at a prespecified level of purity. the recovery process is characterised by the objective of obtaining the product in solution from the production system. purification takes the multiprotein solution and purifies the individual protein product to a high purity level, which is generally more than % pure. in addition, on a large scale, it is necessary to obtain the highest possible yield while minimizing the resources utilised and hence the cost (asenjo, ). the main steps in a large-scale protein separation process are usually not more than eight, they are shown in fig. . the principal unit operations used in dsp are shown in table . at present, the bio-product markets are becoming very competitive. therefore, it is important to choose the optimal flowsheet as early as possible, since once the process has been approved by regulatory agencies, its characteristics are fixed and are expensive to change. thus, it becomes necessary to use rational design tools for this purpose. in order to achieve this we can recognise two different tasks: . selection of separation sequence or flowsheet generations. . design of individual unit operations. in this paper we will center our attention in the first task. selection of separation sequences is a complex procedure in which a design evolves from a preliminary to a final stage in a trial and error fashion by repeatedly revising and refining the initial assumptions and restric- � ii pan american workshop in catalysis and process systems engineering, september – , , santa fe, argentina * tel.: + - - ; fax: + - - . e -mail address: bagajewicz@mailhost.ecn.uoknor.edu (m. baga- jewicz). - / /$ - see front matter © elsevier science ltd. all rights reserved. pii: s - ( ) - m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – tions. for protein recovery processes, the flowsheet generations always have the same type of unit opera- tions (see table ). the problem that has to be solved is, e.g. choosing between alternative operations. for example, for cell disruption it is necessary to select between a homogenizer and a bed mill or for cell separation between cross-flow microfiltration and cen- trifugation. this selection is more or less done using heuristics, i.e. using rules of thumb to arrive at a rapid and reliable specification of equipment type (leser & asenjo, ). another way to design is using appropri- ate mathematical correlations and models for optimising the task. on the other hand, for the protein purification pro- cesses the flowsheet consists of a chromatographic se- quence (see table ). this sequence has to be satisfied with maximum yield and a minimum number of steps ( , or ). this part of the process is undertaken in classical chemical process engineering, finding a rigorous solution using numerical methods like mathematical optimising techniques (jennings, teo, wang & yu, ) or using an expert systems approach (eriksson, sandahl, brewer & osterlund, ; forslund, ). for the selection of an optimal sequence, purely mathematical techniques have limited use in biotechnology because of a lack of useful design equations and databases. the expert sys- tems approach is more attractive because it allows the use of empirical knowledge that is not rigorous in nature and is typical of that used by experts (asenjo & maugeri, ). in this paper, we show the implementa- tion and validation of an expert system for recovery operations and another one for purification processes. fig. . general flowsheet for downstream processing of proteins. table principal steps and unit operations used in downstream processing stepstages unit operations protein reco ery cell separation centrifugation, membrane process (filtration, microfiltration), two phase aqueous partitioning homogenisation, bead milling, chemical and enzymatic lysis and permeablizationcell disruption (only for intracellular proteins) debris separation (only for centrifugation, membrane processes (microfiltration, ultrafiltration), aqueous two-phase partitioningintracellular proteins) concentration ultrafiltration, precipitation protein purification pre-treatment or primary isolation batch adsorption, ion exchange chromatography, affinity adsorption, hydrophobic interaction chromatography, aqueous two-phase partitioning hydrophobic interaction chromatography (hic), high resolution ion-exchangehigh resolution purification chromatography, affinity chromatography, metal chelate chromatography (imac) dye-ligand chromatography hplc polishing of final product gel filtration (gf) hplc, ion-exchange chromatography m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – table potential constraints observationsconstraints the product is only stable in a range of phsrange of ph (e.g. – ) the product is stable only up to certainrange of temperature temperature elimination of proteases, because they couldproteases degrade the protein product inclusion bodies inclusion bodies have to be solubilized and refolded for the production bulk industrial enzymes this is not the case. . how big is the flow rate? bigger than m /h. smaller than m /h. . . characterisation of starting material the second important point is the characterisation of the starting material. the process design will mainly depend on the physical, chemical and biochemical properties of the contaminating materials in the original broth and those of the protein that will constitute the final product. the properties of the starting material will be partially determined by: . its fermentation source. bacterial. yeast. mammalian cell. . how big is the cell concentration? . the type of cultivation medium used. presence of albumin. calf serum. presence of proteases. solid bodies like whole cells or cell debris. . localization of the product. intracellular. extracellular. . physicochemical properties of the product and the contaminant proteins in the final protein mixture in solution. surface charge at different phs and pi. surface hydrophobicity. molecular weight. biospecificity toward certain ligands. . stability of the final product is also of extreme importance as this will affect the types of operations that can be used as well as the conditioning and processing times that can be afforded. . . define possible separation steps and constraints it is necessary to have a database with all possible separation steps (table ) and to consider all potential constraints (table ). . . e aluation of potential process integration finally, the production process should try to inte- grate as much as possible, upstream, fermentation and downstream process. asenjo and leser ( ) have described possible areas and levels of process integra- tion. process integration can involve either the whole process, specific parts of it or particular unit operations as shown in fig. . examples of process integration between fermentation and downstream and between downstream operations are shown in table . . general aspects for designing a protein separation process the design of a protein recovery and purification process shares many characteristics with other engineer- ing design activities. to design a process or an opera- tion requires: . to satisfy of a number of constraints, such as purity, quality, process temperature and ph, desired yield, etc. . to know properties of the materials, for instance chemical and biochemical properties, thermody- namic and fluid properties of the process material. using the information and knowledge of ( ) and ( ) to propose a sequence of equipment interconnected in a particular order. . basic information for designing a separation process the design of a process to economically purify a protein, maintaining a high yield, yet obtaining a virtu- ally pure product while minimizing the cost, requires four main considerations: . defining final product. . characterisation of starting material. . define possible separation steps and constraints. . evaluation of possible process integration. . . defining final product it is necessary to define the final product and to have information on its used. for instance: . how the product is going to be used? industrial. diagnostic. laboratory reagent. therapeutic. questions regarding the use and purity are vital (e.g. , or . % pure) as well as allowable ranges of specific impurity concentrations. with therapeutic proteins any impurities have to be minimized whereas m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – fig. . the possible areas of process integration. fig. . the basic architecture of an expert system. the knowledge showed in the previous sections can be used for defining expert system for the rational selection of downstream protein processes. . expert systems expert systems are programs that attempt to solve problems in a way similar to how a human expert would solve them. they incorporate ‘rules of thumb’ that experts in the field have developed through years of experience. the problems tackled are not necessarily solved in a procedural fashion, they are often vague, complex, and can contain incomplete or inexact infor- mation (nebendahl, ). expert systems contain three basic elements: a knowledge base, an inference engine, and a user inter- face. the architecture of an expert system is shown in fig. . the knowledge base contains the information, which the program uses to reach decisions. the infer- ence engine is the program that manipulates the knowl- edge base to reach these decisions. finally, the user interface allows the program and the user to communi- cate with each other in an effective manner (harmon & king, ). there are software systems, called shells, used for the manipulation of heuristics, databases, and algebraic equations (like design equations). for in- stance, nexpert object™ and clips. there are many examples of the use of expert systems in chemical and biochemical engineering (siletti, ; mulholland, walker, manis, hinriks, buydens, blaffer & schoen- maker, ; jakus, ; forslund, ). . implementation of expert system the implementation of an expert system for separa- tion processes for proteins was divided in two different parts: . a first expert system, called prot – ex, for recovery process selection. . a second one, prot – ex – purification, for purifica- tion process selection. this division was carried out because both systems need different kind of information and the available heuristic knowledge is different, too (asenjo, herrera & byrne, ). prot – ex uses only heuristic knowledge whereas prot – ex – purification needs an important amount of quantitative data to make a good selection. in a consultation both expert systems are integrated and controlled by the main expert system. the main expert system: � receives the information from the user, database or other sources. � organizes and gives the information that each sub- expert system needs for suggesting a sequence. � receives the sequences suggested by each sub-expert system � gives the global sequences suggested to the user or other programme. table examples of possible process integration level of integration action cell immobilisation, cell recycling using membranes, use offermentation-downstream (the adequate design of the fermentation step can provide conditions to facilitate the downstream) expanded bed chromatography after fermentation use of liquid-liquid separation, use of expanded bed for proteininside downstream (recovery and purification) recovery, use of hydrophobic interaction chromatography after precipitation with ammonium sulphate m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – fig. . diagram of main expert system and relation with sub-expert systems prot ex and pro ex purification. the general diagram of the main expert system and sub-expert system with its relation is shown in fig. . . . prot – ex for reco ery process the expert system for the recovery process was im- plemented using only heuristic knowledge. the heuristic rules were obtained from the literature and renowned industrial experts were consulted to clarify specific points for which knowledge is missing or there is am- biguous information. to do this, it was necessary to provide the experts with a questionnaire (table ). for recovery processes, it is only necessary to choose between two or three options. for example, for separa- tion of cells if is necessary to select between centrifuga- tion and cross-flow microfiltration. this selection is based on the following variables: � thermal sensitivity of the product. � shear liability of cells. � size of the cells. � process throughput. � density difference between solid and liquid phase. � cell concentration in the broth. depending of the value of each variable it is possible to choose one or the other option. considering the previous points, the expert system prot – ex was implemented. it has more than six hun- dred logical rules to emulate human reasoning. its implementation was carried out in a commercial pro- gramme, nexpert object™ from neuron data. consul- tations were carried out to validate the expert system prot – ex. the basic information used for the expert system is shown in table . table displays a compari- son between a sequence suggested by prot – ex and a good process for recovery of somatotropin produced in escherichia coli (wheelwright, ). for this example, the recovery sequences suggested by prot – ex and the industrial process were very simi- lar. the difference with the published process is in step , where centrifugation was used instead of microfiltra- tion. for small cells such as e. coli microfiltration usually has clear economic advantage (asenjo & table typical questions presented to biotechnology experts for implementa- tion of the expert system for recovery processes (leser, ) questionaspect therapeutic what would your recommendation be if the product utilisation is therapeutic?product are the common ‘broth conditioningcell separation processes’, like heating, freezing, coagulation, flocculation, enzymatic or other used to facilitate cell separation in recombinant dsp or not? cell disruption would you agree that the best suggestion of a mechanical method for breaking cells is ‘high pressure homogenizer’ for most real situations? precipitation of what are the preferred methods for precipitation of nucleic acids?nucleic acid m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – table basic information for designing a purification process of somatotropin (lienqueo et al., ) defining final product protein’s name somatotropin . (estimated)pi . kda (estimated)molecular weight surface hydrophobicity . m (estimated) . mg/mlconcentration titration curve ph . . . . . . . . . charge (coulomb/molecule) − − . localization of the product intracelullar, inclusion bodies industrialutilization – ph stability range %final purity level not knownspecific affinity characterisation of starting material e. colifermentation source patrick, ). therefore, we considered that the se- quences suggested by prot – ex are adequate and it can be used as a starting point for the selection of a good industrial process. . . prot – ex – purification for selection of purification processes the selection of an efficient purification process is a critical step in downstream processes. these steps are not usually chosen in a rational manner, the common method being ‘trial and error’. therefore, the use of some basic heuristic rules for separation processes has been proposed. these rules are: (asenjo & patrick, ; prokopakis & asenjo, ): . choose the separation process based on the differ- ent physicochemical properties, such as: . . surface charge as a function of ph (titration curve) and pis. . . surface hydrophobicity. . . molecular weight. . . biochemical properties such as biospecificity with different ligands. . . stability at different temperatures and phs. . eliminate those contaminant proteins and com- pounds that are found in greater percentage first. . use a high-resolution step, as soon as possible. the chromatographic techniques ranked according to their efficiency are: . . affinity. . . ion exchange. . . hydrophobic interaction. . . gel filtration. . do the most arduous purification step at the end of the process (final polishing). considering those rules two selection criteria have been defined: seperation selection coefficient (ssc) criterion and purity criterion. the first one, ssc criterion, was developed by as- enjo and collaborators (asenjo, ; asenjo & maugeri, ; leser & asenjo, ). the ssc crite- rion considers that the key parameter is ssc, which caracterizes the ability of the purification operation to separate two proteins. ssc value is defined as: ssc = dfhui ( ) deviation factor (df), this variable relates the differ- ence between a property of the target protein (for example, the dimensionless retention time in a specific chromatographic technique) and the same property of the contaminant. df = �kdproduct − kdcontaminant� ( ) dimensionless retention time (kd), this variable rep- resents the behaviour of the proteins in a separation carried out by gel filtration, ion exchange or hydropho- bic interaction chromatography. mathematical relation- ships for predicting kd have been derived using the physicochemical properties of proteins (lienqueo, leser & asenjo, ). table comparison between sequences suggested by prot – ex and industrial process for recovery of somatotropin (lienqueo et al., ) sequence suggested bysteps industrial process prot – ex centrifugation crossflow microfiltration high-pressurehigh-pressure homogenizationhomogenization centrifugation-pellet wash disk centrifugation solubilizationsolubilization renaturation renaturation microfiltration ultrafiltration concentration and diafiltration m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – fig. . algorithm of ssc criterion. efficiency factor (h ) this parameter gives account of the unequal capability of the different separation pro- cesses to separate different proteins. its value is con- stant for each type of separation and the chromatographic material used and have been mea- sured experimentally (lienqueo et al, ). concentration factor (ui ) its value represents the relative concentration of each contaminant. it will af- fect the selection criteria. in this way contaminants which are present in a high concentration have to be eliminated first. ui = concentration of contaminanti concentration of all contaminants ( ) the amount of contaminant eliminated after a chro- matographic step has been mathematically calculated m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – for different situations (lienqueo, salgado & asenjo, ) then the ssc criterion chooses as best chromato- graphic operation the one that has the highest ssc value (leser, lienqueo & asenjo, ). after deter- mining which chromatographic technique gives the maximum ssc value, it is necessary to calculate the new concentration of all contaminant after the chro- matographic technique selected has been applied and construct a new database of contaminants concentra- tion. with this new database the system calculates the purity level and this value is compared with the level of purity required. the optimal sequence of steps is chosen until the required level of purity is reached. the al- gorithm used is shown in fig. . considering that the most important value is the final purity level and that now we had developed and al- gorithm to calculate the purity after a purification step, fig. . algorithm of purity criterion. m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – table main variables used for ssc and purity criteria (lienqueo et al., ) variable symbol definition ssc criterion ssc = dfhuiseparation ssc selection coefficient df = �kdproduct−kdcontaminant�deviation factor df kd relationships with physicochemicaldimensionless retention time properties of the proteins. (lienqueo et al., ) efficiency factor values for each chromatographich techniques shown in table peak width values for each chromatographics techniques shown in table uiconcentration ui factor = concentration of contaminant i concentration of all contaminantspurity criterion pj pjpurity = concentration of the target protein sconcentration of the proteins present one that has the least number of total steps. consider- ing these criteria an expert system was implemented, prot – ex – purification, and its implementation was car- ried out in the same shell than prot – ex. for testing these criteria, consultations were carried out to validate the expert system prot – ex – purification using both criteria. table and table show the basic information used for purification of bovine serum albu- min (bsa) and b-glucanase, respectively and table and table show a comparison between sequences suggested by prot – ex – purification and experimental sequence for both examples. . . comparison between ssc and purity criteria in both examples the sequences suggested are valid and adequate. nevertheless, in the first case, the se- quence suggested by the purity criterion has one fewer step than that suggested by the ssc criterion, that mean less capital cost. this situation takes place be- cause the ssc criterion considers the contaminant that gives the highest ssc value for one protein only. the purity criterion chooses the optimum chromatographic step considering all the contaminant proteins present. hence, this situation can occur when there are several contaminants present in similar quantities as was the case in this example. on other hand, the purity criterion spends less time and computational resources than the ssc criterion, since the purity algorithm gives the optimal sequences in less time than the ssc algorithm. . . comparison between sequences suggested and experimental sequences in both examples the sequences suggested result in experimentally valid ones and are a good starting point for an industrial process. there are differences between the estimated and experimental final purity level, which are less than %. these differences could originate in different causes: � hydrophobic interaction between the protein, this effect was not considered in the development of correlations for estimated dimensionless retention time for hydrophobic interaction chromatography (hic) (table ). � asymmetric chromatographic peaks, the peak shape could be different to a triangle shape. the differences between estimated and experimental purity level could be minimised: � improving the prediction of dimensionless retention time for hic. � using a gaussian approximation in the representa- tion of the chromatographic peak instead of triangle simplifications. table expressions and parameters used for ssc criteria and purity criteria (lienqueo et al., ) efficiency factor h peak width schromatographic techniques anion exchange . . cation exchange . . . hydrophobic interaction . gel filtration . . we implemented the purity criterion as a possible selec- tion criterion. this criterion compares the final purity level obtained after a particular chromatographic tech- niques has been applied. the purity concept is defined as: purity = concentration of the target protein sconcentration of the proteins present ( ) after determining which chromatographic technique gives the highest purity level, the system chooses this as the technique to use at this step. a sequence of steps is chosen until the required level of purity is reached. finally, the system creates a list with the defined se- quence of operations. the algorithm is shown in fig. the main parameters, variables and correlations used for ssc and purity criteria are summarized in tables and . both criteria consider that the cost of each chromatographic option is equal, except for affinity chromatography. therefore, the optimal system is the m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – . conclusions it is possible to use artificial intelligence tools, as expert systems, for selection of optimal protein separa- tion processes. for implementing an expert system, it is necessary to divide the separation process in two parts, because the information and available heuristic knowl- edge are different in each case. the first part for the recovery of proteins uses only heuristic rules from human experts and literature, the expert system was called prot – ex. the second part of a purification pro- cess is an hybrid expert system, which uses heuristic rules and mathematical correlations and models for selecting the optimal protein purification sequences, expert system called prot – ex – purification. both expert systems, implemented in the shell nexpert object™, were integrated by a main expert system. considering the cases studied, like the recovery of somatotropin, the purification of bsa and the purifica- tion of b-glucanase, the sequences proposed by the expert system are a good starting point for a practical industrial process. nevertheless, there are differences between estimated and experimental purity level, these could be minimized by improving correlations for pre- dicting retention time and approximation of the peak shape. in the specific case of prot – ex – purification, the sequences suggested by the purity criterion have less table basic information for desining a purification process of bsa from artificial mixture (lienqueo et al., ) defining final product protein’s name bovine serum albumin (bsa) . p i . kdamolecular weight surface hydrophobicity . m concentration . mg/ml titration curve . . ph . . . − . − . charge − . . − . (coulomb/molecule) − localization of the extracellular product industrialutilization ph stability range – final purity level % not knownspecific affinity characterisation of starting material artificial mixture (bsa, ovalbumin, thaumatin, soybean trypsin inhibitor)fermentation source hydrophobicitymw (kda)contaminant titration curve (coulomb/molecule) − initial concentration (m)(mg/ml) ph . ph . ph . ph . ph . − . . . . − . − . ovalbumin − . . sbti . . − . − . − . − . . . . . . . . thaumatin . . . table comparison between sequences suggested by prot – ex – purification and experimental process for purification of bsa from a mixture of four proteins (lienqueo et al., ) sequence suggested by prot – ex – purificationstep experimental validation process purity criterionssc criterion chromatography step purity chromatography step puritypurity chromatography step %cation exchange ph . anion exchange ph . % % anion exchange ph . hydrophobic interaction hydrophobic interaction % % hydrophobic interaction % % anion exchange ph . m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – table basic information for designing a purification process of b- , -glucanase (lienqueo et al., ) defining final product enzyme name b- , -glucanase (bgl iia) . pi kdamolecular weight . msurface hydrophobicity . mg/mlconcentration titration curve . . . . . ph . − . charge − . − . − . (coulomb/molecule) − extracellularlocalization of the product utilization industrial ph stability range – %final purity level not knownspecific affinity characterisation of starting material b.subtilus toc fermentation source contaminant mw (kda) hydrophobicityinitial concentration ph . titration curve (coulomb/molecule) − (mg/ml) (m) ph . ph . ph . ph . . . . . . − . − . − . . . . . − . − . − . . . . − . − . − . − . . . − . . − . − . − . . . − . − . . − . − . . . − . . − . − . − . . . . . − . − . − . . . . . − . − . − table comparison between sequences suggested by prot – ex – purification and experimental process for purification of b-glucanase (lienqueo et al., ) sequence suggested by prot – ex – purificationstep experimental validation process purity criterionssc criterion purity chromatography step purity chromatography stepchromatography step purity % hydrophobic interaction %hydrophobic interaction hydrophobic interaction % % anion exchange ph . % anion exchange ph . – %anion exchange ph . steps than the sequences suggested by ssc criterion, then the purity criterion is more recommendable. on other hand, the use of expert systems for the synthesis of downstream protein processes is a clear case of ‘expert amplification’. heuristic knowledge from experts was complemented with databases and design equations. to implement such a solution more globally there is a lack of databases of protein properties. with ad- vances in the area of proteomics and protein engineer- ing in a few years there could be sufficient information for such an approach. an effort should be made to generate such data. acknowledgements financial assistance from fundación andes and vicerrectoria académica, universidad de chile (beca pg/ / ). this project was also supported by proyecto catedra presidencial en ciencias. m.e. lienqueo, j.a. asenjo / computers and chemical engineering ( ) – references asenjo, j. a., herrera, l., & byrne, b. ( ). development of an expert system for selection and synthesis of protein purification processes. journal of biotechnology, , – . asenjo, j. a. ( ). selection of operations in separation processes. in j. a. asenjo, separation processes in biotechnology (pp. – ). new york: marcel dekker. asenjo, j. a., & patrick, i. ( ). large-scale protein purification. in e. l. v. harris, & s. angal, protein purification applications: a practical approach (pp. – ). oxford: irl press. asenjo, j. a., & maugeri, f. ( ). an expert system for selection and synthesis of protein purification processes. in p. todd, s. k. sikdar, & m. bier, frontiers in bioprocessing ii (pp. – ). washington: american chemical society. asenjo, j. a., & leser, e. w. ( ). process integration in biotech- nology. in m. verrall, downstream processing of natural products (pp. – ). chichester: wiley. eriksson, h., sandahl, k., brewer, j., & osterlund, b. ( ). reac- tive planning for chromatography: chemometrics and intelligent laboratory systems. laboratory information management, , – . forslund, g. ( ). designing for flexibility: a case study. expert systems, ( ), – . harmon, p., & king, d. ( ). expert systems: artificial intelligence in business (p. pp). new york: wiley. jakus, w. ( ). artificial intelligence in chemistry. collect czech chemistry communication, , – . jennings, l. s., teo, k. l., wang, f. y., & yu, q. ( ). optimal protein separation. computers & chemical engineering, ( ), – . lee, j. m. ( ). biochemical engineering (p. ). englewood cliffs, nj: prentice-hall. leser, e. w., & asenjo, j. a. ( ). rational design of purification processes for recombinant protein. journal of chromatography, , – . leser, e. w. ( ) prot – ex: an expert system for selecting the sequences of processes for downstream purification of proteins, phd thesis, university of reading, england. leser, e. w., lienqueo, m. e., & asenjo, j. a. ( ). implementa- tion in an expert system of selection rationale for purification processes for recombinant proteins. annals of the new york acadamy of sciences, , – . lienqueo, m. e., leser, e. w., & asenjo, j. a. ( ). an expert system for selection and synthesis of multistep protein separation processes. computers & chemical engineering, , – . lienqueo, m. e., salgado, j. c., & asenjo, j. a. ( ). an expert system for selection of protein purification processes: experimental validation. journal of chemical technology & biotechnology, , – . mulholland, m., walker, n., manis, f., hinriks, h., buydens, l., blaffer, t., & schoenmakers, p. ( ). expert system for re- peatability testing of high-performance liquid chromatographic methods. journal of chromatography, , – . nebendahl, d. ( ). expert systems: introduction to the technology and applications (p. ). new york: wiley. prokopakis, g. j., & asenjo, j. a. ( ). synthesis of downstream processes. in j. a. asenjo, separation processes in biotechnology (pp. – ). new york: marcel dekker. siletti, c. a. ( ). computer aided design of protein reco ery pro - cesses, phd thesis, (pp. ). massachussets institute of technol- ogy, cambridge, usa. wheelwright, s. m. ( ). protein purification: design and scale up of downstream processing (p. ). münchen: carl hanser. . - / send orders for reprints to reprints@benthamscience.net doi: . / , , , - the open bioinformatics journal content list available at: https://openbioinformaticsjournal.com research article an expert system to diagnose spinal disorders seyed m.s. dashti ,* and seyedeh f. dashti department of computer engineering, kerman branch, islamic azad university, kerman, iran department of advanced research, bushehr university of medical sciences, bushehr, iran abstract: objective: until now, traditional invasive approaches have been the only means being leveraged to diagnose spinal disorders. traditional manual diagnostics require a high workload, and diagnostic errors are likely to occur due to the prolonged work of physicians. in this research, we develop an expert system based on a hybrid inference algorithm and comprehensive integrated knowledge for assisting the experts in the fast and high-quality diagnosis of spinal disorders. methods: first, for each spinal anomaly, the accurate and integrated knowledge was acquired from related experts and resources. second, based on probability distributions and dependencies between symptoms of each anomaly, a unique numerical value known as certainty effect value was assigned to each symptom. third, a new hybrid inference algorithm was designed to obtain excellent performance, which was an incorporation of the backward chaining inference and theory of uncertainty. results: the proposed expert system was evaluated in two different phases, real-world samples, and medical records evaluation. evaluations show that in terms of real-world samples analysis, the system achieved excellent accuracy. application of the system on the sample with anomalies revealed the degree of severity of disorders and the risk of development of abnormalities in unhealthy and healthy patients. in the case of medical records analysis, our expert system proved to have promising performance, which was very close to those of experts. conclusion: evaluations suggest that the proposed expert system provides promising performance, helping specialists to validate the accuracy and integrity of their diagnosis. it can also serve as an intelligent educational software for medical students to gain familiarity with spinal disorder diagnosis process, and related symptoms. keywords: spine, expert system, uncertainty factor, spinal disorder, knowledge engineering, knowledge representation. article history received: october , revised: march , accepted: april , . background the development and expansion of information technology have brought about new and essential functions in computer- based decision-making systems. expert systems (es), as an ingredient of artificial intelligence, have a central role in these systems. decisions in an es are made via computers. in these systems, experts’ knowledge in a particular field of science is transferred to a computer, as a result of which experts’ way of thinking in that particular area can be emulated. more speci- * address correspondence to this author at the department of computer engineering, kerman branch, islamic azad university, kerman, iran, e-mail: dashti@iauk.ac.ir fically, the es detects the logical models through which an expert makes a decision, making judgments similar to those of human beings [ , ]. one of the specialized topics in physiotherapy and physical education is the detection of abnormalities. spinal disorders are usually caused by some behaviors, bad habits, or some diseases. the spine, as one of the most vital parts of the human body, continually changes, and in some cases, it may experience anomalies due to its structure and the spinal arrangements [ ]. one of the most severe anomalies involves a deformation of the spine and the upper body [ ]. the side effects of these anomalies include scoliosis, flat back, round back (thoracic kyphosis), swayback (lumbar lordosis), and cervical lordosis [ ]. nevertheless, no https://openbioinformaticsjournal.com http://crossmark.crossref.org/dialog/?doi= . / &domain=pdf mailto:dashti@iauk.ac.ir mailto:reprints@benthamscience.net http://dx.doi.org/ . / the open bioinformatics journal, , volume dashti and dashti medical es has been implemented or designed to detect and prevent spinal disorders. this study suggests an es designed to simulate the knowledge of experts and helps patients detect and prevent their harmful postural habits that could, in some cases, lead to premature death. a medical es is composed of some features that distinguish it from other conventional medical systems. one of the significant differences in these systems as far as obtaining results is concerned, is that they can emulate doctors’ reasoning and inference [ , ]. a medical es also helps to overcome some of the shortcomings that a human expert may display. the most critical shortcomings include outdated information, possible mistakes, geographical limitations, boredom, lack of quick or complete response time, unclear diagnostic processes, lack of flexibility, low accuracy, slow learning, and lack of varied expertise [ ]. patients at any age may develop spinal disorders. spine and intervertebral discs are curved but have different structures. in a healthy person, typically, the spine has four curvatures, including cervical lordosis, thoracic kyphosis, lumbar lordosis, and sacral kyphosis [ ]. these curves, in the bone and muscle structure of an average person, must be aligned at an acceptable angle. however, if they are deformed or experience abnormalities, they will cause anomalies and defects such as boredom, loss of movement functionality, respiratory restriction, reduced capacity of the heart, reduced blood circulation, and congenital disabilities. furthermore, due to the alternations imposed on the person’s appearance, she/he may be psychologically vulnerable, while being exposed to the risk of increased injury and damage as a result of pressure engendered by abnormal curvature [ ]. in the es proposed in this study, five anomalies, namely scoliosis, flat back, round back, swayback (lumbar lordosis), and cervical lordosis, are examined. two different theories form the basis of the proposed system, including backward chaining, the theory of uncertainty; each of them is described separately in the next chapters. the paper is organized as follows: in section , the rationale behind this research is discussed. in section , elements of inference are explained. studies in the literature are reviewed in section . in section , the methodology used in knowledge engineering is discussed. section describes the system evaluation investigated in this study. finally, the paper presents the concluding remarks in section . . rationale despite the development of medical science, industrial growth, and technology expansion in the modern world, there are still people threatened by new disorders, such as postural anomalies [ ]. spinal disorders which might be detected by the proposed expert system are described as follows: scoliosis: a sideways curve in the spine is called “scoliosis” [ - ]. this anomaly is initially c- shaped but develops into an s-like shape in its advanced forms. scoliosis can be detected by various methods [ , ]. the disorder is occasionally very progressive. if the curve increases, due to the pressure it exerts on the lungs, it reduces lung capacity and respiratory functioning, as a result of which some patients experience a slow death [ , ]. flat back: bounded curvature in the spine in the back or the waist, is called “flat back” [ ]. in this anomaly, the upper body switches into a vertical position, and mobility in the spine decreases [ ]. in the flat back, the cartilage and disc are damaged, and the defensive mechanism against imposed blows and spine resiliency both disappear [ ]. round back: increased back curvature leads to a condition known as “round back” [ , ]. severe round back is of two types, namely the irreversible and the reversible [ , ]. this anomaly disfigures the shape of the body, causing physiological effects and tightness in the chest. the tightness can disturb the activity of the cardio-pulmonary system [ ]. swayback: an increase in the back curvature is called “swayback”. in this anomaly, the curvature in the lumbar region increases [ ]. the stomach area is pulled forward, usually accompanied by pain and fatigue in the back [ , ]. swayback leads to pressure on vertebral discs and supposedly engenders back pain, makes childbirth difficult in women, and causes erosion in the bones [ ]. cervical lordosis: increased cervical curvature leads to a condition called “cervical lordosis.” as a result of this anomaly, the individual experiences a forward head posture, in which and the height of the neck appears to be shorter than its standard height [ ]. failure to detect and treat these anomalies may leave harmful impacts on physiological functions, as in the case of the kyphosis effect, which disturbs the respiratory system [ ]. in addition to physical damage, this adverse condition may even bring about psychological and social effects; for instance, kyphosis can be associated with depression [ ]. postural anomalies, apart from hereditary factors, are caused by industrialization, lack of exercise, bad eating habits, and the use of non-standard equipment [ ]. in everyday life, an individual may experience different postural positions, some of which can lead to some anomalies in the long run [ ]. additionally, individuals who use pieces of equipment that do not conform to ergonomic standards, and those who ignore anthropometric dimensions are more likely to develop structural and physiological disorders [ ]. recent longitudinal studies show that the most common form of scoliosis is late- onset idiopathic scoliosis [ ] that leads to a physical disorder, fatigue, and back pain. if left untreated, this disorder can even contribute to the mortality rate of the public population, because strenuous physical activity causes cardiac arrest over time [ , ]. there are two significant motivations behind designing and implementing the proposed es. first, helping experts to diagnose spinal anomalies more accurately; second, determining the potential risk of development of spinal disorders in healthy people. even more, since the integrity of knowledge is ensured in the development of the proposed system, it could be easily used by medical students to gain an diagnose spinal disorders the open bioinformatics journal, , volume understanding of the detection of spinal disorders and related symptoms. . basic elements in this section. first, we discuss the technical architecture of expert systems. next, fundamental theories and inference approaches used in the development of the proposed expert system are thoroughly explained to provide readers with the basics necessary to understand how the proposed approach works. . . rule-based expert system in order to find solutions, conventional computer problem- solving programs use well-structured algorithms, data structures, and crisp reasoning strategies. in the face of the severe problems with which expert systems are concerned, it may be more useful to employ heuristics: strategies that often lead to a correct solution but sometimes fail. conventional expert systems based on rules use human expertise to solve real-world problems that would typically require human intelligence. knowledge of experts is often expressed on a computer in the form of rules or data. based on the problem condition, specific rules and information can be retrieved to solve problems. rule-based expert systems have played a significant role in strategic goal setting, planning, development, scheduling, fault control, diagnosis, and so on in modern intelligent systems and their implementations [ ]. today's users can choose from hundreds of commercial software packages with friendly graphical user interfaces with the technological advances made in the last decade [ ]. conventional computer programs carry out tasks using a decision-making logic containing very little knowledge other than a basic algorithm to solve this particular problem. basic knowledge is often incorporated as part of the programming code so that as knowledge changes, the program has to be updated. small pieces of human knowledge are compiled from knowledge expert systems into a knowledge base; that is used to infer through a problem using appropriate knowledge. a significant benefit here is that using the same program without reprogramming attempts; a particular problem can be solved within the knowledge based domain. besides, expert systems can explain the rationale process and tackle the level of confidence and ambiguity that traditional algorithms do not manage [ ]. some of the critical advantages of expert systems are as follows: replication and maintenance of irreplaceable human experience; ability to deliver a system which is more reliable than human experts in terms of consistency; minimizing the need for human expert’s presence at several locations at the same time (especially in a dangerous environment that is hazardous to human health); solutions can be developed much faster when compared to the training procedure of human experts; figure displays the essential components of the expert system. all relevant information, details, rules, and relations used by the experts are stored in the knowledge base. the knowledge base may incorporate many human experts ' expertise [ ]. a rule is a conditional statement that links the premises to actions or results. fig. ( ). architecture of a typical expert system. knowledge base knowledge acquisition facility user interface explanation facilityinference engine expert with knowledge user the open bioinformatics journal, , volume dashti and dashti another method that is used to collect and store information in a knowledge base is the frame. it links an item or entity to a set of facts or values. frame-based representation of knowledge is a well-suited method for object-oriented programming techniques. also referred to as are expert systems that use frames to store knowledge in the knowledge base are usually referred to as frame-based expert systems. the inference engine aims to search for knowledge based information and relationships and provide answers, predictions and suggestions in the way a human expert might provide. the inference engine should find and compile the correct facts, definitions, and laws. two types of inference approaches are widely used– backward chaining is the practice of beginning with hypotheses and moving backward to the facts that support them [ ]. forward chaining begins with the evidence and progresses to the conclusions [ ]. the facility of explanation allows a user to understand how the expert system achieved those outcomes. the purpose of the knowledge acquisition facility is to provide an effective and secure means to collect and store all knowledge based components. expert user interface software is very often used to model, upgrade, and use expert systems. the user interface's purpose is to make the use of an expert system simpler for designers, users, and administrators. . . inference in rule-based expert systems a rule-based expert system is made up of if-then rules a set of facts, and an interpreter controlling the execution of the rules, fed with the facts. such if-then rule statements are used to formulate the conditional statements which form the full base of knowledge. a single if-then rule presumes the form 'if x is a then y is b' and if-part of rule 'x is a' is referred to as the antecedent or premise while the following part of rule ' y is b' is referred to as the consequence or conclusion for rule-based systems, there are two specific types of inference engines: forward chaining and backward chaining. the initial facts are first processed in a forward chaining system and continue to use the rules to draw new conclusions from those facts. the hypothesis (or solution/goal) we are trying to reach in a backward chaining system is processed first, and we continue to search for rules that would allow us to infer this hypothesis. new sub-goals are also set for testing as the processing progresses. primarily, forward chaining systems are driven by data, whereas backward chaining systems are goal-driven. in circumstances where information is costly to obtain, forward chaining strategy is particularly appropriate. if there is not enough information about what the result of inference might be, or if there is some particular hypothesis to be validated, forward chaining systems may not be efficient. backward chaining inference is useful in conditions where the amount of data is potentially very high and where it is of value to have some unique features of the system under consideration [ ]. . . backward chaining inference in the case of backward chaining, the primary concern is to align the conclusion of a rule against some known goal. so the' then' (consequent) part of the rule is generally not expressed as an action to be taken but rather as a condition, which is valid if the antecedent part(s) is correct. the backward chaining inference is analogous to the validation testing of the hypothesis in human problem-solving. for instance, a health- care specialist might suspect a patient's problems, which he/she then tries to prove by checking for specific symptoms. this reasoning style is designed by a goal-driven quest in an expert system and is called backward-chaining [ ], [ ]. it is a theoretical top-down model that starts with a goal or hypothesis and searches for rules to validate the hypothesis. it tries to balance the variables that lead to relevant data facts and shows that the inference moves backward from the intended goal to establish facts that would fulfill the goal [ ]. the implementation of backward chaining in a rule-based expert system is as follows: ( ) the system checks the memory to see if the target has been added. another knowledge base may have already proven the goal, so this phase is required. the algorithm reviews its set of rules, and if the goal has not been proven before, it continues to look for one or more that contains the goal in its then portion. this type of rule is called the goal rule. ( ) then the method looks at whether the target rule premises are listed in the memory. if the premises are not specified, new goals or sub-goals are to be checked, and other rules can be used to support them. ( ) the system continues in this recursive way until it discovers a premise not provided by any rule; it is called a ' primitive'. the algorithm asks the user for details about it when a primitive is identified. this knowledge is then used by the system to prove both the sub-goals and the original goal. . . certainty factor earlier rule-based expert systems assumed that all current information was either absolutely true or false. however, in the real world, there is often uncertainty associated with the set of rules and the data provided by the user [ ]. expert systems use the certainty factor (cf) to implement the theory of uncertainty, which produces a possible output value of certainty. thus, knowledge engineers consider some degree of uncertainty generated by uncertain occurrences. this consideration is specifically essential in the diagnosis of some diseases with uncertain consequences [ ]. the certainty value in this work is within the range of to . in the proposed inference algorithm, certainty factor (cf)s are assigned both to the premises and antecedents. as the rule's premise is uncertain because of uncertain facts, and the conclusion is uncertain because of the rule's specification, the following formula is used to estimate the conclusion's certainty factor: ( ) in some situations, there is more than one rule which supports a given conclusion. in this case, each of the rules can be fired and added to the cf. if a fact supports a conclusion and a rule fires for the favor of that conclusion, then the following equations will be used to evaluate the fact-related current cf: diagnose spinal disorders the open bioinformatics journal, , volume ( ) . literature review until now, many different approaches have been introduced to analyze spine condition and diagnose anomalies and vertebrae defects. detection of spinal anomalies is generally possible by non-invasive or invasive methods. some invasive methods include x-ray images, fluoroscopic, ct, and mri scans [ ]. non-invasive approaches fall into two categories. the first batch category includes the use of kyphometer, inclinometer, flexible ruler, spinal pantograph, electro-goniometer, spinal mouse. the second category, however, relies on non-contact methods, such as new york test and observation methods [ ]. one of the oldest methods, which is considered to be a common practice among professionals, is cobb's method. in addition to the risks of exposure to x-rays, time-consumption, and financial costs, this method is prone to errors caused by patients’ physical movement during the imaging procedure. this invasive process can also cause bone cancer in men, as well as breast cancer and abortion in women in some cases [ ]. the approach proposed by a study [ ] is a good example of non-invasive methods. in , a method was introduced for detecting anomalies of the spine, drawing on the image process, and the markers installed on the spinous process [ ]. the authors tested this method on forty male students of birjand university. in this method, markers were installed on the naked body, and images of the individual's movement were recorded through the motion analyzer camera. comparing their method with a flexible ruler, the researchers observed results that showed a pearson correlation coefficient of % for kyphosis and % for lumbar lordosis [ ]. this method involves some disadvantages, as it: requires special equipment including a motion analyzer and markers; requires the individual to be naked, which may not be the ideal option in all cultural contexts; requires background and proper lighting; lacks a specially designed software to be publically available; involves a high markers installation error, especially when they are set up by non-experts; has a non-inclusive algorithm, and it needs marker installation; an inclusive and comprehensive algorithm, nevertheless, could function without the use of the motion analyzer device and the markers. however, with the increasing power of computers and the availability of mobile devices to people, new opportunities were created in the field of health and medicine. first, it helped to increase the knowledge of people and communities about health-related issues and keys to a healthy life. second, new multidisciplinary branches in the field of health and medicine, namely bioinformatics, health informatics emerged. these fields relied on artificial intelligence and mathematical techniques as their basis for computation and reasoning. until now, many approaches have been introduced to address the challenging task of diagnosing orthopedic diseases, and more specifically, spine disorders and vertebral defects. methods in this area are roughly divided into two classes: methods based on machine learning and conventional rule-based approaches. . . methods based on machine learning the first group of approaches can grasp complex, non- linear relationships in the existing data [ ]. another study [ ] proposed a new approach based on random forests to diagnose osteoarthritis and to provide an easy interpretation of results by using a continuous regression output. in another work [ ], authors classified osteoarthritis subjects using knee joints ' sodium mris, which were transformed into radiant d acquisitions before and after the fluid suppression technique was applied. every patient was then defined using main features extracted from target image regions. a probabilistic boosting tree classifier was introduced in order to achieve a reliable segmentation approach to d vertebra ct spine images [ ]. information on all vertebras shapes was exploited by using statistical shape modeling techniques that supported the segmentation process. a weighted neighbor distance approach was manipulated in a few studies [ , ] along with a hierarchy combination of algorithms, which were first introduced in another study [ , ]. this approach represented morphology and was applied for osteoarthritis diagnosis in mris of articular cartilage scans. in practice, this program was an image classifier that extracted global features from a training set of images, and discriminately assigned weights to them. in a work, an svm based approach for determining needle entry site was developed. authors applied midline detection, and template matching approaches to ultrasound spine images; in order to find the best classification features. in this research, more than , images were analyzed, and the accuracy degree of % on the training set and % on the test set were respectively obtained. a variation of the svm model, named least squares svm (ls-svm), was manipulated by a study [ ] in order to distinguish among scoliosis curve types. these curve types were obtained from d trunk images by using optical digitizers. when obtained, the d images were divided into horizontal slices. then, all the slices were broken down into patches, and extraction of geometric thoracic and lumbar descriptors was completed, which were later used as classification features after a thorough dimensionality reduction with principal component analysis. in a research [ ], a computer-aided diagnosis system was developed to diagnose lumbar inter-vertebral discs degeneration anomaly the open bioinformatics journal, , volume dashti and dashti automatically. the proposed method analyzed and extracted the features from both t and t weighted mr images in order to acquire different types of feature sets. a generative model was proposed to predict the progression and shape of idiopathic scoliosis affected spines, using d spine reconstruction of a set of x-ray images [ ]. the researchers modeled multiple probabilistic structures, a geometric space with adequate structures and transform operators where high-dimensional data are reduced in size while retaining high-dimensional properties. a model based on deep learning and clustering techniques was proposed to detect adolescent idiopathic scoliosis (ais) automatically [ ]. in order to optimize the encoding-decoding of d spine model vectors, a neural network with an auto-encoder stack was trained. in two recent studies [ , ], a neural network was used, which was trained with ultrasound images for automatic detection of optimal vertebra level and percutaneous spinal needle injection plane. a neural network was designed and trained to partition ultrasound images into a multi-scale patch series recursively was designed [ ]. in each iteration, hadamard coefficients [ ] were manipulated to transform standard wave-like ultrasound signatures into signatures of region-orientation correlations to identify distinguishing features of specific spinal patterns. then, input images were classified for both epidural and facet joint injections as either belonging to or not belonging to the target plane. in a reported work [ ], a real- time scanner system was introduced. this scanner was implemented by using a convolutional neural network and a finite state transducer. the convolutional neural network was trained using a transfer learning methodology, where an inter- domain transfer of information was manipulated as a precondition for predicting accurately. thus, the convolutional neural network was able to detect and assign probabilities to three key positions along with the patient's vertebral scanning – sacrum, intervertebral gaps, and vertebral bones. in a more recent neural-based approach [ ], two deep learning pipelines were employed; one for the cervical and the other for lumbar vertebra detection, using annotated clinical mris with information labels for each spinal vertebra as training data. two cnns were used for each pipeline for the identification of more general and specific vertebral features. moreover, a component-based graphic model was built based on a layered graph to eliminate false positives and properly mark each vertebra. each layer of this graph shows the previously observed vertebrae and configurations of identified and marked vertebrae and measured the shortest path between them with the distance function based on mean and adjacent covariance matrices. despite the ability of ml models, this capability is not the complete solution to any inquiry which a healthcare provider may make at the time of treatment. ml algorithms could overfit “predictions in false data correlations”, [[ ]] and could identify predictors which “are not causes” despite their predictive power. . . rule-based approaches for a long time, rule-based medical systems have been a reliable tool for health practitioners to validate their assumptions [ ]. while providing a high degree of accuracy, these type of systems do not rely on any training data. the other advantage of a rule-based medical system, especially the medical rule-based expert system is transparency, being easier to understand the reasoning from which conclusions are derived. until now, few rule-based works have been reported in the field of orthopedics. authors believed the classification of cervical spine defects is the basis for surgical therapy and the recommendation for therapy [ ]. they reported an automatic rule-based classifier. this classifier required evidence from patient records, converted into a standardized form to facilitate data exchange and multimodal interoperability. a computer-based expert system was introduced to discover ergonomic risks for work-related musculoskeletal disorders. the proposed expert system utilized a knowledge base to classify risk factors into two central knowledge based units, in addition to four secondary knowledge based units that were related to the work environment and organization factors. this work suggested that a knowledge base system may be useful to assess a full-body assessment accurately. in a recent study, an electronic health records (ehr) retrieval system is developed [ ]. in particular, a retrieval system was introduced as a model for ehrs to support the medical decision-making process in managing spine defects using the information extracted from textual data on patient records. the patient cases were categorized according to the classes of cervical spine defects, and the classification was based on the rules obtained from the relevant defect classification scheme. the classifier was applied to medical documents in a retrospective study. in another study [ ], the authors presented a new -d segmentation, classification approach for structural changes in the hippocampal dendritic spines. an interactive rule-based module was leveraged for the classification of spines. morphological features described essential attributes of the segmented spinal shapes; then, spines were categorized automatically into one of four classes: stubby, filopodia, mushroom, and spine head. a rule-based expert system for diagnosis of neck diseases was reported in a study [ ]. the proposed expert system was based on forward chaining and decision trees inference; this system took user response in the form of yes/no to complete the chain of inference. while lacking technical and scientific rigor, the authors claimed that their proposed approach is helpful in the diagnosis of neck pain diseases. though many expert systems have been developed to help specialists with spine-related issues until now, no es has been implemented for diagnosing and estimating the risk of development of spinal abnormalities. . knowledge engineering method the implementation of the proposed es is displayed graphically in fig. ( ). the cycles in fig. ( ) reflect the processes of level , and the indices next to them describe the processes of level , which are directly related to level processes. in the proposed methodology, each process is repeated for any number of times until the medical expert system is formed in high quality and performance. the methodology is explained in this section. diagnose spinal disorders the open bioinformatics journal, , volume fig. ( ). phases of knowledge engineering methodology to build the proposed expert system. . . knowledge acquisition the necessary resources of knowledge acquisition for system design included the following items: ( ) four orthopaedists; ( ) two physiotherapists; ( ) a physical education instructor; ( ) articles, reports, and four research projects, including all related signs and symptoms. to design this system, the relevant knowledge for the detection and prevention of any deformities, as well as their symptoms and side effects, were identified by interviewing experts and using questionnaires. fig. ( ) explains the interview process thoroughly; moreover, a filled questionnaire to detect swayback abnormality is represented in table ) as a sample. both of these are crucial parts of the knowledge acquisition process in the early stages. • feasibility study • stating the reasons and motives • goals • determining resourcesassessment • acquisition of knowledge • inference algorithm design • designing the pruning method acquisition of knowledge and system design • implementation of the inference engine • user interface implementation • implementation of minor features in expert system implementation of expert system • functional testing • user-friendly testing • possible bug fixing • expert system evaluation system testing • system properties description documentation • fundamental changes and the growth of knowledge support the open bioinformatics journal, , volume dashti and dashti fig. ( ). an overview of interview process. table . sample of the filled questionnaire by an expert in the first review session. type of abnormality: round back definition: round back is a spinal abnormality in which an excessive outward curve of the spine results in an abnormal rounding of the upper back. typically, the thoracic spine should have a natural kyphosis between to degrees, while postural or structural abnormalities can result in a curve that is beyond this range. patients of all ages might be diagnosed with round back. nevertheless, this condition is common during adolescence, when the bones rapidly grow. diagnosis: -wall test -shortness of hamstring muscle test -checkerboard -plumb-line -spinal mouse -x-rays -mri scan side effects: - intervertebral disk herniation - pain in trapezius muscle and chest -disk space narrowing -pelvic obliquity -muscle imbalance -fatigue - tear or wear of the lumbar (lower) spine -loss of height -mild to severe back pain -difficulty standing straight upright -limited physical functions -digestive problems -body image problems -breathing problems causes: -degeneration of adjacent intervertebral discs -overstretched or weak upper back muscles -fractures -osteoporosis -scheuermann's disease - imheritence and birth defects -inheritance -syndromes - cancer and cancer treatments -tuberculosis -paget's disease - spina bifida - muscular dystrophy -neurofibromatosis next, the experts (orthopaedists) were asked to classify the symptoms based on their importance and rate of occurrence in patients with the spinal anomaly. after the classification of symptoms, experts assigned a unique certainty factor value to each symptom. this value determines the degree of certainty of observing a symptom related to an anomaly in unhealthy patients (table ). finally, the certainty factors assigned by the experts were summed, and the average value was considered to be a single certainty factor. in the third column of table , the final certainty factor can be seen. this factor was derived from the stages mentioned above. similarly, after classifying the symptoms associated with each other, the certainty effect factor in the fourth column of table was obtained through maximizing the certainty factor of the symptoms via formula . this factor denotes the cf of the rule premises and serves as a threshold to confirm the observance of symptom in a patient. fig. ( ) schematically illustrates certainty effect values related to flat back anomaly. session • experts were asked to fill the questionaries. highlights of interview session were recorded (a sample of filled questionna shown in table ). results • by reviewing the highlights, symptoms of each disorder w discovered. (a sample might be seen in the second column of tab session • with regards to dependencies between symptoms, experts asked to assign a certainty factor as a probability value between and one, to each symptom. results • average of certainty factor values given by the experts to symptom was calculated and selected as that symptoms' cert factor (column three of table represents a sample of cert factor value.) session • in order to ensure no piece of information is missing, experts requested to review the complete list of symptoms related to abnormality. in case of any alterations, the calculations were m from the beginning. results • the maximum certainty factor value of each symptom was select as the maximum certainty value (column three of table ). finally equation was used to calculate the value of the certainty effect factor (column four of table ). a graphical display of certainty eff values, related to flat back abnormality is shown in fig . diagnose spinal disorders the open bioinformatics journal, , volume table . sample of symptoms and related factors of flat back abnormality. # symptoms certainty from the perspective of an expert probability cumulative probability probability amendment symptoms class flat thoracic spine (loss of natural curve in the upper back) % . . . a flat lumbar spine (loss of natural low back curve) % . . . a reduced flexibility of the spine % . . . b loss of intervertebral disk space % . . . c degenerative disk disease % . . . d compression fractures % . . . e digestive problems % . . f table . sample of the certainty effect factor calculated using the certainty factor in table for the flat back anomaly. # symptoms class maximum certainty from the perspective of an expert in class certainty effect uncertainty effect cumulative certainty effect a % . . . b % . . . c % . . . d % . . . e % . . . f % . . fig. ( ). a sample of certainty effect factor for classified symptoms of the flat back anomaly. ( ) . . the proposed algorithm of inference engine after knowledge acquisition, inference methods and decision-making processes were determined based on the existing knowledge. to emulate experts’ thinking and to avoid the redundant occurrence of the rules, a hybrid algorithm was formulated for inference to obtain an excellent performance in terms of completeness, optimality, time complexity, and space complexity. the algorithm was an incorporation of the bci [ ] and uncertainty [ ]. the most frequently used method of the open bioinformatics journal, , volume dashti and dashti the uncertainty principle in the mess relies on the certainty factor [ ]. however, in the proposed hybrid algorithm, another factor, known as the “certainty effect factor” was used. applying this factor in the bci would make it possible for the system to function without storing users’ trail questions and answers. moreover, in every stage of the decision-making process, one could provide the feedback of statistical results by knowing the previous step of questions and answers. besides, one could prune the chain through a binary decision tree. because ess, especially medical diagnosis systems, are, in most cases, multi-agent, dynamic, inaccessible, uncertain, and continuous, the method proposed in this study drew on the certainty effect factor. this factor helped to simplify the system design by integrating symptoms into some classes and by removing redundant symptoms across the classes. this algorithm frames a “certainty memory” which functions similar to the human thinking process. the certainty memory sorts the rules in descending order according to the value of their symptoms’ certainty effect factor at each moment. the inference process begins when the user specifies the type of abnormality he/she wants to be diagnosed. next, a question is asked, and the user responds to that question by entering a numerical value within a range of and . if the input value is equal or higher than the present certainty effect value of the symptom in certainty memory, it is concluded that examined symptom is observed in the patient with a high degree of certainty. then, the system will remove the original value of that symptom’s “certainty effect” value. next, the user input value of that symptom will be replaced as the new “certainty effect factor” while the premise is satisfied. however, if the user response is any value smaller than the existing value, the predefined “certainty effect” value remains unchanged. that is due to the low chances of observing that symptom in the patient, and as a result, that premise will not be satisfied. afterward, all the satisfied premises are matched with their antecedents, and goal-related rules are fired. as soon as the rules are fired, the certainty factor (cf) of each rule is calculated based on its premises. at each stage, when a new rule is fired, the average sum of certainty factor of rules will be calculated as “certainty degree” of the chain of inference. equation is used to calculate this factor, which represents the likelihood of existence or risk of developing an anomaly in a person. while delivering a good level accuracy, this approach made efficient reasoning possible. the proposed inference algorithm and approximate responses would require fewer rules than the conventional method of the decision tree and explicit responses. the flowchart of this algorithm can be seen in fig. ( ). fig. ( ). flowchart of inference engine algorithm. start determining the certainty memory displaying certainty memory to the user user’s response in working memory update the knowledgebase and certainty memory is there any rule left end certainty effect percentage exit yes no determine the goal run goal selection rule running next rule in knowledgebase display certainty memory to user change in remaining rules’ conditions preparing knowledgebase and arraigning the rules in descending order according to certainty effect factor diagnose spinal disorders the open bioinformatics journal, , volume ( ) the advantages of this inference algorithm are as follows: in the proposed algorithm, the user’s questioning- answering process can be stopped at any point, while reaching an acceptable answer at any stage of the decision-making process, because the events are independent. there is no need to store the path because the certainty memory can compute the relevant statistics at any stage. the possibility to return to previous questions in this algorithm can be accomplished. . . implementation of knowledge base and user interface after knowledge acquisition was completed, and the inference algorithm was developed, knowledge engineers implemented the knowledge base, using java drools library. on the other side, system designers must realize that ordinary users are not software engineers or knowledge engineers who can work in a complex environment of an es. there is no doubt that systems with an inconvenient interface cannot be functionally efficient in actual practice. in this study, after the knowledge base was deployed, the graphical interface was designed by javafx. it has communicational features such as text-to-speech tool. fig. ( ) illustrates the proposed es during the questioning-answering process initiated by the user. during the evaluation, all the users could conveniently work with this system. the customized version of the es is compatible with these operating systems: windows xp, windows , windows , windows , windows server, windows vista, mac, linux, and open solaris. fig. ( ). a sample view of the proposed expert system during question-answering process. . . diagnosis in practice the developed expert system relies on the degree of certainty in a series of responses made by the user to determine the order of questions that have to be asked. after each response, user input is taken into account as a certainty amount, and the system state is updated. then, the system provides the most suitable question to the user. the question answering process ends at the point where no related question is left in the working memory. a flow of questions and answers to diagnose scoliosis abnormality in a -year-old female patient is shown in table . note that the patient was diagnosed with scoliosis abnormality after a close check-up by experts. the open bioinformatics journal, , volume dashti and dashti table . sample of question answering to diagnose scoliosis abnormality. # questions user input/degree of certainty what is the patient’s sex? female how old is the patient? how tall is the patient? (in cm) how much does the patient weight? to calculate weekly physical activity, answer following questions: average heart rate during activity: average daily physical activity average weekly physical activity low low does the patient’s spine have a sideways curve? % does the patient have an “s”- or “c”-shaped spine? % does one shoulder looks higher than the other? % does one shoulder blade stick out more than the other? % does the patient usually carry heavy objects using only one hand? or play tennis or throw spears? % are the hips uneven? % does the plumb line not pass within . cm of the posterosuperior corner of the s vertebral body ? % is one of the legs longer than the other? % are the abdominal muscles weak? % result: patient is diagnosed with scoliosis abnormality with certainty degree of % . results the developed expert system was evaluated in two separate phases. in the first phase, participants or real-world samples (patients diagnosed with spinal abnormalities and also healthy people) were used to examine the accuracy of the system. whereas, in the second phase, four orthopaedists evaluated the system using medical records of patients suffering from spinal disorders; based on their knowledge and experience. both phases are discussed as follows. . . phase : evaluation based on real-world samples two random statistical samples (groups of participants) were selected to evaluate the accuracy of the proposed es results. four orthopaedists were asked to introduce the es to the participants, to gain feedback about the functioning of the es. the first sample included individuals with spinal anomalies, while the second class of samples comprised of people who had a healthy spine. the evaluation process at this stage included predicting the likelihood of existence or chance of development of a spinal disorder in an individual. in doing so, individuals were selected to form the samples. one hundred twenty of them were already diagnosed with one of the spinal anomalies by the experts, while of them seemed to be in reasonable condition. each expert examined only samples from the group diagnosed with spinal disorders and samples from a healthy group of individuals. every one of the experts ran the system as a total number of times for unhealthy samples, and five times for each healthy sample in the second group. the reason behind repeating the question- answering process for the healthy samples was to determine the likelihood of the development of any of the five spinal disorders in that patient in the future. this could help the experts to prevent the development of spinal disorders in healthy patients by providing precautionary actions or giving the necessary tips to them. finally, the system was run for the total number of executions, and results were gathered. table shows the statistical information about the participants with anomalies, and table , represents the information related to healthy/normal individuals. x-, s and cv denote mean or average probability, standard deviation and coefficient of variation respectively. as table represents, evaluation results show a minimum average accuracy value of . for the correct detection of cervical lordosis in unhealthy patients. also, the system was successful in true detection of kyphosis disorder with a maximum average accuracy value of . in samples group suffering from the spinal anomaly. table demonstrates the mean value of the likelihood of developing any of the five anomalies in healthy samples. figs. ( and ) show the error bar for the detection of anomalies in both healthy and unhealthy groups of participants. table . the results of the expert system in examining the sample with anomalies. # anomaly name x- s cv scoliosis . . . flat back . . . kyphosis. . . . cervical lordosis . . . swayback . . . diagnose spinal disorders the open bioinformatics journal, , volume fig. ( ). the error bar, and phase evaluation results of the expert system for the samples with anomalies. fig. ( ). the error bar, and phase evaluation results of the expert system for the samples without anomaly. the open bioinformatics journal, , volume dashti and dashti table . the results of the expert system in examining the sample without anomalies. # anomaly name x- s cv scoliosis . . . flat back . . . kyphosis . . . cervical lordosis . . . swayback . . . . . phase : evaluations based on medical records four orthopaedists participated in evaluating the system. experts were provided with medical records of patients diagnosed with spinal abnormalities in earlier examinations. the records comprised of instances per each anomaly. all the records were segmented and placed under the related abnormality class. for each spinal disorder, cases were selected by each expert. experts were then asked to review the given medical records. features of these cases were later fed to es as inputs through question and answering process. each orthopaedist ran the system times for each anomaly, as a sum of executions by four experts for each anomaly and total executions of for all the anomalies. experts determined to what extent the results of the expert system are accurate in the real world applications, and how properly it can detect an abnormality. as table and fig. ( ) show, evaluation results from the four experts are very close, with minimum average accuracy value of . for true detection of cervical lordosis and swayback, and maximum average accuracy of . for true detection of kyphosis spinal disorder. table . evaluation results of the system by experts. # anomaly name expert no. expert no. expert no. expert no. x'- s cv scoliosis . . . . . . . flat back . . . . . . . kyphosis . . . . . . . cervical lordosis . . . . . . . swayback . . . . . . . fig. ( ). the error bar, and phase evaluation results of the system for each anomaly by experts. diagnose spinal disorders the open bioinformatics journal, , volume . . cut-off values in the evaluation procedure when the four experts completed the two phases of evaluation, they were asked to give their feedback about the system performance. based on their opinions, there are three different situations: first when a sample is correctly diagnosed, second when a sample needs closer examination, third when a given sample is healthy. then, experts were asked to describe their degree of certainty about system’s accuracy in each of these three cases in terms of numerical values. table shows a complete list of cut-off values for the detection of spinal abnormalities, which were obtained based on the experts’ feedback. threshold values were categorized into three classes: table . cut-off values estimated based on experts’ feed back. # anomaly name expert no. expert no. expert no. expert no. cut-off- true positive detection scoliosis . . . . . flat back . . . . . kyphosis . . . . . cervical lordosis . . . . . swayback . . . . . true negative detection scoliosis . . . . . flat back . . . . . kyphosis . . . . . cervical lordosis . . . . . swayback . . . . . -tpd or true positive detection: the minimum cut- off value showed that an examined sample is definitely suffering from any type of spinal disorders. -tnd or true negative detection: the maximum cut- off value showed that a given sample is healthy. -rfi or requirement of further examinations: any value lower than tpds or higher than tnds is regarded as a rfi, suggesting that a sample with diagnosis probability within this span of numbers requires close examinations. for each anomaly, the estimated average of values given by experts was considered as the final threshold value. . discussion experts participating in the evaluation of the system unanimously agreed over the integrity of the knowledge and validity of assessments. however, the proposed work is still subject to a few limitations. first, while experts found the evaluations to be reliable, they still argued over the degree of correctness of the results. for example, once the expert system detected scoliosis abnormality with a minimum certainty of % in a sample, all the four orthopaedists were confident that the examined sample was suffering from scoliosis. meanwhile, there are situations where the system cannot detect abnormalities with a high degree of accuracy. for instance, in the examination of patients, where scoliosis abnormality had been diagnosed with accuracy degrees between % and %, at least one out of four experts believed further physical examinations are required before reaching a final judgment. this condition is very similar to the real-world cases where an orthopaedist fails to make a definitive diagnosis and starts further investigations. at this stage, the health specialist may seek or ask for additional details and information. in cases where abnormalities were detected with certainty degree equal or below %, all experts agreed that the patient was healthy; nonetheless, that certainty value still reflected the risk of developing anomaly in the examined sample. knowing that experts are prone to errors, primarily due to prolonged activity, the purpose of this research was to create an expert system in the form of intelligent software to help experts diagnose spinal anomalies more accurately. evaluations show that the proposed approach was successful in detecting spinal anomalies and to assess the risk of anomaly development. conclusion and future studies this study proposed a novel, non-invasive method for detecting and preventing spinal anomalies. the expert system designed and implemented in this research can be used as a reliable assistant by medical experts, specially orthopaedists. this expert system helps specialists to validate the accuracy and integrity of results of spinal disorders diagnosis. it may also serve as an intelligent educational software for medical students to learn the process of spinal disorder diagnosis, and discover the symptoms and variables related to each spinal anomaly. by suggesting an optimal algorithm for the inference engine to emulate experts’ inference, the proposed method could successfully reason based on an optimal number of rules in a flexible question-answering process. it could also bring about advantages in comparison with the malfunctioning of sole use of conditional probabilities. in the conducted evaluation, the system was able to successfully detect anomalies in an individual by using anthropometric dimensions, everyday life habits, anatomy, and physiological data. the extensive use of this expert system by specialists may help a more accurate diagnosis and prevent the spread of spinal anomalies, leading to a healthier life. the present research may be further improved by using fuzzy inference or neural networks. besides, other factors of anthropometric dimensions the open bioinformatics journal, , volume dashti and dashti of healthy participants, anomalies, and medical information by researchers may be incorporated to yield better results. list of abbreviations ais = adolescent idiopathic scoliosis bci = backward chaining inference cf = certainty factor cnn = convolutional neural network ct = computed tomography ehr = electronic health records es = expert system ls = least squares mes = medical expert system ml = machine learning mr = magnetic resonance mri = magnetic resonance imaging rfi = requirement of further examinations svm = support vector machine tpd = true positive detection tnd = true negative detection author’s contribution seyed mohammad sadegh dashti and seyedeh fatemeh dashti conceived the presented idea, accomplished the process of knowledge acquisition, developed the inference algorithm and encoded the knowledge to the knowledge-base, designed and developed the user interface, supervised carrying the experiments out and discussed the results. seyed mohammad sadegh dashti developed the theoretical formalism and verified the analytical methods. . seyedeh fatemeh dashti performed the analytic calculations and the numerical simulations and wrote the whole manuscript with support from seyed mohammad sadegh dashti. consent for publication not applicable. funding none. conflicts of interests the authors declare that there is no conflict of interest regarding the publication of this article. acknowledgements authors would like to greatly appreciate dr. azita yazdani from the department of health information management, tehran university of medical sciences, for providing the medical records without which this research could not be evaluated. references jibril i, agajo j. development of a medical expert system for[ ] hypertensive patients diagnosis: a knowledge-based rules; advances in electrical and telecommunication engineering. aete . chaudhuri sb, rahman m. design of a medical expert system (mes)[ ] based on rough set theory for detection of cardiovascular diseases. progress in advanced computing and intelligent engineering. advances in intelligent systems and computing . [http://dx.doi.org/ . / - - - - _ ] ghasemi g;the effect of eight weeks of nasm exercises on sway[ ] back of high school female students. the scientific journal of rehabilitation medicine . heidari moghaddam r. reviewing the role of major thalassemia major[ ] in spinal abnormality development reviewing the role of major thalassemia major in spinal abnormality development . tolouei a. developing an expert system to detect blood cancer. j[ ] health manag . jibril i z, agajo j, ajao l a, kolo j g, inalegwu o c. development[ ] of a medical expert system for hypertensive patients diagnosis: a knowledge-based rules advances in electrical and telecommunication engineering ; elaine n. marieb;human anatomy & physiology. pearson education,[ ] inc . khoubi m. corrective and therapeutic exercises for treatment of spinal[ ] abnormalities. physical sciences publication . keshvari f, mirdarikvandi m. applicable corrective cxercises .[ ] daneshmandi m, sedaghati p. new corrective approaches to treatment[ ] of parkinson’s disease. boshra publications . machida m. neurological research in idiopathic scoliosis.[ ] pathogenesis of idiopathic scoliosis . [http://dx.doi.org/ . / - - - - _ ] mao sh, qian bp, shi b, zhu zz, qiu y. quantitative evaluation of[ ] the relationship between comp promoter methylation and the susceptibility and curve progression of adolescent idiopathic scoliosis. eur spine j ; ( ): - . [http://dx.doi.org/ . /s - - -y] [pmid: ] akazawa t, kotani t, sakuma t, et al. midlife changes of health-[ ] related quality of life in adolescent idiopathic scoliosis patients who underwent spinal fusion during adolescence. eur j orthop surg traumatol ; ( ): - . [http://dx.doi.org/ . /s - - - ] [pmid: ] nanjundappa s, harshavardhana md. results of bracing for juvenile[ ] idiopathic scoliosis. spine deform . brumitt j. core assessment and training. human kinetics .[ ] ilbeigi s, mehrshad n, afzalpour me, yousefi m. diagnosing spinal[ ] disorders by using markers mounted on spinous excrescences, sports medicine. . daneshmandi h, pourhosseini h. analyzing spinal disorders among[ ] boys and girls journal of motion . uei h, tokuhashi y, maseda m, nakahashi m, nakayama e. multiple[ ] vertebral fractures associated with glucocorticoid-induced osteoporosis treated with teriparatide followed by kyphosis correction fusion: a case report. osteoporos int ; ( ): - . [http://dx.doi.org/ . /s - - - ] [pmid: ] kobets aj, komlos d, houten jk. congenital cervical kyphosis in an[ ] infant with ehlers-danlos syndrome. childs nerv syst ; ( ): - . [http://dx.doi.org/ . /s - - - ] [pmid: ] schule r thomas c. segmental lumbar lordosis: manual versus[ ] computer-assisted measurement using seven different techniques. j spinal disord tech . cressey eric. strategies for correcting bad posture – part .[ ] ericcressey.com . robertson pa, armstrong wa, woods dl, rawlinson jj. lordosis[ ] re-creation in tlif and plif: a cadaveric study of the influence of surgical bone resection and cage angle. spine . [http://dx.doi.org/ . /brs. ] [pmid: ] sayari a, farahani a, ghanbarzadeh m. review and comparison of[ ] structural and aerobic corrective exercises affecting pulmonary function of students with kyphosis. olympics journal . ghafouri f. relation of kyphosis with depression and anxiety[ ] among athletic and non-athlete male students in tehran universities journal of research in sport science . mirbagheri r, ghasemi b. effect of eight weeks of massage therapy[ ] on lumbar lordosis second conference on applicable researches in sport science . mohammadi sh, mokhtarinia hr. investigating the effects of different[ ] working postures on cognitive performance. archives of rehabilitation . http://dx.doi.org/ . / - - - - _ http://dx.doi.org/ . / - - - - _ http://dx.doi.org/ . /s - - -y http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /s - - - http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /s - - - http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /s - - - http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /brs. http://www.ncbi.nlm.nih.gov/pubmed/ diagnose spinal disorders the open bioinformatics journal, , volume [http://dx.doi.org/ . /jrehab. . . ] choubineh a, moudi a. human-design and ergonomics. tehran:[ ] phisnet . smith s, kandel a. verification and validation of rule-based expert[ ] systems. crc press . [http://dx.doi.org/ . / ] ignizio james p. introduction to expert systems: the development and[ ] implementation of rule-based expert systems. . giarratano jc. gary riley; expert systems. pws publishing co .[ ] mohammadi motlagh ha, minaei bidgoli b. design and[ ] implementation of a web-based fuzzy expert system for diagnosing depressive disorder. appl intell . al-ajlan a. the comparison between forward and backward chaining.[ ] int j mach learn comput . [http://dx.doi.org/ . /ijmlc. .v . ] grosan crina, abraham ajith. rule-based expert systems in:[ ] intelligent systems. . [http://dx.doi.org/ . / - - - - _ ] caignya a, coussementa k. a new hybrid classification algorithm for[ ] customer churn prediction based on logistic regression and decision trees. eur j oper res . jose a. a prasad,design and development of a rule based expert[ ] system for aacr: a study of the application of artificial intelligence techniques in library and information field. saarbrücken, germany: vdm verlag . sagheb-tehrani m. expert systems development: some issues of[ ] design process. softw eng notes ; ( ): - . [http://dx.doi.org/ . / . ] shapiro ey. logic programs with uncertainties: a tool for[ ] implementing rule-based systems in: ijcai. . baudryab g, macharis c. range-based multi-actor multi-criteria[ ] analysis: a combined method of multi-actor multi-criteria analysis and monte carlo simulation to support participatory decision making under uncertainty. eur j oper res . fon gt, pitt mj, thies ac jr. thoracic kyphosis: range in normal[ ] subjects. ajr am j roentgenol ; ( ): - . [http://dx.doi.org/ . /ajr. . . ] [pmid: ] mahmoudi f, shahrjerdi s. changes in pain and kyphosis angle[ ] following a corrective exercise program in elderly women: a randomized controlled trial. jrums . kado dm, prenovost k, crandall c. narrative review: hyperkyphosis[ ] in older persons. ann intern med ; ( ): - . [http://dx.doi.org/ . / - - - - - ] [pmid: ] chen jh, asch sm. machine learning and prediction in[ ] medicine—beyond the peak of inflated expectations. n engl j med . [http://dx.doi.org/ . /nejmp ] kotti m, duffell ld, faisal aa, mcgregor ah. detecting knee[ ] osteoarthritis and its discriminating parameters using random forests. med eng phys ; : - . [http://dx.doi.org/ . /j.medengphy. . . ] [pmid: ] madelin g, poidevin f, makrymallis a, regatte rr. classification of[ ] sodium mri data of cartilage using machine learning. magn reson med ; ( ): - . [http://dx.doi.org/ . /mrm. ] [pmid: ] mirzaalian h, wels m, heimann t, kelm bm, suehling m. fast and[ ] robust d vertebra segmentation using statistical shape models in th annual international conference of the ieee engineering in medicine and biology society (embc). [http://dx.doi.org/ . /embc. . ] ashinsky bg, coletta ce, bouhrara m, et al. machine learning[ ] classification of oarsi-scored human articular cartilage using magnetic resonance imaging. osteoarthritis cartilage ; ( ): - . [http://dx.doi.org/ . /j.joca. . . ] [pmid: ] ashinsky bg, bouhrara m, coletta ce, et al. predicting early[ ] symptomatic osteoarthritis in the human knee using machine learning classification of magnetic resonance images from the osteoarthritis initiative. j orthop res ; ( ): - . [http://dx.doi.org/ . /jor. ] [pmid: ] shamir l, orlov n, eckley dm, macura t, johnston j, goldberg ig.[ ] wndchrm - an open source utility for biological image analysis. source code biol med ; : . [http://dx.doi.org/ . / - - - ] [pmid: ] yu s, tan kk, sng bl, li s, sia at. lumbar ultrasound image[ ] feature extraction and classification with support vector machine. ultrasound med biol ; ( ): - . [http://dx.doi.org/ . /j.ultrasmedbio. . . ] [pmid: ] adankon mm, dansereau j, labelle h, cheriet f. non invasive[ ] classification system of scoliosis curve types using least-squares support vector machines. artif intell med ; ( ): - . [http://dx.doi.org/ . /j.artmed. . . ] [pmid: ] oktay ab, albayrak nb, akgul ys. computer aided diagnosis of[ ] degenerative intervertebral disc diseases from lumbar mr images. comput med imaging graph ; ( ): - . [http://dx.doi.org/ . /j.compmedimag. . . ] [pmid: ] kadoury s, mandel w, roy-beaudry m, nault ml, parent s. -d[ ] morphology prediction of progressive spinal deformities from probabilistic modeling of discriminant manifolds. ieee trans med imaging ; ( ): - . [http://dx.doi.org/ . /tmi. . ] [pmid: ] thong w, parent s, wu j, aubin ce, labelle h, kadoury s. three-[ ] dimensional morphology study of surgical adolescent idiopathic scoliosis patient from encoded geometric models. eur spine j ; ( ): - . [http://dx.doi.org/ . /s - - - ] [pmid: ] pesteie m, abolmaesumi p, ashab ha, et al. real-time ultrasound[ ] image classification for spine anesthesia using local directional hadamard features. int j cars ; ( ): - . [http://dx.doi.org/ . /s - - - ] [pmid: ] hetherington j, lessoway v, gunka v, abolmaesumi p, rohling r.[ ] slide: automatic spine level identification system using a deep convolutional neural network. int j cars ; ( ): - . [http://dx.doi.org/ . /s - - - ] [pmid: ] beauchamp kg. applications of walsh and related functions: with an[ ] introduction to sequency theory. academic press . forsberg d, sjöblom e, sunshine jl. detection and labeling of[ ] vertebrae in mr images using deep learning with clinical annotations as training data. j digit imaging ; ( ): - . [http://dx.doi.org/ . /s - - -x] [pmid: ] obermeyer z, emanuel ej. predicting the future—big data, machine[ ] learning, and clinical medicine. n engl j med ; ( ): - . [http://dx.doi.org/ . /nejmp ] [pmid: ] abu-nasser b. medical expert systems survey; international journal[ ] of engineering and information systems. ijeais . deng y, groll mj, denecke k. rule-based cervical spine defect[ ] classification using medical narratives. stud health technol inform ; : . [pmid: ] pavlovic-veselinovic s, hedge a, veselinovic m. an ergonomic[ ] expert system for risk assessment of work-related musculo-skeletal disorders. int j ind ergon . [http://dx.doi.org/ . /j.ergon. . . ] deng y, denecke k. patient records retrieval system for integrated[ ] care in treatment of cervical spine defect in vldb workshop on data management and analytics for medicine and healthcare. basu s, plewczynski d, saha s, et al. dspan: semiautomated -d[ ] segmentation, classification and analysis of hippocampal dendritic spine plasticity. bioinformatics ; ( ): - . [http://dx.doi.org/ . /bioinformatics/btw ] [pmid: ] naser ss. almursheidi sh;a knowledge based system for neck[ ] pain diagnosis; world wide journal of multidisciplinary research and development. wwjmrd . © dashti and dashti. this is an open access article distributed under the terms of the creative commons attribution . international public license (cc-by . ), a copy of which is available at: (https://creativecommons.org/licenses/by/ . /legalcode). this license permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. http://dx.doi.org/ . /jrehab. . . http://dx.doi.org/ . / http://dx.doi.org/ . /ijmlc. .v . http://dx.doi.org/ . / - - - - _ http://dx.doi.org/ . / . http://dx.doi.org/ . /ajr. . . http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . / - - - - - http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /nejmp http://dx.doi.org/ . /j.medengphy. . . http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /mrm. http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /embc. . http://dx.doi.org/ . /j.joca. . . http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /jor. http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . / - - - http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /j.ultrasmedbio. . . http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /j.artmed. . . http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /j.compmedimag. . . http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /tmi. . http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /s - - - http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /s - - - http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /s - - - http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /s - - -x http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /nejmp http://www.ncbi.nlm.nih.gov/pubmed/ http://www.ncbi.nlm.nih.gov/pubmed/ http://dx.doi.org/ . /j.ergon. . . http://dx.doi.org/ . /bioinformatics/btw http://www.ncbi.nlm.nih.gov/pubmed/ https://creativecommons.org/licenses/by/ . /legalcode an expert system to diagnose spinal disorders [objective:] objective: methods: results: conclusion: . background . rationale . basic elements . . rule-based expert system . . inference in rule-based expert systems . . backward chaining inference . . certainty factor . literature review . . methods based on machine learning . . rule-based approaches . knowledge engineering method . . knowledge acquisition . . the proposed algorithm of inference engine . . implementation of knowledge base and user interface . . diagnosis in practice . results . . phase : evaluation based on real-world samples . . phase : evaluations based on medical records . . cut-off values in the evaluation procedure . discussion conclusion and future studies list of abbreviations author’s contribution consent for publication funding conflicts of interests acknowledgements references material recognition using tactile sensing emmett kerr, tm mcginnity and sonya colemana,b,a aintelligent systems research centre, university of ulster, magee campus, northern ireland. e-mail: ep.kerr@ulster.ac.uk, sa.coleman@ulster.ac.uk bcollege of science and technology, nottingham trent university, uk. e-mail: martin.mcginnity@ntu.ac.uk abstract identification of the material from which an object is made is of significant value for effective robotic grasping and manipulation. characteristics of the material can be retrieved using different sensory modalities: vision based, tactile based or sound based. compressibility, surface texture and thermal properties can each be retrieved from physical contact with an object using tactile sensors. this paper presents a method for collecting data using a biomimetic fingertip in contact with various materials and then using these data to classify the materials both individually and into groups of their type. following acquisition of data, principal component analysis (pca) is used to extract features. these features are used to train seven different classifiers and hybrid structures of these classifiers for comparison. for all materials, the artificial systems were evaluated against each other, compared with human performance and were all found to outperform human participants’ average performance. these results highlighted the sensitive nature of the biotac sensors and pave the way for research that requires a sensitive and accurate approach such as vital signs monitoring using robotic systems. keywords: tactile sensing, signal processing, neural networks, artificial intelligence, classification, material classification . introduction humans can quickly gain information about properties of an object or material simply by viewing them from different angles; for example they can estimate how it might feel to touch or how heavy it may be. this is due to our highly sophisticated visual capabilities and ability to adapt prior knowledge preprint submitted to expert systems with applications october , learned from similarly shaped and known objects. however, there are some properties which are difficult to detect by vision alone, for example thermal conductivity, compressibility or a determination from what material an ob- ject is made. in order to learn these characteristics and remember them for future interactions, physical manipulation of an object is required. distin- guishing between objects and materials of different compressibility, tempera- ture and texture can be achieved by performing complex manipulation tasks such as squeezing or rubbing. these complex manipulation tasks can inher- ently be performed extremely well by humans due to our very sophisticated tactile perception. in order to produce the tactual perceptions required to learn about ob- ject properties, humans perform various types of movements when interacting with an object. experimental psychologists have identified six general types of exploratory movements: pressure to determine hardness, static contact to determine thermal properties, lateral sliding movements to determine surface texture, enclosure to determine global shape and volume, lifting to deter- mine weight, and contour following to determine exact shape (lederman and klatzky, ). humans can complete these exploratory movements very fast and rapidly evaluate the object leading to a possible identification. one of the main applications of an artificial tactile sensing system is expected to be in robotic applications. in order to retrieve the necessary tactual perceptions, a robot must explore an object in a similar manner to that of a human. however, in comparison to humans, these exploratory movements can be slow for a robot and typically involve the evaluation of a great volume of acquired data, thus being computationally expensive. it would be a useful skill for a robot if it could complete a preliminary evaluation of the object by quickly identifying the physical nature of the material (e.g. metal, wood, plastic etc.) through completing a small number of initial basic actions, thus removing the need for extensive manipulation. these initial basic actions could include the retrieval of thermal information upon initial contact with a material which may provide an indication of its physical nature, e.g. wood, metal or plastic. furthermore the texture of a material can be identified by sliding a robotic hand or finger along the material, i.e. determining if it is rough or smooth. we use an artificial fingertip to acquire data relating to the thermal prop- erties and surface texture of materials, that are first analysed to initially identify to which group the material belongs (e.g. wood, metal, plastic etc.) and subsequently to determine the specific individual material (e.g. alu- minium, copper, pine etc.). building on the work presented in (kerr et al., , a) and (kerr et al., b) the contribution of this work is the im- plementation of six further classifiers, hybrid configurations of the classifiers and a comparison of the performances of these classifiers. the classifiers implemented and compared are a one stage ann (kerr et al., a), a two stage ann (kerr et al., b), a one stage svm, a two stage svm, a hybrid ann and svm approach, gmm, lda, nb, k -nn and mlp. utili- sation of these techniques to classify materials from tactile data resulted in an increase in classification accuracy and system efficiency in comparison to previous work (kerr et al., b) resulting in an approach capable of a fast initial identification of the material group and/ or individual material. these classifiers are also evaluated against human performance when identifying the same test materials as reported in (kerr et al., b). therefore, in this paper we propose an approach for material classification using a combination of two of the three key characteristics, namely surface texture (represented through vibration data) and thermal characteristics. although one weakness of the approach is that it does not achieve accuracies as high as that of (xu et al., ), the strengths include the fact that it is computationally efficient and it demonstrates high classification accuracies across materials, some of which are very similar, unlike many methods presented in the literature where the test materials can be very different. a range of common classifiers together with novel combinations of classifiers are evaluated and proven to be accurate and efficient algorithms for material classification. furthermore, the system is tested with previously unlearned materials in order to assess its robustness. the remainder of this paper is organised as follows: section presents an overview of related research in tactile sensing-based material classifica- tion. section describes the experimental set-up, including an overview of the equipment used and the learning algorithms used for the robot system. section presents an evaluation of the performance of the various artifi- cial systems developed and compares their performance to that of human subjects. section discusses the artificial system and human participants’ performance. conclusions and plans for future work are given in section . . background and related research three of the key characteristics that are critical in performing an effi- cient grasp on an object are texture, compressibility and thermal properties. such characteristics can be obtained from tactile sensing. materials from dif- ferent classes may have similar compressibility; for example neither acrylic or medium density fibreboard (mdf) are very compressible but are from different classes of materials, i.e. plastic and wood. drimus et al. ( ), de- veloped a sensor and presented a technique used to classify between ten rigid and deformable household objects. the objects were squeezed in a gripper fitted with force sensors and a k -nn classification algorithm was utilised to classify the different objects, based on their compressibility. classification accuracy of up to % was achieved for classification between the ten objects ranging from a rubber ball to a plastic bottle. the algorithm struggled to distinguish between similarly compressible objects even though the objects were very different, for example a “bad orange” and an empty bottle were found to have similar compressibility and therefore the algorithm failed to classify these objects correctly. another approach which uses images de- rived from collected tactile sensor data is that of pezzementi et al. ( ), which utilises computer vision algorithms to allow tactile sensor readings, representing compressibility, to be viewed as images. the tactile array sen- sor explores the object images representing contact made on the array are produced. six feature descriptors including the well known scale invariant feature transform (sift) (lowe, ) and mr- (varma and zisserman, ) are used to extract features from these images. the bag-of-words (bow) approach (jurie and triggs, ), commonly used in computer vi- sion methods, is then used to learn and classify the objects based on these features. luo et al. ( ), also present an approach which utilises tactile images and a bow approach to object recognition. an optimised bow ap- proach and dictionary were produced by considering different outputs of the classic bow with different dictionary sizes for different views of objects. the classification performance of this optimised approach was compared against the performance of the classic bow model utilising objects. it was found that the proposed approach outperformed the classic bow model particularly for classification of objects with similar features. exploiting other characteristics such as the object’s surface texture or its thermal properties may help to distinguish between very different ma- terials which have similar compressibility, leading to improved classification accuracy. related literature reports methods that are capable of classifying materials using surface textile and achieving high performance rates, i.e. is the object rough or smooth? fishel and loeb ( ) used the biotac sensors with a bayesian exploration method for the intelligent identification of tex- tures. classification from a set of textures was achieved. this approach related the textures to human style descriptions frequently used in psy- chophysical literature exploring texture discrimination (i.e. sticky/slippery, rough/smooth). the focus was to find the most useful exploratory motions on the surface of the material, relating to six general types of exploratory movements that humans make when tactually exploring objects, as outlined by experimental psychologists lederman and klatzky ( ). ho et al. ( ) assessed the ability of their sensors to classify between three different ma- terials (denim, a photo and tape) based on their surface texture. classifi- cation accuracy of % was achieved between two very different materials, i.e. denim and a photo. however, the method struggled to distinguish ac- curately between two similar materials, namely the photo and tape. jamali and sammut ( ) propose a method which used an artificial silicon tactile sensor with polyvinylidene fluoride (pvdf) films to slide along materials and collect vibration data. a total of seven materials were tested includ- ing sponge, carpet, wood, two tiles of different roughness and two pieces of vinyl of different roughness. a bayes classifier was trained with the fourier coefficients of the sensor output data retrieved from the sliding motion over each material. it was shown that the authors’ classifier performed well for classification between dissimilar surface textures, however, their approach was unable to distinguish between two types of tiles, as the texture of the surfaces were similar. jamali and sammut ( ) extended this work to eval- uate multiple classifiers and added a majority voting algorithm. in a similar way to their previous work, data were retrieved from the fingertip making contact with test materials. the test materials included the same seven as in their previous work with the addition of a second type of carpet (differ- ent roughness), making eight material surfaces in total. whilst the fingertip was being rubbed over each surface, different textures induced different in- tensities of vibration in the silicon fingertip. as in their previous work, the data from the fingertip were pre-processed and the fourier coefficients of the sensor outputs were used as inputs to various classifiers that were imple- mented in the weka machine learning toolkit (hall et al., ). it was found that the best performance of %± % was achieved using the naive bayes tree (nbtree). the system struggled most when trying to classify between materials of similar roughness. the approach proposed by chathu- ranga et al. ( ) is another example of a system which produced promising results but still struggled to classify between similar material surfaces. the authors developed a biomimetic fingertip and used it to discriminate between seven different materials based on surface texture. using anns as classifiers, accuracies of up to % were achieved. the thermal characteristics of a material is another modality that can be assessed and used to identify a material/object. although one can generally tell the difference between materials that feel hot and cold, one would struggle to discriminate between similar materials on the basis of their actual temper- atures. what human fingers detect thermally is not the absolute temperature of the material alone, but the effect of thermal conductivity and diffusivity. this is due to the fact that the human finger is at a consistent temperature, approximately ◦ celsius, and therefore will be different from the ambient temperature of most objects encountered. the nerves of the finger detect the flow of heat to/from the source. this suggests that human temperature sens- ing may be emulated by thermal conduction sensing (monkman and taylor, ). an approach which analyses the thermal properties of various ma- terials is described in (kerr et al., ). the materials are classified into initial groupings and further identified individually via the data gathered from test materials using the syntouch r© biotactm sensor. pca was ap- plied to identify relevant features from the static temperature and thermal conductivity data that were used to train an ann. the system was compared with humans and was found to outperform human performance. a success rate of % was obtained by the system when classifying the material group (i.e. plastic, wood, metal etc.) and % was achieved when identifying the individual material. when similar experiments were conducted using hu- mans, results showed classification rates of % for group identification and % for individual material identification. however, even though the system achieves better performance than the humans, the classification rates are poor which is potentially caused by the fact that only thermal conductivity data are being used. additionally, the system is computationally expensive due to the large number of input features necessary to accurately represent the data and the substantial number of neurons in the hidden layer. bhat- tacharjee et al. ( ) present another approach that focusses on material classification. their approach uses a tactile sensor that is capable of measur- ing solely thermal properties and they focus on classification within a short period of time, for instances where longer term contact is not viable. three classification algorithms are considered, svm+pca, k -nn+pca and hid- den markov models (hmm). it was found that svm+pca performed best with classification rates of % and up to % with longer contact times across materials with varying initial conditions and -fold cross valida- tion. although the classification rates achieved are impressive, the materials tested varied widely in thermal conductivity and the purpose of the sensor is solely to determine thermal properties; it therefore cannot measure material properties such as texture or compressibility, thus limiting the applications for which it can be used. rather than classifying materials by analysing just one property, a com- bination of compressibility, texture and/ or thermal properties can be used to achieve higher rates of classification accuracy. kerr et al. ( a), used a combination of two of these properties, extending their work in (kerr et al., ) by considering surface texture (vibration) as an additional modality alongside the thermal properties of a material. additionally, the authors re- designed the system to improve efficiency by selecting a reduced number of principal components, reducing the number of neurons in the hidden layer and hence training the system faster than previously. the one stage ann system in (kerr et al., a) obtained a classification rate of . % for ma- terial group classification and . % for individual material identification. conversely, the humans achieved classification rates of . % and . % respectively. a two stage ann approach to the same material classification experiment is presented in (kerr et al., b). the output from the network in the first stage is used as an input to the second stage, which is a set of networks, one for each material group. this second stage is used to identify each individual material, within a specified group. despite not being as effec- tive as the one stage ann method presented in (kerr et al., a), this two stage approach obtained a classification rate of . % for material group identification and . % for individual material identification. bhattachar- jee et al. ( ) continued their earlier research and presented a method for distinguishing between contact with inanimate objects and humans, using a novel portable handheld device developed by the authors consisting of three tactile sensing modalities: a force sensor to detect contact, a heat-transfer sensor that is actively heated, and a small thermally-isolated temperature sensor (wade et al., ). data was collected from the arms of human subjects from different locations on the right arm and objects consisting of similar objects from different bathrooms (bhattacharjee et al., ). the effect of varying durations of contact made with the human subjects and the objects was also evaluated. a svm was used to learn to classify between an object and human in the first instance and it achieved an aver- age classification accuracy of . % when contact was held for . seconds, . % for . second of contact and . % for . seconds of contact. high classification accuracies were also achieved when generalising to new contact locations within the same bathroom with an average of . % for . sec- onds of contact, . % for . second and . % for . seconds. however, when generalising to new environments, i.e. similar objects in different bath- rooms, accuracies were lower achieving . % for . seconds of contact, . % for . second and . % for . seconds. like many other meth- ods it struggled to identify similar objects especially when the environment changes. hoelscher et al. ( ) use a biotac fingertip sensor to collect several features from a range of objects to develop a system capable of identifying material and recognising objects. the objects are made of mate- rials including plastic, metal, stone ceramic, paper etc. they collect data by making static and lateral contact on each material with the biotac sensor. the authors then compare seven different methods of extracting features from their processed data, including temporal data, pca reduced dimensionality raw data, pressure features, electrode features, physically motivated features, temperature features and mean features. the authors compare two genera- tive (nb and gaussian) and two discriminative (svm and random forests) classifiers to evaluate identification of the materials and objects from the collected data. the best performing method consisted of an svm classifier with dimensionality-reduced mean values of filtered data, achieving a classi- fication accuracy of material and object identification of over %. xu et al. ( ) presented an algorithm which considered three key properties of the material to enable classification; compliance (compressibility), texture and thermal conductivity. to collect data for these three key properties a bio- tac fingertip was mounted on a shadow robot hand. based on a bayesian exploration approach in the authors previous work (fishel and loeb, ) three exploratory movements were selected and used to explore ten test ma- terials. a classification accuracy of % using materials was achieved with the only failure being that a damp sponge was classified as a feather. the approach in (xu et al., ) may have obtained % classification, however this was while assessing substantially different materials, varying from a brick to a light feather. therefore, the materials have very different proper- ties allowing for more straight forward classification. furthermore, all three tactile properties were used making the system both slow and computation- ally expensive. it may be concluded that using just one material property does not allow for desirable classification of materials and combining modalities can lead to high computation cost. the ideal scenario would be to design a system that can perform classification with accuracy similar to that of (xu et al., ), using similar materials, with reduced computational burden and hence run in real-time (or close to). an overview of the best performing machine learning methods for the application of material identification using tactile sensing can been seen in table . it demonstrates that the biggest difficulty in classification of materials is between similar materials or materials of similar roughness. although, this is to be expected, it is still an ongoing challenge within the field. . methodology . . experimental set-up . . . tactile sensor the biotac fingertip is a tactile sensor that is shaped like a human fin- gertip and is filled with an incompressible liquid, giving it similar compliance to a human fingertip (lin et al., ; kerr et al., ). the sensor is ca- pable of detecting a full range of information: micro vibrations, forces and temperature, similar to a human finger. figure shows a cross section view of the biotac fingertip. figure : cross section view of biotac fingertip tactile sensor (syntouch, ) there are various sensors on the fingertip that enable retrieval of the aforementioned sensory information. an array of electrodes measures the force applied on the fingertip by reading the impedance between each electrode and four common excitation electrodes. as the impedance over the sensing electrode increases, the measured voltage decreases, demonstrating table : table comparing tactile sensing-based material identification methods research group machine learning method sensor type accuracy achieved (%) description xu et al. ( ) probability matrices multi-modal tactile sensor (biotac) . very different materials were used. continuous exploration and three modalities required therefore high computational costs. hoelscher et al. ( ) svm multi-modal tactile sensor (biotac) . struggled to distinguish between similar materials. required training from all materials in the group drimus et al. ( ) knn single mode piezoresistive tactile sensor array . classification across just objects. struggled to identify between similar objects. chathuranga et al. ( ) ann multi-modal tactile sensor . classification across just materials, of which were fabrics. ho et al. ( ) ann single mode electro-conductive soft skin sensors . reported to be a very computationally expensive approach. jamali and sammut ( ) nbtree single mode silicon tactile sensor with pvdf films . struggled to distinguish between materials of similar roughness. bhattacharjee et al. ( ) svm single mode thermistor-based tactile sensor . longer contact with materials required for maximum accuracy and test materials varied widely bhattacharjee et al. ( ) svm multi-modal hand-held tactile sensor . classification accuracies achieved across already learned data, accuracy drops to . % when tested against new data from the same materials. contact in that area of the finger. having a electrode array allows the exact point of contact on the fingertip to be derived. furthermore, it enables multiple points of contact to be derived at any instance. inside the core of the fingertip there is a pressure sensor located in the rigid core. this sensor reads pressure changes in the conductive fluid located between the hard core and the elastomeric skin. the vibration sensor generated two values, a dc pressure signal (pdc) that is obtained using a low pass filter and an ac pressure signal (pac) that is obtained using a band pass filter. the pac data represent the vibration strength recorded when the biotac finger makes contact with materials or objects. static temperature (tdc) and thermal flow (tac) data can also be collected from using the biotac sensor as there is a thermistor at the tip of the finger. tac data represent the rate that heat is transferring from the fingertip to an object. this is made possible by the thermistor which is located in the tip of the fingertip. using the biotac fingertip, two actions were performed on test materi- als to classify them. the first action was a press action where the fingertip was pressed into the material at a constant force. the second action was a slide action where the fingertip was dragged across the surface of the ma- terial at a constant force. both actions aim to replicate two of the actions outlined by experimental psychologists lederman and klatzky ( ), that a human completes when first investigating an unknown material, namely static contact to determine thermal properties and lateral sliding movements to determine surface texture. the press action enables collection of the tac and tdc data whereas the slide action enables the collection of pdc, tdc, pac and tac; these data are used for classification in section . . . . . materials classified consistent with previous work (kerr et al., a,b), we use mate- rials with the slide and press actions. we have intentionally selected some materials to be similar to determine if the approach can readily classify the material type and then identify the specific material. the materials are acrylic (ac), rough acrylic (acr), copper (cr), aluminium (al), rough copper (crr), rough aluminium (alr), redbrick (r), glossy cardboard (gc), plain cardboard (pc), soft foam (s), carpet (ct), doormat (d), mdf (m) and pine (p). figure presents pictures of each of the materials used. analogous with (kerr et al., a,b), the materials used in the experi- ments were divided into six groups corresponding to the six different material figure : samples of the materials used in the experimental set-up. table : table showing the individual materials and the groups to which they belong. material group acrylic (ac) plastic (p) rough acrylic (acr) plastic (p) copper (cr) metal (me) aluminium (al) metal (me) rough copper (crr) metal (me) rough aluminium (alr) metal (me) redbrick (r) masonry (ma) glossy cardboard (gc) cardboard (c) plain cardboard (pc) cardboard (c) soft foam (s) fabrics (f) carpet (ct) fabrics (f) doormat (d) fabrics (f) mdf (m) wood (w) pine (p) wood (w) types. the individual materials and the groups to which they belong are shown in table . the artificial system was then evaluated on its ability to classify each material individually. . . data collection analogous with work carried out in (kerr et al., ; xu et al., ), the biotac fingertip was powered on and left to rest for - minutes to allow it to reach its steady state temperature (approximately ◦c, ◦c above ambient). a rig including a motorised arm and turntable was designed and built for the experiment. two stepper motors (one for the arm and one for the turntable) were connected to an arduino unotm board and a graphical user interface (gui) was developed using python to control the motors. an image of the experimental rig and gui can be seen in figure (a) and (b) respectively. the fingertip was placed at the end of the motorised arm so that it could be moved towards or away from the material. additionally, the material was placed on the motorised turntable to allow for the material to be moved below the finger to replicate a sliding action. to collect data corresponding to a thermal exploratory movement, the fingertip is pressed onto a material using a constant force of n, measured (a) (b) figure : a) image showing the experimental rig; b) screen shot of the developed gui. by an ati nano -axis f/t sensor. to allow time for the heat flow to stabilise, it is necessary to collect seconds of data for the press action. for the slide action, the finger tip was pressed on the material using a constant force of . n (also measured with the aforementioned force sensor), and moved along the material surface for approximately cm. the biotac data are recorded at hz and in a sequence, with the pac values and the individual electrode values alternating (i.e. pac data are recorded at hz). all other values such as tdc (static temperature), tac (thermal flow rate) and pdc (static vibration) are given at the end of each sequence of values. an example of this sequence can be seen in figure . all values are sampled with -bit resolution. these values are extracted and individual datasets for pac, pdc, tac and tdc data are obtained using matlab. figure : diagram showing the sequence of data from the biotac fingertip . . pre-processing to obtain high quality thermal conductivity (tac and tdc) time series data, we use only the data collected after the contact force has reached its maximum, this was found to occur approximately seconds after contact. the tac and tdc datasets, collected during the press action, each con- tain data points. additionally a set of data points for each of tac, tdc, pac and pdc is obtained during the slide action. in line with other work presented in the literature, initial evaluation of using three modalities (i.e. thermal properties, vibration data and impedance values to represent compressibility) was conducted. it was hypothesised that due to the vast quantity of electrode data collected from the biotac fingertip, the process- ing time would increase greatly when identifying materials. initial visual analysis of the electrode values (voltage related to impedance) and initial ex- perimentation was conducted to quantify the additional time it would take to include analysis of compressibility when classifying the materials in this experiment. as demonstrated in figure , it is clear that all materials be- have relatively similar during the press action with the exception of softer and rougher materials such as the softfoam and doormat. however, these dif- ferences are also clearly identified when analysing the vibration data (kerr et al., a) and therefore it was felt that any additional benefit from in- cluding compressibility data in this experiment was minimal to none. in fact, preliminary results indicated a decrease in accuracy when the electrode data were considered as part of the algorithm. furthermore, the computation time to include the compressibility data was analysed. it took approximately times longer to train one fold of a svm network when the vast compress- ibility data was included in comparison to when only thermal and vibration data are considered. therefore, considering the decrease in accuracy in pre- liminary investigations with compressibility data being considered, together with the increase in computation time, we use only thermal and vibration data in the experiments in this paper. initially we trained the classifiers with raw tac, tdc, pac and pdc data collected directly from the biotac sensor, however these resulted in poor classification accuracy such as % training accuracy for the ann with % testing accuracy. hence, to improve the characteristics of the data we applied principal component analysis (pca) to reduce the dimensionality of the inputs. we applied pca using both eigen-decomposition of the sample’s covariance and singular value decomposition (svd). with the sensor data, using the eigen-decomposition based approach generated the best inputs for the classifiers, evidenced by the training and testing accuracies over multiple experiments. hence, the presented approaches and results use pca with eigen-decomposition. in (kerr et al., ), pca was applied to a combined set of tac and tdc data ( values), however in this paper we apply pca to each individual modality (pac, pdc, tac, tdc) first. in this way, the principal components matrix for each material is formed from dif- ferent modalities tailored for either the press or slide action experiment. as described in section . , the principal components matrices are used to train the numerous classifiers and hybrid combinations of classifiers evaluated. figure : graph showing initial analysis of compressibility data from the biotac fin- gertip . . classifiers the data collected from the fingertip during the press and slide action ex- periments were normalised and pca using eigen-decomposition was applied to extract features. the optimal number of principal components (pcs) for each dataset was determined empirically as three. hence in total there are inputs to a classifier (six modalities and three pcs per modality, -bit values) for each experiment. to obtain the classification accuracy of each classifier in each experiment, the classified test data form an output ma- trix which is compared with a matrix containing the actual material labels ( classes for classification of materials into the groups and classes for classification of the individual materials). average classification accuracy is obtained using fold cross validation. the matlab package (matlab, ) and r package (r core team, ) were used to implement the range of classifiers. each approach used for classification is described in more detail in section . . to section . . . . . . one stage support vector machine (svm) support vector machine is a useful and well established technique used in machine learning to learn and classify data. the main objective of a svm is to separate a set of inputs by identifying a line, or hyperplane in n-dimensional space. each of the input points is known to belong to one of two classes. if it is possible for the algorithm to find a suitable hyperplane that will minimise future misclassification, then the data are easily classified based on what side of the hyperplane they reside. however, it is common that it is not possible to linearly separate the data. in this instance the svm algorithm is required to map the input data to a higher dimensional space through the use of a kernel function (kf). although there are many types of kfs, five are evaluated in this work and their performances compared. within each kf there are various parameters that can be changed to optimise the performance of the classifier. one common parameter is the soft margin (sometimes known as the box constraint ), c. the purpose of the soft margin is to modify the svm algorithm to tolerate training errors by intuitively tolerating a few outliers on the wrong side of the hyperplane. deciding on the size of the margin requires parameter tuning and optimisation. the kfs evaluated for the classification of materials and their respective parameters that were optimised are: • linear kf, quadratic kf and multilayer perceptron (mlp) kf: soft margin value (c) is optimised. • polynomial kf: two parameters are optimised in this kf; the soft mar- gin value (c) and the order of polynomial (p). various combinations of values were evaluated in order to find the best performing combination of parameters. • radial basis function (rbf) kf: two parameters are optimised in this kf; the soft margin value (c) and the gamma parameter for the gaussian rbf (γ). the gamma parameter is a positive number rep- resenting the scaling factor of the rbf. various combinations of values were evaluated in order to find the best performing combination of parameters. svm is more commonly used as a binary classifier. however, binary classification is not suitable for the material classification task outlined in this work; the standard svm algorithm has been adapted to enable multi- classification for material classification. two common ways of utilising binary svm classification techniques to allow for multi-class classification are by either having a sequence of numerous one-versus-one binary classifiers or by having a multiple one-versus-all classifier. the approach utilised in this work is a series of one-versus-all classifiers in which the system is trained with each class classified against the samples of all the other classes, to develop a svm multi-class classifier. two separate experiments were conducted: one for the classification of the materials into their six respective groups and one for the classification of the fourteen individual materials. all kfs were evaluated and their parameters optimised for both experiments. for each optimised kf implemented and evaluated, three different methods for calculating the separating hyperplane were also implemented and evaluated. the three standard methods evalu- ated were sequential minimal optimization (smo), quadratic programming (qp) and least-squares (ls) method. table outlines the best performance achieved from each kf for material group classification, the best performing method for calculating the separating hyperplane and the respective opti- mised parameters. it can be seen that the two best performing kfs with a classification rate of . % are the quadratic kf and the polynomial kf. although the two kfs performed equally well in terms of classification, the polynomial kf was found to be marginally more efficient than the quadratic kf at achieving the highest classification rate. the quadratic kf took . seconds to train for one fold of data whereas the polynomial kf took . seconds. therefore, the polynomial kf with a soft-margin value of . and a polynomial order of was used in the svm implemented in this approach. table : table comparing the svm kf classification accuracies. kernel function best performing method optimised parameters classification accuracy linear qp c = . . % quadratic smo c = . . % mlp ls c = . . % polynomial smo c = . p = . . % rbf qp c = . γ = . . % the svm system was used for both experiments with the only difference being the number of groups/ classes (g) that the svm had to classify. for the first experiment, i.e. classifying the materials into their six groups, g = and for the second experiment involving individual materials, g = as described in section . . . . . . two stage svm in contrast to having two separate experiments for the classification of the materials into their groups and the classification of individual materials, as described in section . . , this section explains a two stage approach to material classification. this approach utilises the multi-class svm classifica- tion algorithm explained in section . . to firstly classify the materials into their groups and then uses this information to further classify the materials individually using a mixture of binary and multi-class svm classification in the second stage. the same svm design as outlined in section . . is used in the first stage of this two stage approach. for the second stage, svm systems are trained, one for each material group, some of which have different numbers of outputs/ groups for the individual materials within them, requiring the need for binary classification in some cases and multi-class classification in others. these svm systems are saved and later used for testing the data input at the second stage. similar to the development of the svm in the one stage approach, five different kfs with a range of parameters and hyperplane calculation methods were evaluated for the second stage svm classification system. the three equally best performing kfs in the second stage were the quadratic kf, the polynomial kf and the rbf kf, with the most efficient proving to be the polynomial kf. therefore, the polynomial kf was used in the second stage svm. upon completion of material group classification in the first stage, the results of the classification are used as the inputs to the second stage. a model of the two stage svm approach can be seen in figure . if a material was correctly classified into a specific group in the first stage then this activates the second svm to classify the individual material within this material group. however, if the material group is incorrectly classified in the first stage then it is recorded as an incorrect classification in the final output matrix and is not classified in the second stage to improve efficiency. likewise, if the material is classified incorrectly in the second stage in comparison with ground truth (i.e. it has been correctly classified figure : diagram showing the -stage svm approach used for material classification in its group but incorrectly classified individually), then it is recorded as an incorrect classification in the output matrix. . . . one stage ann we use a back propagation ann as illustrated in figure . during train- ing, the number of neurons in the hidden layer and the number of training epochs were varied and evaluated. this determined that the ideal structure for the ann contained neurons and required epochs for training. the network structure illustrated in figure was applied in both exper- iments differing numbers of outputs (n). in the first experiment, n = as there are six material groups and in the second experiment n = corre- sponding to fourteen individual test materials, as described in section . . . . . . two stage ann the two stage approach (kerr et al., b), used a series of back prop- agation anns, each with one hidden layer, where the output of the first ann is the input of the second to allow for group classification first followed by individual classification. this differs from the one stage ann approach outlined section . . as the two networks work in series rather than two individual networks, one classifying materials into groups and one classify- ing individual materials. the two stage approach structure can be seen in figure . the anns used in the two stage approach are similar in structure to the ann used in the one stage approach explained in section . . . the first ann is used to classify the materials into groups and the second stage of the ann is comprised of six individual anns for each material group (i.e. plastic, metal, masonry, fabrics, paper and wood). these anns were all trained individually with their respective outputs relating to how many materials there are in each group, for example plastic has two outputs for the two plastic materials whereas metal had four outputs for the four metal materials in that group. after training these networks were saved and then used for classifying the resulting output from the first ann. in order to avoid false positives being identified, a threshold was empirically identified and set for testing the materials within the group. if the output of the neuron fired for the material sample being tested is not greater than . , then it was considered to be a failed classification. figure : diagram showing the ann used for material classification figure : diagram showing the two stage ann used for material classification . . . svm and ann hybrid system a hybrid approach of the svm classifier for material group classification and separate ann classifiers for the individual material classification within each group has been developed. the structure of the hybrid system can be seen in figure . figure : diagram showing the svm+ann hybrid approach used for material classifica- tion a svm classifier, as described in section . . , is used first to classify the materials into their material group (i.e. using the polynomial kf with a poly- nomial order value of and a soft margin value of . ). upon classification into their groups an output matrix is formed and the values in this matrix initiates the second stage of the hybrid approach with six anns for indi- vidual material classification within each of the material groups. the anns are constructed as described in section . . (i.e. with one hidden layer of neurons and trained for epochs). as with the other two stage ap- proaches outlined in this section, in order to improve efficiency only materials correctly classified into their group are classified in the second stage. . . . gaussian mixture models (gmms) gmm based on the maximum likelihood (ml) estimation using expec- tation maximisation (em) was also used for classification. the gaussian probability density function (pdf) is a bell shaped curve defined by two pa- rameters, mean and variance. the gaussian distribution is commonly used for approximating a class model shape in a selected feature space. it is as- sumed that the class model is truly a model of one basic class, however if the actual pdf is multi-modal it fails. gmm is a mixture of several gaussian distributions and is therefore able to represent different subclasses within a class. the resulting probability density function is defined as a weighted sum of the gaussian distributions that make up the gmm. the pdf of the gmm can be used to calculate the likelihood of any new input data within each class and identify the class for which the pdf generates the maximum likelihood and therefore classify the data. the em algorithm is an iterative method for calculating the maximum likelihood distribution in gmms (dempster et al., ). gmms are calculated to represent each material (cluster centres are ini- tialised randomly and the k -means algorithm is used for calculating conver- gence). the em algorithm is then used to calculate the maximum likelihood and the associated class label is selected. these classified labels are used to form an output matrix. as the cluster centres are initialised randomly runs of each classification (i.e. material groups and individual materials) were completed and an average classification accuracy calculated. . . . linear discriminant analysis (lda) lda finds a linear combination of features that characterises or separates two or more classes of objects or events by creating the linear combination which yields the largest mean differences between the classes (martinez and kak, ). it is a generalisation of fisher’s linear discriminant and it is a widely used method in statistics, pattern recognition and machine learning. lda can be used for binary or multi-class classification. in this work it is used for multi-class classification to identify materials individually and into their groups. . . . naive bayes naive bayes is a family of algorithms used for training classifiers capable of binary or multi-class classification. rather than the classifier taking all features of a class into account collectively to describe the class, all naive bayes classifiers assume that the value of a particular feature is completely independent of any other feature in the class meaning they can be trained very efficiently during a supervised learning task (rish, ). in this work it is used firstly to classify materials individually and secondly classify them into their groups. . . . k-nearest neighbour (k-nn) k -nn is a non-parametric method which can be used for classification or regression. in this work is it used for multi-class classification of materials. when used for classification. an object is assigned to the class which is most common amongst its k nearest neighbours following a majority vote. the number of neighbours is defined by the user and is normally a small positive integer, for example if k = then the object is assigned to the class of the single nearest neighbour or if k = then the most popular class amongst the three closest neighbours is the class that the object will be assigned to. a range of k values was tested to identify the most accurate number of neighbours to take into account for maximum classification accuracy. the k value that proved to be most effective for classifying the materials individually and into their groups was k = . . . . multi-layer perceptron (mlp) multi-layer perceptron (mlp) is a feed forward ann which uses back- propagation to train the network. a mlp consists of multiple layers of adaptive weights that are trained to map sets of input data onto a set of appropriate outputs (bishop, ). a range of numbers of layers and neu- rons on each layer was evaluated in order to find the optimal structure for the mlp. the structure of the mlp used in this work to classify the mate- rials into their groups and individually was a network consisting of layers containing neurons each. . results in order to evaluate the artificial system’s performance, analysis com- paring the algorithms was conducted. in addition, the artificial system was evaluated against human performance. in this section, the results of the human evaluation experiments are presented. the initial analysis of the extracted data collected from the fingertip during the experiments is also presented. furthermore, the performances of the aforementioned classifiers are presented, evaluated against each other and against human performance. . . human evaluation to evaluate each artificial system, they are compared against human per- formance in identifying the the same set of materials. the participants were evaluated on identifying the materials both individually and identify- ing which group each material belonged to. initially they were evaluated by only being allowed to press into the materials with their finger and assess its thermal properties and secondly they were evaluated using both a press and slide action (to evaluate the surface texture of the material). the findings from these experiments can be found in (kerr et al., ) and (kerr et al., a) respectively. . . system evaluation a comparison of the results obtained for all of the classification systems presented and the average classification accuracies of the human participants are shown in table . as specified in section . , the results for the artificial system are calculated by computing the average of the classification rates using -fold cross validation. as shown in table , svm has outperformed the other approaches for both the classification of the materials into their groups and for classification of the individual materials. the two stage svm has proven to be capable of maintaining the group classification accuracy for the classification of the individual materials in its second stage. as the svm only has to use binary classification in three out of the six groups and the most materials to classify between in the other groups is four, then it was possible to achieve such classification accuracy in the second stage using svm. the second best performing algorithm is the svm and ann hybrid approach. the svm and ann hybrid approach using the quadratic kf achieved . % classification. using the polynomial kf within the hybrid approach proved to be marginally more accurate, achieving . % accu- racy. the gaussian mixture models (gmm) approach was also a highly performing classifier with . % and over % for material groups. how- ever, it can be seen in table that aside from the fact that they perform table : table comparing the experimental results. modalities utilised run time (to train one fold) material group individual materials human participants - thermal only (kerr et al., ) n/a . % . % human participants - thermal and vibration n/a . % . % one stage ann - thermal only (kerr et al., ) . secs . % . % one stage ann - thermal and vibration . secs . % . % two stage ann - thermal and vibration . secs . % . % one stage svm - thermal and vibration . secs . % . % two stage svm - thermal and vibration . secs . % . % hybrid svm and ann - thermal and vibration . secs . % . % gmm - thermal and vibration . secs . % . % lda - thermal and vibration . secs . % . % nb - thermal and vibration . secs . % . % k -nn - thermal and vibration . secs . % . % mlp - thermal and vibration . secs . % . % with slightly less accuracy in comparison with the two stage svm algorithm, like the other systems, the gmm system also takes significantly longer to train (taking . secs to train for one fold of data) than the svm system (taking . secs to train for one fold of data). interestingly the lda and nb classifiers both performed better when classifying the materials individually rather than into their groups. this is contrary to what was expected as there are outputs when classifying into individual materials compared with only six outputs for group classification. as a consequence, lda and nb both performed marginally better than gmm for classifying the individual ma- terials with . % accuracy however performed poorer than gmm for the classification of material groups with . % accuracy for lda and . % accuracy for nb. k -nn was found to perform optimally when k = with a fast training time of . sec for one node, however in comparison with the other classifiers it is still lacking in accuracy for classifying materials individ- ually and into their groups. mlp was found to be relatively slow to train and was out performed by the majority of the other classifiers. the training times presented are based on training one fold of data for each classifier and the training was completed in matlab on a desktop pc with a intel r© xeon r© cpu e - @ . ghz processor and gb of ram. it is demon- strated in table that the novel two stage svm and hybrid svm and ann approaches have outperformed all standard approaches such as gmm, lda or singular svm and ann approaches. these architectures have proven to be more accurate than “off-the-shelf” architectures with the novel two stage svm system being the best performing and most efficient system for the task of material classification when compared with others. in order to assess robustness, completely unseen objects and materials were presented to the system. seven previously completely unseen objects in total were presented namely, a plastic bottle, a ceramic cup, a glass cup, a metal cup, a metal fork, a plastic fork and a paper plate. three datasets for each object were tested. there is no “any other material” option in the system so every material will be classified as one of the materials outlined in table . interestingly, materials that the system had not previously been trained on, i.e. ceramic and glass, were identified as a material with similar properties. the glass was identified as plastic in the majority of cases and the ceramic was classed as metal in all cases. in some instances the system struggled to identify the materials on which the system had been trained. the paper plate was only correctly identified as paper in out of the instances. however, this may be due to the thin plastic film coating on paper plates that makes them water retardant, as shown in figure (a). this made it difficult to identify the plate as paper and indeed it was also incorrectly identified as plastic in both other instances. the plastic fork was also correctly identified as plastic in of the three instances. this may be due to the fact that this specific metal fork was manufactured to be “metal like”, i.e. look like a metal fork as shown in figure (b). this means that there is a fine film on the surface of the fork making it difficult to identify as plastic, similar to the difficulty faced with identifying the paper plate. indeed, in both other cases the fork was identified as being metal. (a) (b) figure : a) image showing two of the previously unlearned objects; a) the paper plate; b) the “metal-like” plastic fork however, the metal cup was identified as the correct material in at least of the instances and both the plastic bottle and metal fork were correctly identified as their respective material in all instances. this illustrates the robustness of the system and with further training for more materials the system is capable of correctly identifying previously learned materials. . discussion from the results of the human evaluation carried out in (kerr et al., a), it was found that the average accuracy was . % when identifying the material groups and . % when identifying individual materials. a detailed analysis of these findings can be seen in the confusion matrices in figure (b) and (b). it can be seen from table that all of the artificial systems presented have outperformed the human participants (when perfor- mance is averaged across all test materials), in some cases by over % for material group classification and by over % for individual material classi- fication. this is due to the fact that is easier for a human to determine what group a material would belong to (i.e. identify that the material is definitely metal) than it is for them to identify a specific material within that group (i.e. is it copper or aluminium). it can be seen from the confusion matrix in figure (a) and (b) that both the artificial system and the human partic- ipants successfully classified the soft foam in every instance and did not get it confused with any other material. it can also be seen from figure (a) that where the artificial system classified the pine material correctly on % of the instances, the highest misclassification was with mdf, another ma- terial with similar properties and in the same material group. likewise, for classifying mdf, the largest misclassification was with pine. this misclas- sification can also be witnessed as a common misclassification amongst the human participants, with % of the participants correctly classifying mdf but in almost % of the instances the human participants misclassified mdf as pine as shown in figure (b). it is shown in figure (a) that a common misclassification of the artificial system was a confusion between acrylic and glossy cardboard, with glossy cardboard being classified correctly for % of the instances however in just over % of the instances it was misclassified as another material with similar texture and thermal conductivity properties: acrylic. interestingly, the identification of the rough materials was evidently chal- lenging for the human participants, with only out of being able to identify all three rough materials. for example, one of the most challenging materials for humans to identify was the rough copper, and it was commonly misclassified as redbrick, as shown in figure (b). this is due to the fact (a) (b) figure : a) confusion matrix showing the percentage accuracy of the artificial system (using stage svm) for classifying all the materials individually; b) confusion matrix showing the percentage accuracy of the human participants for classifying all the materials individually, collated from the findings in (kerr et al., a). that the redbrick is perceived as cold and rough by the human participants, in a similar manner to how metals are perceived. therefore, this is one of the instances where the artificial system outperforms the human participants, as the system can analyse thermal conductivity much more accurately than the human participants and therefore not misclassify it as a masonry material. this is shown in figure (a) where the artificial system classifies the rough copper in just over % of the instances with the only misclassification be- ing with another rough metal: aluminium. this common misclassification is also evident when classifying the materials into their groups, as shown in figure (a) and (b). the artificial system outperforms the human par- ticipants when classifying materials in both metal and masonry groups. the majority of instances of misclassification for the human participants when classifying masonry objects is confusing it with a metal object, as both are perceived as cold. from the results of the human evaluation reported in (kerr et al., a), it was evident that the introduction of texture analysis by sliding their finger along the material has increased the accuracy of the human participants for classifying materials, particularly when identifying the material groups. it should be noted that due to the nature of some materials, humans can in- tuitively take other characteristics into account, for example compressibility (i.e. soft foam is much more compressible than the other materials). the artificial system was trained to analyse only the thermal properties and sur- face texture of the material and not compressibility. this may have given the human participants a slight advantage in identifying some materials. this is evident when comparing the accuracy of the artificial system for classifying fabrics as a material group ( %) with the accuracy of the human partici- pants ( %), as shown in figure (a) and (b). the human participants have no misclassification of the fabrics when identifying their material group. it can be assumed that this is due to the fact that the human participants could identify the compressibility of the objects like soft foam and carpet whereas the artificial system was not trained to analysis this modality. how- ever, even though the artificial system did not consider a key modality for humans, namely compressibility, overall it still outperformed humans. . conclusion and future work a range of classifiers and combinations of classifiers that were imple- mented, tuned and evaluated for the classification of materials into their (a) (b) figure : a) confusion matrix showing the percentage accuracy of the artificial system (using stage svm) for classifying all the materials into groups; b) confusion matrix showing the percentage accuracy of the human participants for classifying all the materials into groups, collated from the findings in (kerr et al., a). respective groups and the classification of individual materials has been pre- sented. this classification was based on a combination of the thermal and surface texture properties of the materials. it is proven that a two stage svm approach performs best and is the most efficient method in completing the material classification task. it was found that the artificial system’s average performance across all test materials exceeded that of human participants’ average performance, with the two stage svm approach outperforming hu- mans by over %. however, as the system does require training it is still slower at classification than the human participants due to humans’ inherent sophisticated tactile sensing. there are two immediate future directions of this work. firstly, the pre- sented tactile based methods for material classification could be used in col- laboration with vision based methods. if an estimate of the physical size of an object is determined using image processing, then the volume of the object can also be calculated. the tactile sensing methods presented in this paper can identify the material from which the object is made and armed with the volume of the object, an estimate of the object’s weight can be calculated. the incorporation of vision would further strengthen the identification of objects and materials. to train a robotic system on a vast array of objects and incorporate vision, a significant dataset of tactile data, static image and video data should be collected for numerous materials, objects, clothing and human skin. a deep learning structure could then be utilised for the robotic system to learn the vast dataset of materials and objects. this information could be used to help determine the effort and grasping method required to successfully lift previously unlearned objects. this would provide a useful skill to assistive robots in a home environment when learning the affordances of objects they have not previously encountered. secondly, the success of the challenging task of material identification presented in this paper proved invaluable in providing familiarity with the use of the biotac fingertip and the analysis of the data it presents. it became apparent that the high sen- sitivity of the biotac sensor means that it may be capable of measuring biomedical parameters, namely human vital signs such as a human pulse. therefore, this also directs the research in this paper towards the automated collection and analysis of human vital signs based on tactile perceptions. references bhattacharjee, t., wade, j., chitalia, y., and kemp, c. ( ). data-driven thermal recognition of contact with people and objects. in ieee haptics symposium, haptics , philadelphia, pa, usa, april - , , pages – . bhattacharjee, t., wade, j., and kemp, c. ( ). material recognition from heat transfer given varying initial conditions and short-duration contact. in proceedings of robotics: science and systems, rome, italy. bishop, c. ( ). neural networks for pattern recognition. oxford univer- sity press. chathuranga, d., ho, v., and hirai, s. ( ). investigation of a biomimetic fingertip’s ability to discriminate fabrics based on surface textures. in advanced intelligent mechatronics (aim), ieee/asme interna- tional conference on, pages – . dempster, a., laird, n., and rubin, d. ( ). maximum likelihood from incomplete data via the em algorithm. journal of the royal statistical society, series b, ( ): – . drimus, a., kootstra, g., bilberg, a., and kragic, d. ( ). classification of rigid and deformable objects using a novel tactile sensor. in advanced robotics (icar), th international conference on, pages – . fishel, j. and loeb, g. ( ). bayesian exploration for intelligent identifi- cation of textures. frontiers in neurorobotics, ( ). hall, m., frank, e., holmes, g., pfahringer, b., reutemann, p., and witten, i. ( ). the weka data mining software: an update. sigkdd explorations, ( ): – . ho, v., araki, t., makikawa, m., and hirai, s. ( ). experimental in- vestigation of surface identification ability of a low-profile fabric tactile sensor. in intelligent robots and systems (iros), ieee/rsj in- ternational conference on, pages – . hoelscher, j., peters, j., and hermans, t. ( ). evaluation of tactile feature extraction for interactive object recognition. in proceedings of th ieee-ras international conference on humanoid robots, pages – , seoul. ieee. jamali, n. and sammut, c. ( ). material classification by tactile sensing using surface textures. in robotics and automation (icra), ieee international conference on, pages – . jamali, n. and sammut, c. ( ). majority voting: material classification by tactile sensing using surface texture. robotics, ieee transactions on, ( ): – . jurie, f. and triggs, b. ( ). creating efficient codebooks for visual recog- nition. in computer vision, . iccv . tenth ieee interna- tional conference on, volume , pages – . kerr, e., mcginnity, t., and coleman, s. ( ). material classification based on thermal properties - a robot and human evaluation. in ieee international conference on robotics and biomimetics, december - , shenzhen, china, pages – . kerr, e., mcginnity, t., and coleman, s. ( a). material classification based on thermal and surface texture properties evaluated against hu- man performance. in ieee international conference on control, automation, robotics and vision, icarcv , singapore, dec . kerr, e., mcginnity, t., and coleman, s. ( b). tactile approach to material classification - evaluated with human performance. in irish machine vision and image processing , northern ireland, pages – . lederman, s. and klatzky, r. ( ). hand movements: a window into haptic object recognition. cognitive psychology, ( ): – . lin, c., erickson, t., fishel, j., wettels, n., and loeb, g. ( ). signal processing and fabrication of a biomimetic tactile sensor array with ther- mal, force and microvibration modalities. in robotics and biomimetics (robio), ieee international conference on, pages – . lowe, d. ( ). object recognition from local scale-invariant features. in computer vision, . the proceedings of the seventh ieee interna- tional conference on, volume , pages – . luo, s., liu, x., althoefer, k., and liu, h. ( ). tactile object recog- nition with semi-supervised learning. in liu, h., kubota, n., zhu, x., dillmann, r., and zhou, d., editors, intelligent robotics and applica- tions, volume of lecture notes in computer science, pages – . springer international publishing. martinez, a. and kak, a. ( ). pca versus lda. pattern analysis and machine intelligence, ieee transactions on, ( ): – . matlab ( ). version . . (r a). the mathworks inc., natick, massachusetts. monkman, g. j. and taylor, p. ( ). thermal tactile sensing. robotics and automation, ieee transactions on, ( ): – . pezzementi, z., plaku, e., reyda, c., and hager, g. ( ). tactile-object recognition from appearance information. robotics, ieee transactions on, ( ): – . r core team ( ). r: a language and environment for statistical com- puting. r foundation for statistical computing, vienna, austria. rish, i. ( ). an empirical study of the naive bayes classifier. technical report. syntouch ( ). the syntouch website. varma, m. and zisserman, a. ( ). a statistical approach to texture classification from single images. int. j. comput. vision, ( - ): – . wade, j., bhattacharjee, t., and kemp, c. ( ). a handheld device for the in situ acquisition of multimodal tactile sensing data. corr, abs/ . . xu, d., loeb, g., and fishel, j. ( ). tactile identification of objects using bayesian exploration. in ieee international conference on robotics and automation (icra) . cloud computing service composition: a systematic literature review expert systems with applications ( ) – contents lists available at sciencedirect expert systems with applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a review cloud computing service composition: a systematic literature review - /$ - see front matter � elsevier ltd. all rights reserved. http://dx.doi.org/ . /j.eswa. . . ⇑ corresponding author. e-mail addresses: amin.jula@gmail.com, aminjula@ftsm.ukm.my (a. jula), elan@ftsm.ukm.my (e. sundararajan), zalinda@ftsm.ukm.my (z. othman). amin jula a,⇑, elankovan sundararajan b, zalinda othman a a data mining and optimization research group, centre for artificial intelligence, faculty of information science and technology, universiti kebangsaan malaysia, ukm bangi, selangor, malaysia b centre of software technology and management, faculty of information science and technology, universiti kebangsaan malaysia, ukm bangi, selangor, malaysia a r t i c l e i n f o a b s t r a c t keywords: cloud computing service composition systematic literature review quality of service parameter qos research objectives importance percentage of quality of service parameters the increasing tendency of network service users to use cloud computing encourages web service vendors to supply services that have different functional and nonfunctional (quality of service) features and provide them in a service pool. based on supply and demand rules and because of the exuberant growth of the services that are offered, cloud service brokers face tough competition against each other in providing quality of service enhancements. such competition leads to a difficult and complicated pro- cess to provide simple service selection and composition in supplying composite services in the cloud, which should be considered an np-hard problem. how to select appropriate services from the service pool, overcome composition restrictions, determine the importance of different quality of service param- eters, focus on the dynamic characteristics of the problem, and address rapid changes in the properties of the services and network appear to be among the most important issues that must be investigated and addressed. in this paper, utilizing a systematic literature review, important questions that can be raised about the research performed in addressing the above-mentioned problem have been extracted and put forth. then, by dividing the research into four main groups based on the problem-solving approaches and identifying the investigated quality of service parameters, intended objectives, and developing environ- ments, beneficial results and statistics are obtained that can contribute to future research. � elsevier ltd. all rights reserved. . introduction exposes a very large number of similar single services to different in recent years, the importance of affordable access to reliable high-performance hardware and software resources and avoiding maintenance costs and security concerns has encouraged large institution managers and stakeholders of information technology companies to migrate to cloud computing. the birth of giant trust- worthy clouds has led to a dramatic reduction in apprehension to- ward such an approach. there are two challenges to address from the standpoint of the significance of all of the needed service accessibilities and efficient allocation possibilities. first, anticipating all of the possible re- quired services is extremely difficult, particularly for software ser- vices. designing and providing simple and single fundamental services by different service providers will be considered constitu- tive and constructive parts of complicated required services and can be utilized in encountering this problem. the second challenge is in selecting the optimum required single services, which are pro- vided by different service providers with different quality of ser- vice (qos) attributes; an optimal combination for forming a complicated service must be composed. addressing this challenge as an optimization problem is an np-hard problem because it service providers in the cloud. service composition is one of the best approaches that has been proposed by researchers and applied by cloud providers; this ap- proach can consider both of the mentioned challenges simulta- neously. selecting appropriate services from a service pool, addressing service composition restrictions, determining the important qos parameters, understanding the dynamic character- istics of the problem, and having rapid changes in the properties of the services and network are some important issues that must be addressed in this approach to assure the service users’ satisfaction. in the early s and in the years before applications in cloud computing, service composition was introduced and investigated for web services (kosuga et al., ; milanovic & malek, ; schmid, chart, sifalakis, & scott, ; singh, ). different arti- ficial and evolutionary algorithms (ai & tang, ; canfora, di penta, esposito, & villani, ; liao et al., ; luo et al., ; tang, ai, & ieee, ; wang, ; zhao, ma, wang, & wen, ; zhao et al., ) and classic algorithms (gabrel, manouvrier, megdiche, murat, & ieee, ; gao, yang, tang, zhang, & society, ; gekas & fasli, ; liu, li, & wang, ; liu, wang, shen, luo, & yan, ; liu, xiong, zhao, dong, & yao, ; liu, zheng, zhang, & ren, ; torkashvan & haghighi, ) have been applied extensively to solve the problem. design- ing workflows and frameworks for the composition of single ser- vices to achieve specific goals is another approach that has been http://crossmark.crossref.org/dialog/?doi= . /j.eswa. . . &domain=pdf http://dx.doi.org/ . /j.eswa. . . mailto:amin.jula@gmail.com mailto:aminjula@ftsm.ukm.my mailto:elan@ftsm.ukm.my mailto:zalinda@ftsm.ukm.my http://dx.doi.org/ . /j.eswa. . . http://www.sciencedirect.com/science/journal/ http://www.elsevier.com/locate/eswa a. jula et al. / expert systems with applications ( ) – observed in the field (chen, li, & cao, ; he, yan, jin, & yang, ; song, dou, & chen, ; song, dou, & chen, ). service composition techniques were first applied in cloud com- puting systems in (kofler, ul haq, & schikuta, ; zeng, guo, ou, & han, ). afterward, a substantial effort was made in this area. the number of ongoing studies in cloud computing service composition is rapidly increasing due to the increasing ten- dency of researchers within different areas of expertise to address the problem. a glimpse of reliable published works shows that researchers who are interested in this promising area face a tre- mendous number of novel ideas, mechanisms, frameworks, algo- rithms and approaches, and further expanding the scope of problems. furthermore, several existing datasets, effective qos parameters and implementation environments with different fea- tures and effects should be recognized. hence, and due to the ab- sence of related surveys, a systematic review on cloud computing service composition is necessary and will help facilitate future re- searches. a systematic review in which the most important aspects of the accomplished researchers must be investigated, and useful information and statistics must be extracted. this paper provides a systematic literature review on state-of- the art approaches and techniques in addressing cloud computing service composition. the discussed advancements and develop- ments of this topic provide useful information to motivate further investigations in this area. identifying the different objectives of performing cloud computing service composition studies and the reasonable and purposeful classifications of such divergent ap- proaches and mechanisms is a major achievement of the review. furthermore, this study extracts all of the considered qos param- eters, introduces the most significant and the least considered parameters and calculates the importance parameter percentages, aiming to eliminate barriers to future research efforts, such as pro- posing comprehensive and reliable mathematical models for calcu- lating composite service qos values. the goals of this research also include declaring the appropriate, investigated qos datasets, utiliz- ing software in generating problems for evaluating methods and discussing the most widely used implementation environments. because the investigated subject is very extensive, it is impossi- ble to include all relevant topics. hence, some related subjects do not fall within the research scope of this review but will be briefly mentioned. providing network services requires common lan- guages and protocols and is closely related to service composition. however, services that provide details will not be considered in this work. strong and independent studies are required to address research methodologies and experimental and statistical perfor- mance evaluation strategies. describing the accurate significance of the parameters for real cloud customers is also beyond the scope of this report but can be achieved by conducting comprehensive studies, utilizing interviews and questionnaires and adopting appropriate statistical methods. the structure of this paper is organized as follows. after the introduction, the main aims of the paper and research questions will be defined and described in section . briefly, cloud computing and its characteristics will be explained in section . cloud com- puting service composition (ccsc) will be defined in section , including its challenges and classified applied approaches. in sec- tion , an extensive discussion on the objectives of the investigated research, their approaches, and utilized datasets is provided, and effective information and statistics are extracted for future re- search. the final sections of the paper contain the conclusions and references. . survey goals and execution survey goals and research questions are described in section . , and the statistics on published and presented papers in different journals and conferences are presented in section . . the authors of the present paper have profited from ‘‘guidelines for performing systematic literature reviews in software engineering’’ (kitchenham & charters, ) and have also used (garousi & zhi, ) to conduct and perform this research. . . survey goals and research questions the present research aims at collecting and investigating all of the credible and effective studies that have examined ccsc. more specifically, the extraction of salient features and methods of pa- pers will be considered, and their characteristics will be described. to achieve the above-mentioned goals and identify the methods that have been selected by researchers for their studies and result assessment methods, case studies are covered for which new methods are proposed and datasets and benchmarks are used. most researchers have considered qos parameters and proposed objective functions and user trends that are important in designing these functions. the following research questions (rqs) are raised. rq . what are the main goals of the researches? rq . what is the proposed approach and what are the methods used? how have the researchers conducted the research? rq . what datasets or benchmarks are used and what case studies are considered? rq . what evaluating procedures have been used to assess the results in each paper? rq . what other research has been considered in each paper to compare the results? rq . what qos parameters are accounted for? rq . how can the user requirements and tendencies be consid- ered in the applied objective function? . . publication statistics in this study, attempts have been made to examine all of the pa- pers that have been published on ccsc in particular and that in- clude novel methods or interesting ideas. to achieve this goal and to answer the research questions in section . , papers that were published from to december were selected from different high-level refereed journals and prestigious international conferences and are considered in section . in each study, if it was required to be familiar with some concepts and methods and to read further on the topic, other books and papers are proposed and referred to. the result of this effort is a comprehensive collec- tion of resources that can provide an acceptable level of concepts and information about the service composition problem in cloud computing and the different views of addressing this problem that are introduced in the literature. . cloud computing . . cloud definition there are different definitions for cloud computing in the liter- ature, many of which do not cover all of the features of the cloud. in one attempt, vaquero et al. attempted to extract a comprehensive definition using different compliments (vaquero, rodero- merino, caceres, & lindner, ). efforts have been made to stan- dardize the definition of the cloud, in which we accept the cloud definition provided by the national institute of standards and technology (nist) (peter mell, ). the nist cloud computing definition: ‘‘cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, a. jula et al. / expert systems with applications ( ) – servers, storage, applications, and services) that can be rapidly provi- sioned and released with minimal management effort or service pro- vider interaction. this cloud model is composed of five essential characteristics, three service models, and four deployment models’’. the cloud cannot be considered to be a new concept or technol- ogy that arose in recent years; instead, its root can be found in what john mccarthy described as the ability to provide computa- tional resources as a ‘‘utility’’ (hamdaqa & tahvildari, ). based on standard material presented by nist, cloud computing is com- posed of five main characteristics, and two other characteristics are added based on the literature, with three spanning service models and four models of deployment, which will be described in some detail in the following sections (see fig. ). . . cloud computing characteristics on-demand self-service. a user can request one or more services whenever he needs them and can pay using a ‘‘pay-and-go’’ method without having to interact with humans using an online control panel. broad network access. resources and services that are located in different vendor areas in the cloud can be available from an extensive range of locations and can be provisioned fig. . cloud computing, characteristics, deployment m through standard mechanisms by inharmonious thin or thick clients. the terms ‘‘easy-to-access standardized mechanisms’’ and ‘‘global reach capability’’ are also used to refer to this characteristic (hamdaqa & tahvildari, ; yakimenko et al., ). resource pooling. providing a collection of resources simulates the behavior of a single blended resource (wischik, handley, & braun, ). in other words, the user does not have knowl- edge and does not need to know about the location of the pro- vided resources. this approach helps vendors to provide several different real or virtual resources in the cloud in a dynamic manner. rapid elasticity. fundamentally, elasticity is another name for scalability; elasticity means the ability to scale up (or scale down) resources whenever required. users can request different ser- vices and resources as much as they need at any time. this char- acteristic is so admirable that amazon, as a well-known cloud service vendor, has named one of its most popular and commonly used services the elastic compute cloud (ec ). measured service. different aspects of the cloud should automat- ically be controlled, monitored, optimized, and reported at sev- eral abstract levels for the resources of both the vendors and consumers. odels, service pool, and types of services and users. a. jula et al. / expert systems with applications ( ) – multi-tenacity. this concept is the fifth cloud characteristic that is suggested by the cloud security alliance. multi-tenacity means that it is essential to have models for policy-driven enforcement, segmentation, isolation, governance, service lev- els, and chargeback/billing for different consumer categories (espadas et al., ). auditability and certifiability. it is important for services to pre- pare logs and trails to make it possible to evaluate the degree to which regulations and policies are observed (hamdaqa & tahvildari, ). . . cloud computing service models definition (service). a service is a mechanism that is capable of providing one or more functionalities, which it is possible to use in compliance with provider-defined restrictions and rules and through an interface (ellinger, ). definition (platform). a platform is a fundamental computer system that includes hardware equipment, operating systems, and, in some cases, application development tools and user inter- faces on which applications can be deployed and executed. definition (infrastructure). infrastructure refers to underlying physical components that are required for a system to perform its functionalities. in information systems, these components can contain processors, storage, network equipment, and, in some cases, database management systems and operating systems. software as a service (saas). a software or application that is executing on a vendor’s infrastructure is recognized as a service provided that the consumer has limited permission to access; the provision is through a thin client (e.g., a web browser) or a program interface for sending data and receiving results. the consumer is unaware of the application provider’s infra- structure and has limited authority to configure some settings. platform as a service (paas). in this service model, the service vendor provides moderate basic requisites, including the oper- ating system, network, and servers, and development tools to allow the consumer to develop acquired applications or soft- ware and manage their configuration settings. infrastructure as a service (iaas). the consumer has developed the required applications and needs only a basic infrastructure. in such cases, processors, networks, and storage can be provided by vendors as services with consumer provisions. . . cloud computing deployment models public cloud. this approach is the major model of cloud comput- ing; here, the cloud owner provides public services in the vast majority of cases on the internet based on predefined rules, pol- icies, and a pricing model. possessing a large number of wide- spread world resources enables providers to offer a consumer different choices to select appropriate resources while consider- ing the qos. private cloud. a private cloud is designed and established to pre- pare most of the benefits of a public cloud exclusively for an organization or institute. setting up such a system because of the utilization of corporate firewalls can lead to decreased security concerns. because the organization that implements a private cloud is responsible for all of the affairs of the system, facing abundant costs is the blind spot of establishing a private cloud. community cloud. based on their similar requirements, con- cerns, and policies, a number of organizations establish a com- munity and share cloud computing to be used by their community member’s consumers. a third-party service pro- vider or a series of community members can be responsible for providing the required infrastructure of the cloud comput- ing. lowering costs and dividing expenses between community members along with supporting high security are the most important advantages of a community cloud (dillon, chen, & chang, ). hybrid cloud. a combination of two or more different public, pri- vate, or community clouds led to the creation of a different cloud model called hybrid cloud, in which constitutive infra- structures not only keep their specific properties but also require standardized or agreed functionalities to enable them to communicate with each other with respect to interoperabil- ity and portability on applications and data. . the cloud computing service composition problem (ccsc) . . ccsc definition fast development in the utilization of cloud computing leads to publishing more cloud services on the worldwide service pool. because of the presence of complex and diverse services, a single simple service cannot satisfy the existing functional requirements for many real-world cases. to complete a complex service, it is essential to have a batch of atomic simple services that work with each other. therefore, there is a strong need to embed a service composition (sc) system in cloud computing. the process of service introduction, requesting, and binding, as shown in fig. , can be accounted for in such a way that service providers introduce their available services to the broker to expose them to user requests. however, users also send their service re- quests to the broker, who must select the best service or set of ser- vices on the basis of user requirements and tendencies; the broker wants providers to bind selected services to the users with respect to predefined rules and agreements. increasing the number of available services causes an increase in the number of similar-function services for different servers. these similar services are located in different places and have dis- tinct values in terms of the qos parameters. for this reason, sc ap- plies appropriate techniques to select an atomic service among the different similar services that are located on distinct servers to al- low the highest qos to be achieved according to the end-user requirements and priorities. because of intrinsic changes in cloud environments, available services, and end-user requirements, sc should be designed dynamically, with automated function capabilities. therefore, selecting appropriate and optimal simple services to be combined together to provide composite complex services is one of the most important problems in service composition. the sc problem in cloud computing can be defined as determining what atomic simple services should be selected such that the ob- tained complex composite service satisfies both the functional and qos requirements based on the end-user requirements. be- cause of various and abundant effective parameters and a large number of simple services provided by many service providers in the cloud pool, ccsc is considered an np-hard problem (fei, dongming, yefa, & zude, ). in this paper, it is assumed that every composed service (cs) in the cloud consists of n unique services (uss) and has p qos param- eters. to terminate a cs, a combination of unique services act sequentially in an ordinal workflow (wf), as shown in fig. . fig. . process of service introduction, requesting, and binding. a. jula et al. / expert systems with applications ( ) – define qi (usj) as the value of qos parameter i for unique service j. accordingly, the qos vector of unique service j is defined as eq. ( ): qðusjÞ¼ ðq ðusjÞ; q ðusjÞ; . . . ; qpðusjÞÞ ð Þ however, if wfk is the workflow of csk, then it is possible to define qi (wfk) to obtain the value of qos parameter i for workflow k. then, the qos vector of workflow k is described in eq. ( ): qðwfkÞ¼ ðq ðwfkÞ; q ðwfkÞ; . . . ; q pðwfkÞÞ ð Þ . . ccsc challenges the dynamic nature of cloud environments involves occasional and consciously planned changes; these changes expose cloud computing to different challenges in the sc. the most remarkable challenges are the following: dynamically contracting service providers. the pricing policy of most service providers is determined by service fees based on supply and demand. thus, mechanisms for updating the table of available resource characteristics must be predicted (gutierrez-garcia & sim, ). addressing incomplete cloud resources. optimal service selec- tion by a broker depends on the availability of complete and updated information on the services. facing several changes in the service characteristics could result in the loss of some data (gutierrez-garcia & sim, ; yi & blake, ). describing and measuring qos attributes of network services. a lack of consensus on the definition and description of network services’ qos parameters among worldwide cloud service fig. . cloud service co vendors is still an important challenge for cloud developers. the absence of agreed-upon ways to measure network qos is another problem that is not completely solved and that should be considered (qiang, yuhong, & vasilakos, ). interservice dependency/conflict. dependency or conflicts that exist among two or more services leads to a complicated service composition problem. in real-world scenarios, encountering dependency and conflict among services is quite common and should be considered in sc (strunk, ). security. designing and apprising security rules, policies, and instructions are among the basic responsibilities of cloud ser- vice vendors. however, a principled self-administered frame- work for supplying and provisioning services in which vendors’ security concerns and policies are observed must be provided (takabi, joshi, & gail-joon, ; zissis & lekkas, ). . . current approaches for ccsc the different approaches proposed in the literature can be divided into five distinct categories: including classic and graph- based algorithms (cgbas), combinatorial algorithms (cas), ma- chine-based approaches (mbas), structures (sts), and frameworks (fws). different studies that belong to these categories are investi- gated in sections . . to . . . . . . classic and graph-based algorithms ccsc is a problem in which there are many potential solutions, among which one or a limited number of solutions are optimal. thus, ccsc is known as an optimization problem (anselmi, arda- gna, & cremonesi, ; yu & lin, ). some types of classic algorithms, such as backtracking and branch-and-bound, can be used to solve optimization problems. these algorithms can guaran- tee that an optimal solution will be found solely by taking expo- nential time complexity (neapolitan & naimipour, ). thus, using classic algorithms to solve optimization problems is possible only with some modifications and improvements for decreasing the execution time. to achieve a feasible concrete workflow for ccsc with respect to the consumer qos requirement, the problem is considered to be equivalent to a multi-dimensional multi-choice knapsack problem (mmkp) in which a parameter called happiness that is calculated based on qos parameters is used as the utility (kofler et al., ). to solve the mmkp and obtain the benefits of heterologous multi-processing environments (e.g., grid computing (preve, )), a parallel form of the branch-and-bound algorithm is proposed in which each node of the decision tree is an independent mposition process. a. jula et al. / expert systems with applications ( ) – instance of a main routine that could be assigned to separated computational nodes. to evaluate the algorithm, researchers ran- domly generated four different size problems; they executed the problems in serial and parallel modes and compared the results against one another. it is obvious that the most important problem that the proposed algorithm faces is the exponential complexity execution time. the performance evaluation of the algorithm could also be completed and could be more reliable if the results were compared with similar algorithms’ results on real-world datasets. to rectify the high execution time problem in the method, a two- phase approach is also proposed in kofler, haq, and schikuta ( ). in this approach, the first phase is similar to the previous phase, whereas for second phase, the previous existing solutions are typically reused for similar requests, and some changes are ap- plied to correspond with the new situation. after designing a new cloud resource data-saving method, a matching algorithm called sma is applied in zeng et al. ( ) to check whether the output parameters of a service and the input parameters of another service are compatible with each other. in this algorithm, the service matching problem is mapped to calcu- late the semantic similarity of the different input and output parameters of different services. the matching degrees of each pair of services that are stored in a table lead to the composition of a weighted directed graph in which finding all reachable paths for two nodes can yield all of the possible service composition meth- ods. researchers proposed an improved fast-ep algorithm called fastb+�ep to obtain all possible paths in a shorter amount of time. however, two-step graph building and searching leads to increased execution time and decreasing algorithm performance, especially in cases when there is an increase in the size of the problem and the number of required services. in view of all of the cloud participants (e.g., resource vendors and service consumers) as agents and simulating their activities by a colored petri net (barzegar, davoudpour, meybodi, sadeghian, & tirandazian, ; jensen, ), an agent-based algorithm for ccsc is introduced (gutierrez-garcia & sim, ). in the proposed algorithm, the broker agent receives a consumer agent’s require- ments and attempts to find a service vendor agent for each single service. thereafter, the contract net protocol (smith, ) is used to select proper single services to compose the solution. using the experimental results, it is proven that the proposed method can achieve solutions in linear time with respect to the number of ser- vice vendor agents and that it can consider parallel activities to ad- dress heterogeneous services. a two-phase service composition approach is proposed for dy- namic cloud environments in which the service performance changes in different transactions (shangguang, zibin, qibo, hua, & fangchun, ). in the first phase, the cloud model (li, cheung, shi, & ng, ) is applied to change the qualitative value of the qos parameters to their quantitative equivalent to calculate the uncertainty level. thereafter, mixed integer programming (mip) is utilized in the second phase to find appropriate services in which the binary decision vector is used to determine whether a service is selected. another two-phase method in which mip is also applied to focus on service performance fluctuations in the dynamic cloud environment by solving ccsc is developed in zhu, li, luo, and zheng ( ). to decrease the number of candidate single services, some appropriate services are selected based on the history of sin- gle services and using k-means in the first phase of the proposed method. in the second phase, mip is used to select the best single services among the preliminary selected services. the authors proved that their two-phase approach can outperform hiresome (lin, dou, luo, & jinjun, ). applying linear programming (lp) (korte & vygen, ; vanderbei, ) for optimizing virtual machine resource alloca- tion in physical servers for video surveillance was proposed in hossain, hassan, al qurishi, and alghamdi ( ) for the first time. composing an optimal set of media services to prepare a composite service over virtual machine resource allocation was also consid- ered. to reach this goal, the authors mapped the virtual machine resource allocation problem to the multi-dimensional bin-packing problem and used lp to solve it. they also suggested the use of a customized best-fit decreasing method (kou & markowsky, ) to solve the problem, which considerably increases the probability of finding appropriate results compared to lp but cannot guarantee that an optimal solution will be obtained. to evaluate two pro- posed methods, randomly generated service composition problems were considered, and the results are compared to fractional knap- sack mapping and round-robin allocation execution results. for generating composite services with the lowest execution cost a two-phase method is applied in liu et al. ( ). the first phase includes utilizing a state transition matrix for the analysis of the dynamic process of composite service execution. in this phase, each composite service status was modeled on a state tran- sition diagram, which is used to produce a state transition matrix. the execution cost of each composite service can be calculated using its state transition matrix. in phase two, business process execution language for web services (bpel ws) (huang, lo, chao, & younas, ) is used to find an optimal solution. because the process of bpel ws, which is known as the distributed traffic model, is extremely time consuming, researchers divided it into three parts, each of which is executed by an independent computer. to consider the user preferences and different resource charac- teristics bundled together, liu et al. proposed a three-layer hierar- chical structure; the first layer includes optimal service selection; the second layer is called the criterion layer and includes timeli- ness, stability, and security, based on which services are divided into three categories; and the third layer is designed for the repre- sentation of nine additional qos parameters (liu et al., ). next, the phases of the proposed algorithm (ssup) include generating and normalizing a user requirement matrix, generating a qos weight specification using the analytic hierarchy process (ahp) (jeonghwan, rothrock, mcdermott, & barnes, ), generating and normalizing the service attribute matrix, and performing the service-demand similarity calculation. the experimental results indicate that ssup could outperform ahp when solving the ccsc problem. worm et al. successfully addressing two antithetical aims for the same provider, namely, the aims of revenue maximization and quality assurance (worm, zivkovic, van den berg, & van der mei, ). the authors based their method on three decision cri- teria: service availability at the decision instant, executing cost, and resting time to deadline. with respect to the response time and with an emphasis on service availability, dynamic program- ming (dp) is used to achieve the main goals of the study, in which it is necessary to save intermediate results to select the best service among all available services. in addition, in the case of a deadline, an arrival decision rule would select services with the lowest price for the remaining tasks. the proposed algorithm is executed for static and dynamic service composition but uses real-world data- sets and benchmarks, comparing the results with the previous re- search; calculation time and memory complexity are neglected. zhou and mao proposed a cloud-based semantics approach for the composition of web services utilizing a bayesian decision (zhou & mao, ). the authors applied a bayesian approach to antici- pate the semantics of a web service for which a discovery graph is generated based on a cloud to use for implementation. they also obtained relations that are based on the graph that could encoun- ter a markov chain (song, geng, kusiak, & xu, ). in addition, using an equation for the cosine theorem (xiangkun & yi, ), it is possible to achieve a similarity of services, which helps to a. jula et al. / expert systems with applications ( ) – identify the cloud service interface through web services and de- velop a composition system. the proposed method has been ap- plied to amazon services and can be compared with an existing approach (stewart, ). on the basis of variability modeling (sinnema & deelstra, ), cloud feature models (cfms) are presented as mechanisms to explain cloud services and user requirements together to pre- pare a suitable ground for their cloud service selection process (cssp) (wittern, kuhlenkamp, & menzel, ). a cfm is repre- sented by a directed graph in which the nodes play service roles and the edges denote the relations between services. further- more, the cssp should satisfy user objectives and also respect the requirements. to satisfy these goals, the user must enter the objectives and requirements to start the process, and the cssp uses decision support tools to generate a list of offered services based on the input model. focusing on the dynamic characteris- tics of cloud computing and updating the graph throughout the execution could help the proposed method to enhance its capabilities. a two-step procedure to satisfy users’ qos requirements is proposed (huang, liu, yu, duan, & tanaka, ) in which in the first step, for each user’s request, the proposed method at- tempts to select single services that meet the first two types of user functional requirements and that eliminate the remaining requirements. then, a virtual network of service providers of se- lected services is generated and modeled by a directed acyclic graph (dag). in the second step, in the case of a single user’s qos requirement, it is sufficient to apply an algorithm to deter- mine the shortest path in the dag. however, for the two qos parameters, the problem is changed to a multi-constrained path problem (korkmaz & krunz, ) when this constructing auxil- iary graph fptas (cagf) is used. in the cagf, an auxiliary graph is used in which the two-weight preliminary dag is changed to a one-weight dag (which can be solved as in the previous case) by considering the first qos parameter as a weight and merging the second parameter into the connectivity among the expanded vertices. researchers have also proposed a method for consider- ing more than two qos parameters. the execution time and re- turned path weight (the final qos) are two parameters that are considered when making a comparison with existing dag-based algorithms to determine the performance of the method; however, generating several graphs in the algorithm is a time- consuming activity that increases the execution time. this inter- esting study could also be enriched by addressing known datasets. to obtain the best set of single services for service composition, three algorithms are proposed (qi & bouguettaya, ) for gener- ating a service skyline that can be considered a specific set of ser- vice vendors that other possible sets cannot dominate with respect to the qos parameters (yu & bouguettaya, ). the basic idea of these algorithms is based on reducing the search space. the first algorithm is called opa, and it examines all of the service execution plans, one for each phase, and saves the best found solution. the researchers also applied some improvements on opa for decreas- ing the processing time and the consumed space. dpa is a second algorithm; it uses a tree structure and is based on the progressive examination of service execution plans that are sorted in ascending order based on their score and progressive results output. this pro- gressive method provides the algorithm pipeline capabilities. to overcome the problem of the repetition of nodes and thus over- head, researchers have proposed reducing the parent table data structure. for a third attempt, a linear approach is used to design a bottom-up algorithm (bua) to address the expensive computa- tional costs of dpa for an increasing number of services. evaluating the proposed algorithms by using different models of problems is another advantage of this study. hiresome-ii is proposed (dou, zhang, liu, & chen, ) to im- prove the reliability of composition qos values. with hiresome-ii, the qos history of services is investigated for evaluation instead of what the service providers claim. thus, the context of service com- position is implemented using a two-layer tree structure in which the required service is located in the root, and the nodes are lo- cated in the second layer as candidate services. in addition, the k-means clustering algorithm can be used to categorize the history of each candidate service into two peer groups. inserting these two peer groups into the tree will provide a three-layer tree, the task- service-history record tree, which identifies the best performance history for candidate services. based on the required accuracy, additional history layers can be inserted into the tree. based on experimental results, the performance of the proposed method proved to be superior to the authors’ previous method, hiresome-i, especially in the case of facing abundant candidate services. despite its strengths, with increasing required services, service providers and history volume, this method will face significant increases in execution time due to the use of k-means for categorizing the history of all candidates. in (wu, zhang, zheng, lou, & wei, ) a model for qos-satis- fied predictions in ccsc is proposed based on the hidden markov model (hmm) (li, fang, & xia, ; zeng, duan, & wu, ). the proposed model was to ensure customer satisfaction of the composite service qos, utilizing the basic form of hmm, in which qos parameters are considered as a state-set, and distinct observa- tion-probability functions are each defined for one of the parame- ters. due to the high computational complexity of the functions, researchers also applied a two part, forward–backward structure for reducing the complexity, in which each part includes three stages, initialization, recursion and end. to evaluate the proposed model, a real cloud computing simulation system was constructed that provides more than different services. the model param- eters can be adjusted with a support vector machine-based algo- rithm. the obtained evaluation error rates were small but with respect to the complicated structure of the model, which will face many services in the real world. this model is predicted to be con- fronted with larger error rates and an ineligible execution time. a novel method for qos mapping is proposed in which a set of three rules is used to map qos specifications and guarantees to- gether in the cloud (karim, chen, & miri, ). to overcome the complexity, the analytic hierarchy process (ahp) (ishizaka & labib, ; jeonghwan et al., ) is utilized in which customer prefer- ences, possible solutions and ranking the services, are conditions, search space and the goals, respectively. ahp has also led to specify qos weights to rank and select candidate services. to evaluate the method, a case study was described, and the obtained results were discussed. because the ahp was constructed from a graph that in- cluded many-to-many relationships, increasing the number of re- quired services and search space led to increased method time complexity, which is unacceptable in real environments. two novel cloud service ranking approaches, cloudrank and cloudrank , are proposed in which the basic strategy ranked the services based on the prediction of their qos (zibin, xinmiao, yilei, lyu, & jianmin, ). these two algorithms are composed of three steps, including similarity computation, preference value compu- tation and ranking prediction, and are different with respect to preference value treatment. because of the negative effects of equal preference value treatment in the accuracy of ranking pre- diction in the first algorithm, the confidence value of service qos preference values is defined and used in the treatment process of the second algorithm to remove the effects, reaching more accu- racy. researchers have also provided a new dataset called tpds that includes a response-time and throughput of services provided by users to evaluate the proposed algorithms. the evaluation results indicate that the algorithms outperformed a. jula et al. / expert systems with applications ( ) – previous approaches; however, based on mathematical consider- ations, the time complexity function of the algorithms was o(n - m) + o(n m) + o(n ), where n and m are the number of services and users, respectively. because the number of services is typically much greater than the number of users, apparent increases in the execution time of the algorithms may occur. . . . combinatorial algorithms ccsc is an optimization problem, and a distinctive feature of this problem is being involved in very large search spaces to obtain optimal results. in addition, the importance of achieving an opti- mal solution in a shorter time frame forces researchers to seriously consider the use of combinatorial algorithms (knuth, ; abonyi et al., ). thus, different efforts have been conducted using dif- ferent models of combinatorial algorithms to solve the ccsc prob- lem, and these efforts are investigated in the following studies. to simplify the composition of services in cloud computing without priority weights, an algebra of communication process (acp) (fokkink, ) based on a qos value aggregation algorithm is applied (zhang & dou, ) in which an artificial neural net- work (haykin, ) is used to focus on service consumer require- ments without predefined user priority parameters. a three-layer perceptron (shrme, ) neural network is used in this study by employing back-propagation (wang, wang, zhang, & guo, ) and least-mean-square-error (sayed, ) algorithms to learn the network and adjust the weights, which are initialized ran- domly. in the proposed neural network, the number of neurons for the input layer is equal to the number of qos parameters con- sidered. the number of neurons in the second layer is also equal to the number of neurons in the first layer, and the third layer has one neuron for an output. during the training process, the users’ prior- ity information is entered into the neural network to realize suit- able values for the weights. weight parameter efficiency is evaluated by applying the mean square error (zhou & bovik, ) as the objective function. in this research, datasets and benchmarks are not investigated, and the results are not compared. by dividing the qos parameters into three groups, including ascending, descending, and equal qos attributes, and by using sim- ple additive weighting to normalize the values of those parame- ters, a new model for calculating the qos of composite services (ye, zhou, & bouguettaya, ) is proposed. these authors also applied a genetic algorithm to solve the ccsc problem in which a roulette wheel selection algorithm is used to select chromosomes to execute a crossover operation. the proposed model uses the achievement of qos services as the fitness value. the results ob- tained with the proposed method were compared with different existing algorithms, such as lp and culture algorithm (reynolds, ) and has shown to be more efficient. to achieve the goal of the representation of automatic service compositions for dynamic cloud environments, in jungmann and kleinjohann ( ) the problem is modeled as a markov decision process (arapostathis, borkar, fernández-gaucherand, ghosh, & marcus, ; chang & yuan, ). a set of all possible composi- tions of services and a set of all possible actions for changing com- positions by adding new valid services are generated, and the composite service provider is considered an agent. a reward func- tion is also used to learn the optimal policy and optimal composi- tion for a user request that utilizes utility values, such as previous user feedback and former execution information. the reward func- tion results are used by the agent for subsequent decisions and up- dated using reinforcement learning techniques (kaelbling, littman, & moore, ; liu & zeng, ). a judgment about this research cannot be made due to a lack of appropriate performance evalua- tion and comparison of results. because of the growing importance of networks in the qos of cloud service composition, in klein, ishikawa, and honiden ( ) an approach is suggest for considering the non-network and network qos of services separately. to this end, they estimated the real network latency among the desired services and their applicants with low time complexity by proposing a network mod- el that allowed services that had a lower latency to be selected. researchers also introduced a qos equation to calculate the net- work qos, latency, and transfer rate. in the last phase of the ap- proach, based on the genetic algorithm, a selection algorithm is designed to apply proposed models to generate composite services, and its results are compared with the dijkstra algorithm (neapolitan & naimipour, ) and random selection. the results of this interesting study could be enriched by the use of real-world datasets. with respect to self-adaptivity (denaro, pezze, & tosi, ) in the service provider system, an improved genetic algorithm is pro- posed (ludwig, ) in which a clonal selection algorithm (de castro & von zuben, ) is used instead of tournament selection to select individuals for the crossover and mutation operators in a more discernible manner. the explanation of the proposed algo- rithm and the experimental results in the original paper do not go into detail regarding the researchers’ work on self-adaptivity. in yang, mi, and sun ( ) game theory is used to propose a service level agreement (sla)-based service composition algo- rithm. in this research, sla is defined as quadruple, including sla main information, service vendors and consumer information, ser- vice type and parameters, and a set of responsibilities for service vendors and consumers. to establish an sla, sc is intended to be a multiple dynamic game, called a bid game, in which service ven- dors and service consumers are players that aim to achieve their goals. in this competitive game, every consumer should promul- gate a price for each requested service based on effective parame- ters and other consumers’ proposed prices, and vendors can assign their services according to received proposed prices with respect to the requested level of the quality of services that are agreed upon and signed in the sla. the reliability of this method is limited due to the lack of comparison with other techniques and the lack of comparison with real-world datasets. a parallel form of the chaos optimization algorithm (jiang, kwong, chen, & ysim, ) called fc-paco-rm is proposed to solve the ccsc problem (fei, yuanjun, lida, & lin, ). the researchers attempted to dynamically modify the sequence length based on the evolutionary state of the solutions. they also utilized the roulette wheel selection algorithm before executing the chaos operator to eliminate improper randomly generated solutions and escape from their destructive consequences. because one of the main goals of this study has been to reduce the execution time, parallelization of the proposed algorithm is also considered. to accomplish this goal, a full-connection topology has been selected because of its high searching capability and message passing inter- face (mpi) (barney, ). finally, a novel migration method called way-reflex migration is introduced and applied to reduce commu- nication overhead of fully connected topologies. compared with a genetic, chaos genetic and chaos optimization algorithm, the pro- posed method showed better results in view of the best solution fitness and execution time. in wang, sun, zou, and yang ( ) particle swarm optimiza- tion (pso) with integer array coding is applied to achieve a fast method to solve the ccsc problem. to achieve this goal, the skyline operator with binary decision variables is used to eliminate impro- per services from the search space. because it is necessary to con- trast all of the services pair-wise when using the skyline operator, its execution time is not acceptable for an increasing number of services. thus, the researchers used offline skyline services to de- crease the response time (borzsony, kossmann, & stocker, ). the qws dataset (al-masri & mahmoud, ) and synthetic generator (wang, sun, & yang, ) are used to evaluate the a. jula et al. / expert systems with applications ( ) – algorithms. comparing the results of the proposed algorithm with another service composition called goa (ardagna & pernici, ) demonstrates that the proposed method achieves positive results. simultaneous optimization of the response time and execution cost of the service composition process motivated jula et al. to pro- pose imperialist competitive-gravitational attraction search algo- rithm (icgas) as a new memetic algorithm (moscato & cotta, ) for solving the ccsc problem (jula, othman, & sundararajan, ). because they wanted to address the enormous number of services that were provided by several service vendors, they at- tempted to apply the advantages of the exploration capability of an imperialist competitive algorithm (ica) (atashpaz-gargari & lu- cas, ) and the significant exploitation ability of a gravitation search algorithm (gsa) (jula & naseri, ) simultaneously. in the proposed algorithm, to balance the process of execution of the two algorithms and to increase the performance, the number of population members of gsa is determined to be equal to % of the number of members of the ica algorithms. researchers also introduced a new mathematical model for calculating the qos of the nominated composite services based on user-defined weights for different qos parameters. using a very large real-world dataset to evaluate the algorithm and to compare its results with the re- sults from executing a genetic algorithm, pso and classic ica have demonstrated suitable results for icgas. a negative selection algorithm, which is inspired from a nega- tive selection mechanism in biological immune recognition and can eliminate improper solutions through an execution process, is used in zhao, wen, and li ( ). in the proposed algorithm, the solutions and candidate services are implemented in the form of vector and integer strings, respectively. there are two different applied models for calculating the local and global fitness. to com- pute the fitness of the solutions for a local search, users are respon- sible for defining their criteria with related weights and a set of constraints based on which quality vectors are generated for all of the services for the decision-making process. global fitness is also computed using an equation that is designed based on local fitness. the performance capability of the proposed algorithm is proven by comparing its results with the results obtained from standard particle swarm intelligence and the clonal selection im- mune algorithm. attempting to select the best services based on overall quality of services a model is designed in which customer feedback and the objective performance analysis results are considered as two inputs, and the output is the quality of services (lie, yan, & orgun, ). because customer feedback is composed of linguistic vari- ables, a mapping table is applied to change the inputs to fuzzy numbers. then, two input values that were changed to fuzzy rat- ings and a fuzzy simple additive weighting system (chou, chang, & shen, ) are utilized to obtain the model output. researchers have also applied a filtering mechanism for removing misleading values that are given by unprofessional customers. the method was evaluated with a case study; however, the results have not been compared to other approaches. this method has also not been applied to select a set of services for assessing its effectiveness in service composition. thus, this method cannot present proper effi- ciency in the case of facing many qos parameters, customer feed- backs and requests. . . . machine-based methods in bao and dou ( ) researchers designed finite state ma- chines (fsms) (koshy, ) to consider service correlations and rightful task executing sequences; each fsm is used to implement a group of services that have limitations on their invocation and executing sequence called cws community and another fsm for a desired composite service called the target process. the proposed method is composed of two phases. in the first phase, a composi- tion service tree is created on the basis of cws community and a target process. a pruning policy is also applied to eliminate impro- per paths of the tree and to reduce the processing time. the qos of all paths in the tree are calculated in the second phase based on user requirements, and the path with the highest qos is selected as the final solution. to reach this goal, a dichotomous simple addi- tive weighting technique (liangzhao et al., ) is used in the first part, from which the scaling of paths is performed by dividing them into negative and positive criteria, and in the second part, the total qos of all of the paths is calculated. researchers could prove appropriate efficiency of their proposed methods by compar- ing its executing time with the executing time of the enumeration method (gerede, hull, ibarra, & su, ). . . . structures based on the b+-tree (bayer & unterauer, ), an index struc- ture has been designed by in sundareswaran, squicciarini, and lin ( ) to simplify the process of information insertion and retrie- val for cloud service providers. in the proposed structure, different properties, such as the service type, security, qos, and measure- ment and pricing units, have specific locations to be stored and considered. to increase the speed at which the information man- agement operators are executed and appropriate vendor queries can be found, service vendors with the same characteristics should be stored together in adjacent rows. researchers also proposed a query algorithm based on a designed structure to search the pro- viders’ database for the best vendors; after finding k vendors close to the optimal for vendors with each desired service, a refinement procedure is designed to reduce the number of selected vendors and sort them according to their hamming distances, which entails starting with the optimal and progressing in ascending order to facilitate the selection of better providers. the proposed method is compared with a brute-force search algorithm and has shown al- most times better execution speed for solving the ccsc prob- lem with , service providers. . . . frameworks pham, jamjoom, jordan, and shae ( ) proposed a new frame- work for service composition in which a composition agent is responsible for receiving the request and providing service man- agement. the agent analyzes each request and divides it into the required single services. using a knowledge base, the service dependencies are identified, and a service recovery section extracts similar service information. a composition is successful provided that all of the required single services are available and are used to update the knowledge base. there is a packaging engine that generates a new software package by using existing composition together with the new composition and that registers it in a service catalog. finally, a service delivery section utilizes service catalog information for service provisioning. chunqing, shixing, guopeng, bu sung, and singhal ( ) pro- posed a systematic framework for automatic service conflict detec- tion and supplying policies and user requirements. the first phase of the proposed framework is a conflict analysis section that in- cludes two sub-sections called the comparator and analyzer. the comparator checks the conformity of the policies and user require- ments based on their priorities, and the analyzer is responsible for uncovering the contradictions between the user requirements and their affiliate relations while applying satisfiability modulo theo- ries (smt) (moura & bjørner, ). the filter, allocator, and solver are three parts of the second phase of the framework, which is called the solution derivation. this section finds appropriate single services and eliminates policy-violating services. it determines a set of appropriate single services for each user’s needs and finally concludes with the best composite service with respect to policies a. jula et al. / expert systems with applications ( ) – and requirements. it uses the backtracking algorithm (neapolitan & naimipour, ) for the assigned tasks. according to the definition of trust as a conceptual probability by which a composite service is expected to execute a task as well as expected by the user, a trust-based method and a framework are proposed by xiaona, bixin, rui, cuicui, and shanshan ( ) to solve the ccsc problem. applying the trust in the process of service composition is divided into three parts, including trustworthy in service selection, guaranteeing trust in the composition processes, and trustworthy in binding the generated plan. in the designed framework, the requirement analyzer classifies the requirements of the user into their different elements, including functional and non-functional requirements and expected input–output parame- ters. the service retriever restores information on the services from the resource pool using query requests. inappropriate services are eliminated from the candidate list by a service filter, and the name and type of the remaining services are identified by an wsdl ana- lyzer. the clustering component, template generator, and binding optimizer are responsible for checking the services’ composability, checking the math interfaces, and evaluating the binding plan trust, respectively. cloudrecommender is a cloud-based three-layer structured service composition system that was proposed by zhang, ranjan, nepal, menzel, and haller ( ). the first layer is a configuration management layer in which a cloud service ontology and cloud qos ontology are located because of its two parts for uncovering services based on their functionality and qos parameters; the ser- vices are mapped to a rational model and data structure. applica- tion logic is the second layer, which is implemented to select single services in the form of sql queries to include criteria, views, and stored procedures. the third layer is a widget that divides the user table desired objectives in the investigated researches. reference approach ro ro kofler et al. ( ) cgba p zeng et al. ( ) cgba gutierrez-garcia and sim ( ) cgba zhang and dou ( ) ca p pham et al. ( ) fw wang et al. ( ) cgba p ye et al. ( ) ca zhu et al. ( ) cgba hossain et al. ( ) cgba liu et al. ( ) cgba p liu et al. ( ) cgba worm et al. ( ) cgba zhou and mao ( ) cgba wittern et al. ( ) cgba p jungmann and kleinjohann ( ) ca p ludwig ( ) ca yang et al. ( ) ca bao and dou ( ) mbm p sundareswaran et al. ( ) st chunqing et al. ( ) fw p xiaona et al. ( ) fw p zhang et al. ( ) fw huang et al. ( ) cgba p qi and bouguettaya ( ) cgba wang et al. ( ) ca jula et al. ( ) ca p zhao et al. ( ) ca p dou et al. ( ) cgba wu et al. ( ) cgba p karim et al. ( ) cgba p zibin et al. ( ) cgba wu et al. ( ) fw p lie et al. ( ) ca p p fei tao et al. ( ) ca cgba, classic and graph-based algorithms; ca, combinatorial algorithms; mbm, machin interface into four objects, including the computing resources, storage resources, network resources, and recommendation. this layer is implemented using restful (richardson, ) and several javascript frameworks. a novel framework is proposed for adaptive service selection in mobile cloud computing (wu et al., ). the framework extracts the qos preferences of the customer immediately after receiving a request. then, based on the euclidean distance, some of the nearest customer preference services are selected and suggested to the ser- vice adapter. finally, the service adapter selects the best service among the suggested services to be assigned to the customer, with respect to device context matching and effectiveness of the service option. to reach the context matching service based on the input information, a fuzzy cognitive map model is also utilized in the ser- vice adapter module. the weaknesses of this method include that the proposed framework can only be used to select a single service; furthermore, this strategy has not been compared to other approaches. . discussion . . objectives of the researches to respond to the first research question rq and achieve a comprehensive view of the topic, it is essential to categorize the goals of the researches. objective scrutiny of the considered papers guarantees the existence of nine categories, ro to ro , in which each paper can be placed in one or more categories, as described in table and fig. (a) and (b). based on table , the largest amount of researchers’ attention has been focused on ro , and there is a large difference in terms of attention paid in the ro ro ro ro ro ro ro p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p e-based methods; st, structures; fw, frameworks. a. jula et al. / expert systems with applications ( ) – literature between ro and ro . this difference may be because decreasing the time complexity of the algorithms is a priority of researchers, and user satisfaction is the second-most important objective for researchers. the objective categories of the research are defined as follows: . user requirement satisfaction (ro ). strictly abiding by cus- tomers’ requirements and providing facilities for them to deter- mine or describe their needs would make them more satisfied. . qualitative to quantitative transformation of qos parameters (ro ). when accurately considering the qualitative qos param- eters and applying them in decision making, it is important to transform them into quantitative values using reliable methods. . algorithm improvements (ro ). there are many different heu- ristic and non-heuristic algorithms that are introduced by algo- rithm designers and applied to solve np-hard problems. the following efforts to improve the current algorithms and specify them for ccsc to obtain the best solutions or reduce the execu- tion time is one of the more often investigated objectives in the literature. . introducing data storage and indexing structures (ro ). pos- sessing appropriate and well-defined data structures and dat- abases can play an important role in the design of an efficient algorithm. using a suitable indexing method is also useful in increasing the search speed, especially when the number of cases is very high. . self-adaptability, automaticity, increasing reliability and accu- racy, and quality assurance (ro ). establishing automated and self-adaptable service brokers is unavoidable (denaro et al., ) because of increasing complexity, the number of requests, the number of available services in the pool, their diversity, and the limitation of human abilities. one of the main factors that attracts customers and retains them in utilizing cloud computing is reliability. because of the importance of providing a reliable and self-adaptable service composition organization, the most important part of a cloud is in its direct contact with customers; researchers must consider this direct contact more seriously than before. . proposing an improvised qos mathematical model (ro ). calculating a qos value for composite services requires a math- ematical model in which all aspects, parameters, user require- ments, and tendencies are investigated. to reach this goal, some researchers have attempted to present improved models that focus more on these objectives. fig. . (a) importance percentage of objective categories a . revenue maximization (ro ). encouraging service providers to expose their high-quality services depends on the ability to amass significant profits. thus, revenue maximization can be noted as an important fundamental aim. . optimization of the service discovery process (ro ). if the ser- vice composer policy is not to register services based on prede- fined requirements, then it must discover required available simple services in the network. it is critical to have the type of policy that uses optimal discovery methods. . proposing new frameworks and structures (ro ). reaching some basic goals, e.g., the definition of new roles, requires changes in the framework of the ccsc and in the structures. in some cases, to achieve a goal, there could be a need to design a new frame- work. this scenario has led some researchers to be encouraged to design new frameworks or to change the existing frameworks. . . applied approaches to address rq , all of the proposed or applied approaches are di- vided into five distinct categories, as mentioned in section . . these categories and their statistics are summarized in table and fig. (a) and (b), from which it can be inferred that the highest and lowest percentages of usage can be seen in the classic and graph-based algo- rithms ( %) and combinatorial algorithms ( %), respectively. . . investigated datasets the obtained answers for the third question rq are not prom- ising. possessing different datasets each of them can support sev- eral qos parameters, and the predefined composition problem are useful and unavoidable for evaluating the proposed approaches and comparing their results. unfortunately, the number of datasets that are available to all and in the research domain is very low and is limited to three datasets, qws (al-masri & mahmoud, ), ws-dream (zibin, yilei, & lyu, ) and tpds (zibin et al., ), and an unknown randomly generated dataset rg (shanggu- ang et al., ). researchers have also used a synthetic generator, rarely. the datasets used in each study are listed in table . . . significance of qos parameters and their percentage based on the literature review presented in this paper, the pri- ority and importance percentage of the qos parameters and their nd (b) number of repetitions of objective categories. a. jula et al. / expert systems with applications ( ) – relevant factors have not yet been studied comprehensively. thus, an analysis of the considered qos parameters in the observed pa- pers and their statistics is essential. in this section, we attempted to extract the importance of different qos parameters and their percentage considering this field of research. as specified in table and to answer rq , according to the sig- nificance and priority, most researchers have accounted for differ- ent qos parameters, although others have neglected them. based on the abovementioned parameters and their frequency of occur- rence in the literature, it is important to obtain the most important and effective qos parameters and their importance percentage. to reach this goal, using eq. ( ), the number of occurrences of a parameter has been counted separately and divided by the sum of the number of occurrences of all parameters. the importance per- centage of each parameter is obtained by multiplying its calculated value by . the result obtained is the ratio of the importance percentage of every parameter to all of the other parameters. the number of occurrences of all of the parameters and their impor- tance percentages are shown in fig. (a) and (b). imp percentageðiÞ¼ occurr noðiÞ pparam no j¼ occurr noðjÞ ð Þ table utilized tools, datasets and qos parameters. reference tools dataset qos c kofler et al. ( ) kepler workflow tool, corba, c++ – resp zeng et al. ( ) visual c++ , postgresql . . – resp gutierrez-garcia and sim ( ) not mentioned – not m zhang and dou ( ) not mentioned – not m pham et al. ( ) not mentioned – not m wang et al. ( ) matlab . ws-dream, rg resp relia ye et al. ( ) not mentioned – resp zhu et al. ( ) visual c#.net – not m hossain et al. ( ) not mentioned – resp liu et al. ( ) not mentioned – cost liu et al. ( ) not mentioned ws-dream not m worm et al. ( ) not mentioned – resp zhou and mao ( ) not mentioned – not m wittern et al. ( ) researchers built a tool based on eclipse modeling framework – not m jungmann and kleinjohann ( ) not mentioned – not m ludwig ( ) java – resp yang et al. ( ) not mentioned – not m bao and dou ( ) not mentioned – resp sundareswaran et al. ( ) c – not m chunqing et al. ( ) java, cauldron, zchaff solver – cost xiaona et al. ( ) not mentioned seekda avail zhang et al. ( ) javascript, restful – not m huang et al. ( ) not mentioned – not m qi and bouguettaya ( ) java using synthetic generator resp wang et al. ( ) matlab . , lp-solve . qws, synthetic generator not m jula et al. ( ) visual c#.net ws-dream resp zhao et al. ( ) not mentioned – resp dou et al. ( ) not mentioned – cost, wu et al. ( ) not mentioned – not m karim et al. ( ) not mentioned – cost, relia zibin et al. ( ) planet-lab, axis tpds resp wu et al. ( ) not mentioned – not m lie et al. ( ) not mentioned – not m fei tao et al. ( ) not mentioned – resp func where imp_percentage(i) is the importance percentage of qos parameter i, occur_no(i) is the number of repetitions of qos param- eter i in the investigated papers, and param_no is the number of ob- served qos parameters in the literature. there could be another way to look at the importance percent- age of the qos parameters. it is possible to consider this issue in view of the ratio of the frequency of occurrences of the parameters in the papers. to increase the total number of papers, it is possible to account for all of the papers, including those that did not men- tion any parameters or excluded them. to reach the results, it is sufficient to divide the number of occurrences of each parameter by the total number of papers. the obtained results are shown in fig. (a) and (b) when including and excluding papers that do not consider the qos parameters, respectively. . conclusion and future works showcasing pertinent achievements and findings with a com- prehensive review generally increases the research efforts and pa- pers in that particular scientific field. proposing novel techniques onsidered parameters onse time (rt), cost onse time (rt), availability (avail), reliability (reli) entioned entioned entioned onse time (rt), cost, throughput (throu), reputation (reput), availability (avail), bility (reli) onse time (rt), cost, availability (avail), reputation (reput) entioned onse time (rt), cost entioned onse time (rt), cost, availability (avail) entioned entioned entioned onse time (rt), cost, availability (avail), reliability (reli) entioned onse time (rt), cost, reliability (reli), availability (avail), reputation (reput) entioned ability (avail), durability (dur) entioned entioned onse time (rt), cost entioned onse time (rt), cost onse time (rt), cost, availability (avail), reliability (reli) latency, reputation (reput), entioned response time (rt), security (secur), reputation (reput), availability (avail), bility (reli), durability (dur), data control (dc) onse time (rt), throughput (thr) entioned entioned onse time (rt), cost, reliability reli, energy, trust, maintainability (maint), tion similarity (fs) fig. . (a) approach categories and (b) number of papers in each category. fig. . (a) percentage of repetition of qos parameters and (b) number of repetitions of qos parameters. fig. . (a) percentage of repetition of qos parameters in all investigated papers and (b) percentage of repetition of qos parameters in all investigated papers, excluding those that did not mention any parameters. a. jula et al. / expert systems with applications ( ) – and approaches for addressing cloud computing service composi- tion from different aspects in addition to addressing the lack of comparable activities were the strong motivating factors for pre- paring this systematic literature review. this paper has provided a complete definition of the ccsc in combination with its associ- ated concepts and a comprehensive analysis of the different ap- plied algorithms, mechanisms, frameworks and techniques extracted from authentic published papers, spanning a. jula et al. / expert systems with applications ( ) – – . the achievements of this review shed light on the re- search grounds of ccsc for future studies. this investigation demonstrates that all cloud computing ser- vice composition innovations and improvements can be catego- rized into the following five groups: classic and graph-based algorithms, combinatorial algorithms, machine-based methods, structures and frameworks. the most widely applied category is the classic and graph-based algorithms group ( %), and the least- used categories are machine-based methods and structures ( % and %, respectively). the objectives of these reports can also be di- vided into categories, among which algorithm improvements (ro ) and user requirement satisfaction (ro ) have attracted the most attention ( % and %, respectively), while revenue maximi- zation (ro ) has been the least important objective ( %). counting the number of qos parameter occurrences indicated that different parameters have been considered in the literature, among which service cost and response time are the most repeated ones ( % and %, respectively). calculating the importance per- centage of the qos parameters also revealed that the importance percentages of two mentioned parameters after including and excluding papers that do not consider the qos parameters were % and % and % and %, respectively. to evaluate the pro- posed approaches, qos datasets have been previously identified, ws-dreams , qws, tpds and rg, the first three of which are real-world extracted, and the last of which is generated randomly. synthetic generators and seedka are also two applications that have applied for runtime generating qos values. with respect to the findings in this paper, achieving certain goals is of utmost importance for the planning of future work. aim- ing to prepare an identical, competitive environment for compar- ing the proposed algorithms and approaches, it is indispensable to provide a set of differently sized, standard problems and a com- prehensive qos dataset. the dataset should include a great number of unique services and service providers and encompass cost, re- sponse time, availability, reliability and reputation as significant qos parameters. another essential research goal is to focus on designing comprehensive mathematical models for calculating the qos values composite services that cover all of the involved parameters and their importance percentage. proposing real-time algorithms that can obtain a few optimal composite services for the given requests represent a significant achievement in the field. furthermore, the less considered objectives (e.g., ro , ro , ro , and ro ) should be regarded. finally, acknowledging the stunning growth in mobile computing, future research efforts should be di- rected towards this forward-looking area. references ai, l. f., & tang, m. l. ( ). a penalty-based genetic algorithm for qos-aware web service composition with inter-service dependencies and conflicts. new york: ieee. al-masri, e., & mahmoud, q. h. ( ). investigating web services on the world wide web. in proceedings of the th international conference on world wide web (pp. – ). beijing, china: acm. al-masri, e., & mahmoud, q. h. ( ). discovering the best web service: a neural network-based solution. in systems, man and cybernetics, . smc . ieee international conference on (pp. – ). anselmi, j., ardagna, d., & cremonesi, p. ( ). a qos-based selection approach of autonomic grid services. in proceedings of the workshop on service-oriented computing performance. aspects, issues, and approaches (pp. – ). monterey, california, usa: acm. arapostathis, a., borkar, v. s., fernández-gaucherand, e., ghosh, m. k., & marcus, s. i. ( ). discrete-time controlled markov processes with average cost criterion: a survey. siam journal on control and optimization, , – . ardagna, d., & pernici, b. ( ). adaptive service composition in flexible processes. ieee transactions on software engineering, , – . atashpaz-gargari, e., & lucas, c. ( ). imperialist competitive algorithm: an algorithm for optimization inspired by imperialistic competition. in evolutionary computation, . cec . ieee congress on (pp. – ). bao, h. h., & dou, w. c. ( ). a qos-aware service selection method for cloud service composition. in ieee th international parallel and distributed processing symposium workshops & phd forum (pp. – ). new york: ieee. barney, b. ( ). message passing interface (mpi). in lawrence livermore national laboratory (vol. ). barzegar, s., davoudpour, m., meybodi, m. r., sadeghian, a., & tirandazian, m. ( ). formalized learning automata with adaptive fuzzy coloured petri net; an application specific to managing traffic signals. scientia iranica, , – . bayer, r., & unterauer, k. ( ). prefix b-trees. acm transaction on database systems, , – . borzsony, s., kossmann, d., & stocker, k. ( ). the skyline operator. in data engineering, . proceedings. th international conference on (pp. – ). canfora, g., di penta, m., esposito, r., & villani, m. l. ( ). an approach for qos- aware service composition based on genetic algorithms. new york: assoc computing machinery. chang, w.-l., & yuan, s.-t. ( ). a markov-based collaborative pricing system for information goods bundling. expert systems with applications, , – . chen, l., li, m. l., & cao, j. ( ). eca rule-based workflow modeling and implementation for service composition. ieice transactions on information and systems, e d, – . chou, s.-y., chang, y.-h., & shen, c.-y. ( ). a fuzzy simple additive weighting system under group decision-making for facility location selection with objective/subjective attributes. european journal of operational research, , – . chunqing, c., shixing, y., guopeng, z., bu sung, l., & singhal, s. ( ). a systematic framework enabling automatic conflict detection and explanation in cloud service selection for enterprises. in cloud computing (cloud), ieee th international conference on (pp. – ). de castro, l. n., & von zuben, f. j. ( ). learning and optimization using the clonal selection principle. ieee transactions on evolutionary computation, , – . denaro, g., pezze, m., & tosi, d. ( ). designing self-adaptive service-oriented applications. in autonomic computing, . icac ‘ . fourth international conference on (pp. – ). dillon, t., chen, w., & chang, e. ( ). cloud computing: issues and challenges. in advanced information networking and applications (aina), th ieee international conference on (pp. – ). dou, w., zhang, x., liu, j., & chen, j. ( ). hiresome-ii: towards privacy-aware cross-cloud service composition for big data applications. ieee transactions on parallel and distributed systems, . ellinger, r. s. ( ). governance in soa patterns (white paper). in the northrop grumman corporation for consideration by oasis soa reference architecture team (pp. – ). espadas, j., molina, a., jiménez, g., molina, m., ramírez, r., & concha, d. ( ). a tenant-based resource allocation model for scaling software-as-a-service applications over cloud computing infrastructures. future generation computer systems, , – . fei, t., dongming, z., yefa, h., & zude, z. ( ). resource service composition and its optimal-selection based on particle swarm optimization in manufacturing grid system. ieee transactions on industrial informatics, , – . fei, t., yuanjun, l., lida, x., & lin, z. ( ). fc-paco-rm: a parallel method for service composition optimal-selection in cloud manufacturing system. ieee transactions on industrial informatics, , – . fokkink, w. ( ). introduction to process algebra. springer-verlag new york, inc.. gabrel, v., manouvrier, m., megdiche, i., & murat, c.ieee. ( ). a new - linear program for qos and transactional-aware web service composition. new york: ieee. gao, a. q., yang, d. q., tang, s. w., zhang, m., & society, i. c. ( ). web service composition using integer programming-based models. los alamitos: ieee computer soc. garousi, v., & zhi, j. ( ). a survey of software testing practices in canada. journal of systems and software, , – . gekas, j., & fasli, m. ( ). automatic web service composition based on graph network analysis metrics. in r. meersman, z. tari, m. s. hacid, j. mylopoulos, b. pernici, o. babaoglu, h. a. jacobsen, j. loyall, m. kifer, & s. spaccapietra (eds.). on the move to meaningful internet systems : coopis, doa, and odbase, pt , proceedings ( , pp. – ). berlin: springer-verlag berlin. gerede, e., hull, r., ibarra, o. h., & su, j. ( ). automated composition of e- services: lookaheads. in proceedings of the nd international conference on service oriented computing (pp. – ). new york, ny, usa: acm. gutierrez-garcia, j. o., & sim, k. ( ). agent-based cloud service composition. applied intelligence, , – . gutierrez-garcia, j. o., & sim, k. m. ( ). agent-based service composition in cloud computing. in t. h. kim, s. s. yau, o. gervasi, b. h. kang, a. stoica, & d. slezak (eds.). grid and distributed computing, control and automation ( , pp. – ). berlin: springer-verlag berlin. hamdaqa, m., & tahvildari, l. ( ). cloud computing uncovered: a research landscape. in h. ali & m. atif (eds.), advances in computers ( , pp. – ). elsevier. haykin, s. ( ). neural networks: a comprehensive foundation. prentice hall ptr. he, q., yan, j., jin, h., & yang, y. ( ). adaptation of web service composition based on workflow patterns. in a. bouguettaya, i. krueger, & t. margaria (eds.). service-oriented computing – icsoc , proceedings ( , pp. – ). berlin: springer-verlag berlin. hossain, m. s., hassan, m. m., al qurishi, m., & alghamdi, a. ( ). resource allocation for service composition in cloud-based video surveillance platform. new york: ieee. huang, c.-l., lo, c.-c., chao, k.-m., & younas, m. ( ). reaching consensus: a moderated fuzzy web services discovery method. information and software technology, , – . http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h a. jula et al. / expert systems with applications ( ) – huang, j., liu, y., yu, r., duan, q., & tanaka, y. ( ). modeling and algorithms for qos-aware service composition in virtualization-based cloud computing. ieice transactions on communications, e .b, – . ishizaka, a., & labib, a. ( ). review of the main developments in the analytic hierarchy process. expert systems with applications, , – . jensen, k. ( ). coloured petri nets: basic concepts, analysis methods and practical use ( ). springer-verlag. jeonghwan, j., rothrock, l., mcdermott, p. l., & barnes, m. ( ). using the analytic hierarchy process to examine judgment consistency in a complex multiattribute task. ieee transactions on systems, man, and cybernetics part a: systems and humans, , – . jiang, h., kwong, c. k., chen, z., & ysim, y. c. ( ). chaos particle swarm optimization and t-s fuzzy modeling approaches to constrained predictive control. expert systems with applications, , – . jula, a., & naseri, n. k. ( ). a hybrid genetic algorithm-gravitational attraction search algorithm (hygaga) to solve grid task scheduling problem. in in international conference on soft computing and its applications (icsca’ ) (pp. – ). planetary scientific research center (psrc). jula, a., othman, z., & sundararajan, e. ( ). a hybrid imperialist competitive- gravitational attraction search algorithm to optimize cloud service composition. in memetic computing (mc), ieee workshop on (pp. – ). jungmann, a., & kleinjohann, b. ( ). towards the application of reinforcement learning techniques for quality-based service selection in automated service composition. in services computing (scc), ieee ninth international conference on (pp. – ). kaelbling, l. p., littman, m. l., & moore, a. w. ( ). reinforcement learning: a survey. journal of artificial intelligence research, , – . karim, r., chen, d., & miri, a. ( ). an end-to-end qos mapping approach for cloud service selection. in services (services), ieee ninth world congress on (pp. – ). kitchenham, b., & charters, s. ( ). guidelines for performing systematic literature reviews in software engineering. ( . ed., pp. – ). keele university and durham university. klein, a., ishikawa, f., & honiden, s. ( ). towards network-aware service composition in the cloud. in the st international conference on world wide web (pp. - ). lyon, france: acm. knuth, d. e. ( ). the art of computer programming. volume. a: combinatorial algorithms, part , pearson education india. kofler, k., haq, i. u., & schikuta, e. ( ). user-centric, heuristic optimization of service composition in clouds. lncs, , – . kofler, k., ul haq, i., & schikuta, e. ( ). a parallel branch and bound algorithm for workflow qos optimization. in parallel processing, . icpp ‘ . international conference on (pp. – ). korkmaz, t., & krunz, m. ( ). multi-constrained optimal path selection. in infocom . twentieth annual joint conference of the ieee computer and communications societies. proceedings ( , pp. – ). ieee. korte, b., & vygen, j. ( ). linear programming ( ). berlin heidelberg: combinatorial optimization springer, pp. - . koshy, t. ( ). chapter – formal languages and finite-state machines. in discrete mathematics with applications (pp. – ). burlington: academic press. kosuga, m., kirimoto, n., yamazaki, t., nakanishi, t., masuzaki, m., & hasuike, k. ( ). a multimedia service composition scheme for ubiquitous networks. journal of network and computer applications, , – . kou, l. t., & markowsky, g. ( ). multidimensional bin packing algorithms. ibm journal of research and development, , – . abonyi, j., akerkar, r., alavi, a. h., arango, c., aydogdu, i., brest, j., et al. ( ). list of contributors. in x.-s. yang, z. cui, r. xiao, a. h. gandomi & m. karamanoglu (eds.), swarm intelligence and bio-inspired computation (pp. xv-xviii). oxford: elsevier. li, d., cheung, d., shi, x., & ng, v. ( ). uncertainty reasoning based on cloud models in controllers. computers & mathematics with applications, , – . li, z., fang, h., & xia, l. ( ). increasing mapping based hidden markov model for dynamic process monitoring and diagnosis. expert systems with applications, , – . liangzhao, z., benatallah, b., ngu, a. h. h., dumas, m., kalagnanam, j., & chang, h. ( ). qos-aware middleware for web services composition. ieee transactions on software engineering, , – . liao, j. x., liu, y., zhu, x. m., xu, t., & wang, j. y.ieee. ( ). niching particle swarm optimization algorithm for service composition. in ieee global telecommunications conference. new york: ieee. lie, q., yan, w., & orgun, m. a. ( ). cloud service selection based on the aggregation of user feedback and quantitative performance assessment. in services computing (scc), ieee international conference on (pp. – ). lin, w., dou, w., luo, x., & jinjun, c. ( ). a history record-based service optimization method for qos-aware service composition. in web services (icws), ieee international conference on (pp. – ). liu, f., & zeng, g. ( ). study of genetic algorithm with reinforcement learning to solve the tsp. expert systems with applications, , – . liu, h., zheng, z. b., zhang, w. m., & ren, k. j. ( ). a global graph-based approach for transaction and qos-aware service composition. ksii transactions on internet and information systems, , – . liu, m., wang, m. r., shen, w. m., luo, n., & yan, j. w. ( ). a quality of service (qos)-aware execution plan selection approach for a service composition process. future generation computer systems-the international journal of grid computing and escience, , – . liu, s., xiong, g., zhao, h., dong, x., & yao, j. ( ). service composition execution optimization based on state transition matrix for cloud computing. in intelligent control and automation (wcica), th world congress on (pp. - ). liu, y., li, m., & wang, q. ( ). a novel user-preference-driven service selection strategy in cloud computing. international journal of advancements in computing technology, , – . ludwig, s. a. ( ). clonal selection based genetic algorithm for workflow service selection. in evolutionary computation (cec), ieee congress on (pp. – ). luo, y. s., qi, y., hou, d., shen, l. f., chen, y., & zhong, x. ( ). a novel heuristic algorithm for qos-aware end-to-end service composition. computer communications, , – . milanovic, n., & malek, m. ( ). current solutions for web service composition. ieee internet computing, , – . moscato, p., & cotta, c. ( ). a gentle introduction to memetic algorithms. in f. glover & g. kochenberger (eds.), handbook of metaheuristics ( , pp. – ). us: springer. moura, l. d., & bjørner, n. ( ). satisfiability modulo theories: introduction and applications. communication acm, , – . neapolitan, r., & naimipour, k. ( ). foundations of algorithms ( ). jones and bartlett publishers, inc (december , ). peter mell, t. g. ( ). the nist definition of cloud computing. in n. i. o. s. a. technology (ed.): u.s. department of commerce. pham, t. v., jamjoom, h., jordan, k., & shae, z.-y. ( ). a service composition framework for market-oriented high performance computing cloud. in proceedings of the th acm international symposium on high performance distributed computing (pp. – ). chicago, illinois: acm. preve, n. p. ( ). grid computing: towards a global interconnected infrastructure. springer. qi, y., & bouguettaya, a. ( ). efficient service skyline computation for composite service selection. ieee transactions on knowledge and data engineering, , – . qiang, d., yuhong, y., & vasilakos, a. v. ( ). a survey on service-oriented network virtualization toward convergence of networking and cloud computing. ieee transactions on network and service management, , – . reynolds, r. g. ( ). cultural algorithms: theory and applications. in c. david, d. marco, g. fred, d. dipankar, m. pablo, p. riccardo, & v. p. kenneth (eds.), new ideas in optimization (pp. – ). uk: mcgraw-hill ltd.. richardson, l. ( ). restful web services ( ). o’reilly media. sayed, a. h. ( ). fundamentals of adaptive filtering. wiley. schmid, s., chart, t., sifalakis, m., & scott, a. ( ). flexible, dynamic, and scalable service composition for active routers. in proceedings of the ifip-tc th international working conference on active networks (pp. – ). springer- verlag. shangguang, w., zibin, z., qibo, s., hua, z., & fangchun, y. ( ). cloud model for service selection. in computer communications workshops (infocom wkshps), ieee conference on (pp. – ). shrme, a. e. ( ). hybrid intelligent technique for automatic communication signals recognition using bees algorithm and mlp neural networks based on the efficient features. expert systems with applications, , – . singh, m. p. ( ). physics of service composition. ieee internet computing, , – . sinnema, m., & deelstra, s. ( ). classifying variability modeling techniques. information and software technology, , – . smith, r. g. ( ). correction to the contract net protocol: high-level communication and control in a distributed problem solver. ieee transactions on computers, c- , . song, x., dou, w., & chen, j. ( ). a workflow framework for intelligent service composition. future generation computer systems, , – . song, x. d., dou, w. c., & chen, j. j. ( ). a workflow framework for intelligent service composition. future generation computer systems-the international journal of grid computing and escience, , – . song, z., geng, x., kusiak, a., & xu, c. ( ). mining markov chain transition matrix from wind speed time series data. expert systems with applications, , – . stewart, b. j. ( ). leveraging amazon web services for enterprise application integration ( ). ibm corporation, developerworks, pp. – . strunk, a. ( ). qos-aware service composition: a survey. in web services (ecows), ieee th european conference on (pp. – ). sundareswaran, s., squicciarini, a., & lin, d. ( ). a brokerage-based approach for cloud service selection. in cloud computing (cloud), ieee th international conference on (pp. – ). takabi, h., joshi, j. b. d., & gail-joon, a. ( ). security and privacy challenges in cloud computing environments. ieee security and privacy, , – . tang, m. l., & ai, l. f.ieee. ( ). a hybrid genetic algorithm for the optimal constrained web service selection problem in web service composition. new york: ieee. torkashvan, m., & haghighi, h. ( ). a greedy approach for service composition. new york: ieee. vanderbei, r. j. ( ). linear programming: foundations and extensions. springer london, limited. vaquero, l. m., rodero-merino, l., caceres, j., & lindner, m. ( ). a break in the clouds: towards a cloud definition. sigcomm compututer communication review, , – . wang, j.-z., wang, j.-j., zhang, z.-g., & guo, s.-p. ( ). forecasting stock indices with back propagation neural network. expert systems with applications, , – . http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h a. jula et al. / expert systems with applications ( ) – wang, s., sun, q., & yang, f. ( ). towards web service selection based on qos estimation. internatinal journal of web grid services, , – . wang, s. g., sun, q. b., zou, h., & yang, f. c. ( ). particle swarm optimization with skyline operator for fast cloud-based web service composition. mobile networks & applications, , – . wang, y. w. ( ). application of chaos ant colony algorithm in web service composition based on qos. los alamitos: ieee computer soc. wischik, d., handley, m., & braun, m. b. ( ). the resource pooling principle. sigcomm computer communication review, , – . wittern, e., kuhlenkamp, j., & menzel, m. ( ). cloud service selection based on variability modeling. lncs, , – . worm, d., zivkovic, m., van den berg, h., & van der mei, r. ( ). revenue maximization with quality assurance for composite web services. in service- oriented computing and applications (soca), th ieee international conference on (pp. – ). wu, q., zhang, m., zheng, r., lou, y., & wei, w. ( ). a qos-satisfied prediction model for cloud-service composition based on a hidden markov model. mathematical problems in engineering, , . xiangkun, t., & yi, g. ( ). a cosine theorem based algorithm for similarity aggregation of ontologies. in signal processing systems (icsps), nd international conference on vol. (pp. v – -v - ). xiaona, w., bixin, l., rui, s., cuicui, l., & shanshan, q. ( ). trust-based service composition and optimization. in software engineering conference (apsec), th asia-pacific vol. (pp. – ). yakimenko, o. a., slegers, n. j., bourakov, e. a., hewgley, c. w., bordetsky, a. b., jensen, r. p., robinson, a. b., malone, j. r., & heidt, p. e. ( ). mobile system for precise aero delivery with global reach network capability. in control and automation, . icca , ieee international conference on (pp. – ). yang, y., mi, z., & sun, j. ( ). game theory based iaas services composition in cloud computing environment. advances in information sciences and service sciences, , – . ye, z., zhou, x., & bouguettaya, a. ( ). genetic algorithm based qos-aware service compositions in cloud computing. in j. yu, m. kim, & r. unland (eds.). database systems for advanced applications ( , pp. – ). berlin heidelberg: springer. yi, w., & blake, m. b. ( ). service-oriented computing and cloud computing: challenges and opportunities. ieee internet computing, , – . yu, q., & bouguettaya, a. ( ). computing service skyline from uncertain qows. ieee transactions on services computing, , – . yu, t., & lin, k.-j. ( ). service selection algorithms for composing complex services with multiple qos constraints. in proceedings of the third international conference on service-oriented computing (pp. – ). amsterdam, the netherlands: springer-verlag. zeng, c., guo, x. a., ou, w. j., & han, d. ( ). cloud computing service composition and search based on semantic. in m. g. jaatun, g. zhao, & c. rong (eds.). cloud computing, proceedings ( , pp. – ). berlin: springer-verlag berlin. zeng, j., duan, j., & wu, c. ( ). a new distance measure for hidden markov models. expert systems with applications, , – . zhang, m., ranjan, r., nepal, s., menzel, m., & haller, a. ( ). a declarative recommender system for cloud infrastructure services selection. in proceedings of the th international conference on economics of grids, clouds, systems, and services (pp. – ). berlin, germany: springer-verlag. zhang, x. y., & dou, w. c. ( ). preference-aware qos evaluation for cloud web service composition based on artificial neural networks. in f. l. wang, z. g. gong, x. f. luo, & j. s. lei (eds.). web information systems and mining ( , pp. – ). berlin: springer-verlag berlin. zhao, s. s., ma, l., wang, l., & wen, z. p. ( ). an improved ant colony optimization algorithm for qos-aware dynamic web service composition. los alamitos: ieee computer soc.. zhao, x., wen, z., & li, x. ( ). qos-aware web service selection with negative selection algorithm. knowledge and information systems, – . zhao, x. c., song, b. q., huang, p. y., wen, z. c., weng, j. l., & fan, y. ( ). an improved discrete immune optimization algorithm based on pso for qos- driven web service composition. applied soft computing, , – . zhou, w., & bovik, a. c. ( ). mean squared error: love it or leave it? a new look at signal fidelity measures. ieee signal processing magazine, , – . zhou, x., & mao, f. ( ). a semantics web service composition approach based on cloud computing. in computational and information sciences (iccis), fourth international conference on (pp. – ). zhu, y., li, w., luo, j., & zheng, x. ( ). a novel two-phase approach for qos- aware service composition based on history records. in service-oriented computing and applications (soca), th ieee international conference on (pp. – ). zibin, z., xinmiao, w., yilei, z., lyu, m. r., & jianmin, w. ( ). qos ranking prediction for cloud services. ieee transactions on parallel and distributed systems, , – . zibin, z., yilei, z., & lyu, m. r. ( ). distributed qos evaluation for real-world web services. in web services (icws), ieee international conference on (pp. – ). zissis, d., & lekkas, d. ( ). addressing cloud computing security issues. future generation computer systems, , – . http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h http://refhub.elsevier.com/s - ( ) - /h cloud computing service composition: a systematic literature review introduction survey goals and execution . survey goals and research questions . publication statistics cloud computing . cloud definition . cloud computing characteristics . cloud computing service models . cloud computing deployment models the cloud computing service composition problem (ccsc) . ccsc definition . ccsc challenges . current approaches for ccsc . . classic and graph-based algorithms . . combinatorial algorithms . . machine-based methods . . structures . . frameworks discussion . objectives of the researches . applied approaches . investigated datasets . significance of qos parameters and their percentage conclusion and future works references sddhand, vol . , part , june , pp. - . © printed in india. expert systems and finite element structural analysis - a review b p naganarayana and g prathap structural sciences division, national aeronautical laboratory, bangalore , india ms received april ; revised december ,abstract . finite element analysis of many engineering systems is practised more as an art than as a science . it involves high level expertise (analytical as well as heuristic) regarding problem modelling (e .g. problem specification, choosing the appropriate type of elements etc .), optical mesh design for achieving the specified accuracy (e .g . initial mesh selection, adaptive mesh refinement), selection of the appropriate type of analysis and solution routines and, finally, diagnosis of the finite element solutions . very often such expertise is highly dispersed and is not available at a single place with a single expert. the design of an expert system, such that the necessary expertise is available to a novice to perform the same job even in the absence of trained experts, becomes an attractive proposition . in this paper, the areas of finite element structural analysis which require experience and decision-making capabilities are explored . a simple expert system, with a feasible knowledge base for problem modelling, optimal mesh design, type of analysis and solution routines, and diagnosis, is outlined . several efforts in these directions, reported in the open literature, are also reviewed in this paper. keywords. finite element method ; optimal mesh design; structural analysis; expert system; knowledge base. the recent trend in the field of software development for scientific and engineering applications is to relieve the lay user of the need for expert knowledge necessary for the optimal use of the software as well as the theoretical and mathematical background for using the code . for this purpose, a combination of knowledge-based expert system (kbes) control and the automation of data generation and problem optimisation (e .g . adaptive mesh generation), with the computational method of solving the problem (e .g . finite element method), can make the entire system a powerful, efficient and cost- erective tool for design, analysis and diagnosis of complex problems . today, the finite element method has become an essential (and rather inevitable) tool for modelling and evaluating the physical performance of structural, mechanical and thennal systems in several engineering disciplines -aerospace, mechanical, civil, b p naganarayana and g prathap nuclear, marine etc . it has considerable impact on other engineering fields like electromagnetism and fluid mechanics as well . it has become an integral part of the present day cad/cam cycles, and has even been used to simulate manufacturing processes like casting, forging etc . the scope of the present paper, however, is confined to the role expert systems can play to simplify finite element structural analysis. structural analysis is the process of determining the response of a fully specified structure to its environment . full specification of a structure needs a priori description of the topology of the structural domain, the material properties and the boundary conditions . the loads acting on the structure - including the equivalent loads produced by thermal effects, material changes like shrinkage and creep, initial strain/stress etc . - form the environment. the structural deformation and stresses (or stress resultants) are the standard response quantities of interest . the finite element analysis of a structural system has attained a high level of sophistication and involves a considerable amount of experience and judgement in taking a priori and/or a posteriori decisions regarding modelling of the system with proper elements, data management for optimum mesh size for achieving the desired accuracy, choice of the type of analysis required and choice of the correct combination of the solution tools and finally the diagnosis of the results obtained . a great deal of expertise and a considerable amount of heuristic knowledge are essential for such an exercise . today, with the availability of numerous general purpose finite element programs (e .g . nastran, aska, sap, nisa, ansys, cosmos, marc etc .), coupled with their broad range of capabilities, the need for expertise in finite element analysis has been multipled . often an experienced analyst might have attained expertise in one field of application using only one such general purpose software packages . thus, apart from the novice user, even the experienced user often confronts difficulties while modelling a new class of problems with a new package . thus, one of the major requirements is to accumulate, organise and thereby transfer the expertise of such experienced users to the less experienced ones, for the same class of problems and the same finite element programs, and, eventually aggregate such expertise over many different classes of problems and many different finite element packages . knowledge-based expert systems hold the promise of providing a methodology for such finite element modelling aids . often these are called intelligent pre- and post , processors (adeli ), in the sense that they act as knowledge-based interfaces to the algorithmic programs . architecturally, such expert systems are closely related to the so-called classification or diagnostic systems (harmon & king ) since they are data-driven (either input or generated data). several expert systems have been developed in practice with reference to modelling structural problems using some specific finite element programs (bennett et al ; bennett & engelmore , pp . - ; rivlin et al ; corlett et al ; fenves , ; holt & narayana ; rehak ; taig ; wilson & itoh ; logo& genberg ; chen & hajela ). sacon (bennett et al ; bennett & eng , pp. - ) is one of the earliest . it used the popular diagnostic expert system shell emycin (zumsteg & flaggs ) and interacted with the user for proper application °f the marc finite element program for structural analysis (bennett et al , bennett & engelmore , pp . - ; rivlin et al ; fjellheim & syversen ) . most are knowledge-based consultation systems and are developed for experi mental purposes . their applications are often confined to initial modelling of the problem for particular finite element program . expert systems and finite element analysis another area of the finite element method which demands experience, judgement and heuristic knowledge is that of identifying the singularities present, the locations of stress concentrations and their relative strengths and the designing of an optimal mesh for achieving the specified accuracy. several strategies of adaptive refinement for this purpose (e.g. increasing the number of elements of the same order, h-type, or increasing the order of the element interpolations, p-type, or the combination of the two, hp-type, refinement) are developed in the literature (babuska & rheinboldt ; atluri et al ; babuska et al ; zienkiewicz & zu ). recently, a few attempts have also been made towards developing expert-like systems to identify the singularities in the structure and arrive at optimal refined mesh configurations (babuska & rheinboldt ; babuska & rank ; rank & babuska ; szabo, probe) . these systems mainly deal with the rules for adaptive refinement based on analytical error norms (e .g. strain energy norms, stress gradient norms etc .) and heuristic knowledge about the strength of the singularities and/or stress-concentrations . however, expert systems for such tasks are not yet well-established in practice. in this paper, the role that expert systems can play in general finite element structural analysis is discussed . an overview of several aspects of finite element analysis which need both theoretical/analytical and heuristic expertise is given . finally a proposal for a simple expert-system architecture for such a task is presented. this architecture was the basis for building an experimental expert-system module for problem modelling and optimal mesh design for a special purpose -d finite element package at the national aeronautical laboratory in bangalore (prathap & naganarayana ) . this is discussed briefly as a typical case study . . finite element analysis - scope for expert systems the main areas of finite element analysis which demand a high level of experience and judgement capabilities (figure ) are problem modelling (e.g. problem specification, choosing the appropriate type of elements etc .), optimal mesh design for achieving domain topology analysis pro bi em mode i i i n g optimisation diagnostics do porn specification environment specific of ion error norms p r a b i e m adoptive refinement singularities material properties boundoy conditions ~---{ type of analysis type of solution figure . scope of expert systems in finite element analysis . loading conditions . b p naganarayana and g prathap specified accuracy, selection of appropriate type of analysis and solution routines and, finally, diagnosis of the finite element solutions . . problem modelling modelling any problem requires expertise in selecting and/or development of the optimal theoretical basis (e .g. plate theories, shell theories etc.) and hence the type of elements to be used . considerable knowledge regarding topological configuration and material constitution of the structure, boundary conditions and loading environment is required for this purpose . a decision on the theoretical basis is crucial for making the analysis meaningful, optimally accurate as well as cost-effective. it should be noted that the cost of finite element analysis is directly related to the continuity requirements and the total degrees of freedom used for modelling the problem (table ) . an intelligent use of the available and/or generated knowledge is necessary to decide whether -dimensional, - dimensional or -dimensional modelling is appropriate for the problem, depending on the geometrical configuration and the material constitution of the structure, the boundary conditions and the loading environment . if -d or -d analysis seems sufficient with some loss of accuracy and/or compatibility in the displacement and the stress recovery, one has to decide whether a refined higher order shell, plate or beam theory is necessary for achieving the required accuracy . in addition, the degree of importance and risk associated with the type of application can also dictate the type of theoretical support required . it should be kept in mind that the time and cost of analysis increases sharply with an increase in the total degrees of freedom and continuity requirements for modelling the structure (table ). thus, -d analysis may be unwarranted, for example, when one of the dimensions of the domain is `small' as compared with the others . if this is not done, the cost of computation increases dramatically requiring much more computer time and larger storage requirements . other deciding factors for choosing -d, -d or -d elements are the surface boundary conditions and loading. for example if adjacent surfaces of a prismatic slab are completely or partially supported or loaded, -d analysis may become mandatory. often, depending on the geometrical and material properties of the table . comparison of different theories regarding the total d.o .f. and continuity requirement . total d .o .f. ( nodes per edge of hexa- d .o.f. hedral domain) per node , , , , expert systems and finite element analysis structure, the boundary conditions and the loading environment, one can reduce a -d problem to a -d problem in a plane-stress or plane-strain sense (timoshenko & goodier ) . accordingly, a -dimensional plane-stress or a plane-strain element may suffice for the modelling requirements . depending on the loading and the type of element choosen, only the active components of strain energy need to be computed e.g . for purely in-plane loading on a plane-stress model of a plate, one requires to compute only the membrane component of the strain energy and the in-active degrees of freedom (transverse deflection and all rotations) should be suppressed . such decisions reduce the computer time and cost effectively, in particular, when one deals with large problems . very often, prescribing the boundary conditions in idealised terms, e .g . simple support or clamped, can be unrealistic and unreliable . boundary conditions become much more complex when they become nonlinear. on the other hand, for some problems, one may go for a combination of simple boundary conditions without losing accuracy of solution. similarly, modelling of loads acting on a structure can also be non-unique. normally, the general loads acting upon a structure (e.g. airplane wing, landing gears etc .) do not fit into any standard category . in such cases, an expert may arrive at a decision regarding modelling of the boundary conditions and the loads based on the structure, its environment, his experience and intuition, and his knowledge of the information available in the literature . hence, in such cases, it may be preferable to go according to the rule of thumb followed by experts in the relevant fields of application . recently, daniel ( ) has proposed an expert system for choosing proper boundary conditions for a given structure. sometimes, the global geometry of the structure and the possible local singularities may lead to conflicting opinions about using any particular theory for modelling and/or the method of solution . for example, a -d structural domain with a small, but critical, crack embedded deep in the domain may just require a global -d analysis and a local -d analysis around the crack . sometimes, for a small crack in a relatively large -d or -d structural domain for which only the global behaviour of the structure is of interest, it may be desirable to have a combination of the boundary element and the finite element methods in modelling the problem . this achieves a cheap but sufficiently accurate solution . on the other hand, if the strength of the singularity is low enough, a refined mesh may be suf . tent for achieving the specified accuracy. problem modelling cannot be complete without defining proper correlation (geometrical as well as material) between any two points in the domain . experience is required in choosing the interpolation functions for generating the geometrical configuration of the structure (e .g. automatic volume, surface and line generation) from the minimum data input . this is turn can play a decisive role in choosing the type and order of the elements (as is apparent above) and the type of the basic mesh required. material modelling requires special attention, in particular when the structure is urinated . material correlation in the thickness direction of the laminate should be evolved from the lay-up data (e .g . thickness, orientation and material properties of each layer for each of the orthotropic layers and the global constitutive relationships for each of the anisotropic layers) from which one has to deduce knowledge on the type of the laminate structure e .g . symmetric. antisymmetric, cross-ply etc . accordingly computations can be optimised later by generating only the required data and following only the required steps of analysis . dimension theoretical basis continuity requirement -d elementary c' first order c° third order c° -d elementary c' first order c° third order-i c° 'third order-i] c° -d c° b p naganarayana and g prathap . mesh optimisation mesh optimisation gains importance, in particular, when singularities are embedded in the problem domain. in general, the singularities create error-prone zones in their proximity. this demands data refinement, localised to such zones, for achieving the specified accuracy of solution in an optimal sense. the domain singularities, with reference to solid and structural mechanics, fall in one of the four general categories: material singularities - hard grains, material voids etc . geometric singularities - cracks, notches, voids etc . boundary conditions - point supports, discontinuous supports etc. loading environment - concentrated forces, discontinuously distributed loads etc . often, an expert uses his experience, and the analytical as well as heuristic knowledge available in the literature with reference to both the method of analysis and the problem under consideration so as to optimise the data refinement in the error-prone zones to achieve accuracy of solution in an efficient and cost-effective manner . sometimes it may not be convenient to determine the strength of the singularities analytically . in such cases, the rule of thumb followed by the acknowledged experts in the corresponding field may have to be sought for mesh optimisation . in general, it is impossible to construct an a priori error norm for finite element modelling of a structure of complex geometry . in such cases one has to use a posteriori information from cheap coarse mesh computations of the model, construct the error norms, identify the error-prone zones and optimise the mesh appropriately to achieve the specified accuracy in the most efficient and cost-effective manner. a considerable amount of research and development has taken place in literature with reference to the error norms for finite element analysis of structure with singularities and several adaptive mesh refinement techniques (e .g. h-type, p-type and hp-type) have been proposed to achieve the specified accuracy (babuska & rheinboldt ; atluri et al ; babuska et al ; zienkiewicz & zu ) . recently some efforts have been made toward developing some expert-like systems for co-ordinating such activities in finite element analysis (babuska & rank ; rank & babuska ; szabo (year not given) . however, the resulting expert systems are primitive in the sense that they only consider the analytical knowledge with reference only to the method of analysis (i .e . adaptive finite element methods). . type of analysis and solution another area in which a great deal of expert knowledge is essential is the understanding of how a structure with certain geometric and material properties responds to a given load under given boundary conditions . the major parameters involved in taking such decisions are the material properties, the domain geometry, the boundary conditions and the loads. one has to have a clear understanding of the independent as well as the collective effects of these parameters on structural behaviour before deciding on the type of analysis required . knowledge of the microstructural as well as the macrostructural properties of th e material is required in deciding whether the material behaviour is linear or nonlinear , elastic or plastic, compressible or incompressible etc . one has to have an a priori knowledge of whether there are certain zones in the domain which require nonlinea r elastic or plastic analysis though the global structural behaviour is linearly elastic for example, the zone in the proximity of the crack tips in an elastic domain ma y expert systems and finite element analysis require plastic analysis in a local sense, in particular, under fatigue load . such knowledge becomes important, for the design/analysis of failure of a structural component. sometimes the loads can also lead to such requirements, though the material is linearly elastic, depending on their magnitude . normally a material database is required for such purpose. the next important aspect is about the geometric nonlinear behaviour of a structure . based on the structural geometry, boundary supports and the type of loads on the structure, one has to decide whether the structure is undergoing relatively small or large deflections and accordingly choose the theory for the analysis . the characteristic properties that influence geometric nonlinearity of a structure are mainly the structural geometry and magnitude and direction of the load . a good example of geometric nonlinearity is seen in the buckling of structures . often such problems are appro- ximately solved as an eigenvalue problem for certain simple cases (e.g. buckling of beams) with reasonably accurate solutions . on the other hand, the true geometric nonlinear considerations often become mandatory for analysing buckling of complex structures like shells . normally one needs a posteriori information about the deflections and stresses for taking a decision on nonlinear analysis for a structure, in particular if deformation or strain under load is large . if the deflection at any point of the structure is of the same order as the minimum characteristic structural dimension (e.g . thickness of a plate) one may need geometric nonlinear analysis . on the other hand, if any of the stress components is close to the corresponding elastic design limits for the material (e.g. corresponding yield stress) material nonlinear analysis may be required . again this material non linearity can fall into any category like elasto-plastic, visco-elastic, plastic etc. depending on the structure and its environment . often, an expert uses his past experience and depends on several heuristic rules for taking such decisions . apart from the magnitude, location and direction of the loads which normally influence the geometric (sometimes material) nonlinearity, the type of the loads with respect to time also influences the type of analysis required . expert decisions have to be taken about whether a static analysis is enough or a dynamic analysis is required depending on the type of loads . if dynamic analysis is required, the type of analysis again influenced by the type of application as well as the duration of application the load (e .g . eigenvalue analysis, frequency response analysis, impact load analysis, tier analysis, fatigue damage analysis etc .). the presence of certain singularities like a crack may require special knowledge h as fracture mechanics - which may again be linear or nonlinear - to be rporated into the analysis system . the analysis may differ for different types of ks . the analysis of crack initiation and propagation becomes very important, in cular, for enduring dynamic loads causing fatigue in the material structure . depending on the type of numerical method and the type of analysis one has to an optimum combination of solution routines . exact or direct solution hniques may be desirable in conventional finite element modelling of linear beems, probably enhanced with an `out-of-core' algorithm for very large problems . the other hand, iterative solution routines may be preferable for conventional e element modelling of nonlinear problems (e .g . newton-raphson method) and adaptive unite element modelling (e .g. element-by-element conjugate gradienth with hierarchical preconditioning of the stiffness matrix) . for dynamic analysis may need additional solution routines like time-step integration, eigenvalue tion etc . depending on the type of analysis required . 'tall y, expertise is needed in determining the tools to be used for tackling the b p naganarayana and g prathap problem in numerical and logical manner, for example, what combination of computer languages, what sort of data management, how the already available software can be optimally used etc . today, most of the expert systems are written in c-language owing to its flexibility and efficiency in handling both numerical as well as logical operations . blackboard architecture (nilssion ; nii ) is popular for the management of data and computational as well as logical software . . diagnostics the usual practice of analysis, especially for a complex problem, is to start with a . simplified approximate model and then choosing optimally efficient combinations of more sophisticated tools of analysis after an intelligent diagnosis of the initial results. interpretation of the final results also demands high levels of intelligence and expertise. thus, it is often the diagnosis and interpretation of the finite element results for a complex problem that calls for high levels of intelligence, expertise and judgement capabilities . on the other hand, this is the most difficult part to code, since such knowledge is normally heuristic and changes from person to person, context to context and finally organisation to organisation (depending on the policies) . in addition, most of the expertise on diagnosis is highly problem-dependent and cannot be generalised for all finite element applications as such. therefore, developing a general purpose expert system for diagnosis of finite element results may not be a practical proposition and in fact hardly any effort has been documented in this area in the open literature . thus one may have to confine the system to tackling specific problems like analysing rocket tanks (desaleux & fouet ) etc . using an appropriate finite element package. . expert systems as discussed earlier, if the expertise has to be transparent to a novice user of an analysis package, it becomes mandatory to coordinate the knowledge regarding the package and the field of interest and make it readily available . again, as explained earlier, an expert system is a very useful tool for such an exercise. problem in addition, a knowledge-based expert system combined with automated p modelling and problem optimisation can make the analysis process very efficient and cost-effective when the process is of repetitive nature (figure ) . such a system can have many advantages over conventional numerical analysis . once the expert system is calibrated by the expert, he can release the software for use to less qualified assistants and coordinate several different tasks simultaneously. since the algorithmic comp tions are also optimised, the total time and cost of the whole analysis process is drastically reduced . finally, such systems can be effectively used to train novice users in specific fields . various interpretations and definitions of expert systems can be found in the literature . for example : "an interactive computer program incorporating judgement, experience, rules of thumb, intuition and other expertise to provide knowledgeable advice about "variety of tasks" (feigenbaum ). r ocedure to "an intelligent computer program that uses knowledge and inference p ertisef or solve problems that are difficult enough to require significant human exp their solution" (gasching et at ). expert final analysis report expert systems and finite element analysis ( problem assignment ) / ( results and diagnostics ) calibration establish es no ledge analysis module i automated data generation / refinement i expert system use figure . role of expert systems in finite element analysis. "an expert system solves real-world, complex problems using a complex model of expert human reasoning, reaching the same conclusions that the human expert would reach if faced with a comparable problem" (weiss & kulikowski ). keeping the dominantly computational nature of the finite element analysis in mind, the last interpretation for the expert system appears to be most appropriate in the present context . in this section, the aspects involved in building an expert system are discussed with particular reference to finite element analysis of structural problems . it is convenient to handle an expert system if its frame is divided into a domain- dependent part, the knowledge base and a domain-independent part, the inference engine (adeli ; levine et al ). the inference engine forms the context, working memory and decision paths for the system while the knowledge-base provides the technical informations and rules for reaching the specified goals in a particular domain of application, as an expert would do. essentially, the inference engine coordinates different tasks (both computational and logical) and the knowledge base, normally constituted by database and rulebase, forms the body of knowledge for executing the tasks identified by the inference engine . the general configuration for an expert system for finite element analysis is shown )n figure . it consists of an inference engine having two parts . the logical modulec oordinates the information provided by the knowledge-base, database, rule-base and user interaction enhanced with an expert advice facility to select the appropriate finitee lements, methods/types of analysis, a coarse initial mesh etc . essentially, this module produces an optimum initial model for the given problem . on the other hand, the conputational module coordinates the computational, algorithmic and graphical toftware through a recurring cycle of finite element analysis, error estimations and mesh refinement resulting in an optimally accurate solution for the problem . this iiodule essentially involves problem optimisation and solution . finally the system n be enhanced with the assistance of post-processing and diagnosis facilities . other major modules in such an expert system are the advise/help, explanation and subordinates know ledge acquisition module novice or experts database b p naganarayana and g prathap rule-base l diagnostics inference mechanism knowledge base advice / help facility gicai module use r procedural software computational module expi anation facility knowledge acquisition t explanation facility learning facility post - processor diagnostics advice/help facility figure . a typical expert system for finite element analysis . learning facilities. figure shows the details of the design of such an expert adaptive finite element structural analysis system. . knowledge base knowledge can be categorised into two types. surface knowledge consists of rules regarding the necessary tasks, data input and the sequence of the tasks for the given particular problem (e .g. determining relative dimensions of the structure) . this can be normally handled by a rule-based expert system through a user-interactive logical questionnaire. on the other hand, deep knowledge involves knowledge about the use r procedural software line, surface and -- volume generator mesh generator i . element conne t"' boundary and load conditions f .e . package error estimators solution routine graphics pack ag e post-proces sor figure . blackboard architecture for expert system in adaptive finite element analysis . expert systems and finite element analysis underlying problem structure (e.g . error estimation and recognition of critical zones) . each can be declarative (facts about objects, events, e .g . whether strain energy density at a point is high or low) or procedural (information about courses of action e .g. the mesh refinement strategy depending on the error norms) . surface knowledge is often heuristic while deep knowledge is analytical . the process of building a knowledge-base involves acquisition of relevant and reliable knowledge from the experts and other sources like the literature etc . and its representation in a form that the inference engine can process in an efficient manner. . a knowledge acquisition : the most difficult challenge in building a knowledge- base is to acquire reliable knowledge relevant to the domain under consideration . this requires careful survey of different reliable sources of knowledge such as (adeli ): ( ) technical literature (e .g . books, manuals, journal articles etc .); ( ) domain experts (e .g. interviews, circulating questionnaires, example problem-solving sessions etc .) ( ) experimental data (e .g. numerical or laboratory experiments) ; ( ) input data for the particular problem . again the knowledge acquired can be domain knowledge (i .e. the knowledge related to the specific area of application e .g . finite element structural analysis) or problem knowledge (i .e . knowledge related to a particular problem in the domain e.g. the dimensions of a structure) . however, they can be either deep or surface and declarative or procedural . domain knowledge is normally independent of the particular ptoblems of application in the domain. since such knowledge does not vary from problem to problem it is often called static knowledge . for the same reason, this can be in-built in the system permanently in the form of rules, nets or frames so that the end-user need not have any control on it. in general, the first three sources mentioned above are exploited to acquire such knowledge. some of the examples relevant to the finite element analysis are rules for selection of appropriate theoretical bases for structural modelling, proper finite element mesh, type of analysis and solution etc . on the other hand, problem knowledge varies from problem to problem (hence often called dynamic knowledge) and it is generated from the input data for a particular problem in the domain in the user-interactive mode . this class of knowledge gains particular importance in the present context . some of the examples are whether the structure is 'thin' or 'thick', flat or curved, isotropic or anisotropic, whether a plane stress or a plane strain model is required etc. it can be observed that input data regarding the structural topology, material constitution and the loading environment are pre-requisites for generating such knowledge . this can be achieved through an efficient user interface . a typical flow of logic in modelling a complex structure with simple hexahedral substructures and determining the relative characteristic dimensions of each of the substructures is shown in figure . such knowledge is essential while i nterpreting domain knowledge (e .g. selecting the appropriate theoretical basis and type of elements) later . . b knowledge representation : an equally important task is to represent the k nowledge that is acquired in an efficient form that can easily be addressed by the inference mechanism. knowledge can be represented in a program in two ways - inference mechanism knowledge or data generator problem modelling oblemoptimisation geometry _ dimension determination coarse mesh (initial ) ---- material boundary _ condition _ i ooding conditions ------ _ laminate properties mesh generator (automated) theoretical support finite element analysis method of analysis error estimation computational -- type of error analysis analysis knowledge base computational - tools mesh refinement database or _ rule base coarse mesh rules optimal solution b p naganarayana and g prathap reduce your structure to parts of simpler geometry, preferably of standard shapes if nonstandard, the part should be bound by to surfaces is feed the standard dimensions locate the part in the -d cartesian space para metric generation of the rest of the surfaces flat faces create substructure volume determine characteristic dimensions the port of tandard shape y yes yes are all the surfaces i f standard geometry) ye~ ino o do at least opposite faces have standard shape ) dge mid points defined no e serendipity quadratic surfaces i ye no no a it o many surfaces bind the port ? t parametric generation all the surfaces i i surface centroid def ye figure . flow of logic for geometric data generation . procedural and declarative . procedural representation is normally used in conventional algorithmic programming and is very efficient . but the knowledge involved is highly context-dependent and embedded in the code, thus making the knowledge opaque - unintelligible and difficult to modify . on the other hand, declarative representation encodes knowledge as data so that it is context-independent and is therefore easier to comprehend and modify. while semantics are distributed over the code in procedural representation, they are collected in one place in declarative representation. several approaches of declarative knowledge representation are available in the al literature (hayes-roth et al ; o'shea & eisenstadt ; brachman & levesque ) . the major ones are briefly described below . rule-based production system - this has been the most favoured representational approach for building expert systems . a production system (brownston et al ; o'shea & eisenstadt ; van melle , pp . - ) is a collection of several rules each of which consists of an ordered pair of symbols (if and then). the first constitutes the premises or the conditions while the second constitutes the actions . this is the most popular and simplest method . the problem reduction method (barr & feigenbaum - ; cohen & feigenbaum ; ishizuka et al ; rich ; winston ; charniak & mcdermott ) is often useful for such knowledge representation so that a complex problem can be subdivided into simpler problems , described hierarchically, each having a subgoal . logic-predicate calculus - this is basically an extension of the so-called propositio n logic where a proposition can be either true or false. a first order predicate calculus (ishizuka et al ; ogawa et al ) consists of a number of components or propositions (like predicate symbols, variable symbols, function symbols and constant symbols) which are interconnected through the logical connectives (like and, or, not, equivalent, implies etc.). this type of knowledge representation is very c onvenient, in particular, with declarative languages like prolog . red of log rang ion quadratic surfaces expert systems and finite element analysis semantic networks and triplets - semantic network, normally, involves nodes and arcs (links). nodes represent objects, concepts or situations, their attributes and values . the arcs link these nodes and define the nodal heirarchy . if the net consists of only one object, and its attribute and value, the three nodes form the so-called semantic triplet or oav triplet (adeli ). frames - sometimes frame-like structures are found to be convenient for knowledge representation, in particular for representing sequences of events in expert systems (minsky ; aikins ; yao & fu ). a frame is composed of its name and several slots. each slot may have a number of facets. a slot in a frame may be an attribute of the object or it may be a relation. a relation slot is used to link different frames together . in the present context, for such a dominatingly analytical exercise, a rule-based production system appears to be most promising . . c representing uncertain knowledge -interval modelling: often the available knowledge about the domain and the problems of interest in the domain are uncertain and incomplete . usually, such knowledge is qualitative e .g. strength of a singularity is high or plate thickness ratio is low etc . in fact, the intelligence of an expert (and hence of an expert system) is reflected in his ability to handle such knowledge . in such cases the so-called approximate reasoning or inexact reasoning or reasoning under uncertainty(zadeh , pp. - ; blockley ; ruspini ; ishizuka et al ; negoita ; lecot & parker ) should be employed for representing and processing the available uncertain knowledge . interval modelling (wong ) is one of the most popular mathematical tools used for handling uncertainty . different mathematical measures which use interval modelling concept are illustrated in wong ( ) . some of them, which are useful for the present exercise, are briefly discussed here . (a) simple intervals -this needs minimum information as seen in figure a . the value of a parameter is estimated to be within aft interval [a, b] and nothing more is known . n (x) (d) (e ) µ(x) x o a o figure . different types of fuzzy knowledge representation : (a) simple interval, (b) triangular set, (c) rectangular set, (d) trapezcidal set, (e) normal set . x x b p naganarayana and g prathap (b) probability measures - probability measure implies exact knowledge of the uncertainty . although the value of the parameter is uncertain, there is certainty about the character of its randomness (e .g . normal curve) . the probability measure is built on a simple interval and the probability of an event a, say p(a) is known at any point in the interval . it should be noted that p(a) implies that probability of the event (not a) is also certain and p(not a) = i - p(a). (c) possibility measures and fuzzy sets - possibility measures are again built upon a combination of simple intervals as seen in figures b to e . the possibility of an event u(a) is known at any point in these interval and the resulting set is generally called fuzzy set . here, the stringent presumption of probability theory is relaxed by introducing the concept of ignorance (shafer ) that the probability of event a should not say anything about the probability of (not a) unless there is explicit evidence supporting the latter i .e . µ(not a) is independent of µ(a). this measure is often identified with the so-called evidence measure . depending on the variation of the possibility factor µ(a) associated with the event/object a the fuzzy set can be of any shape . for example, a triangular set (figure b) can be used for a pessimistic representation of the event a that at only one point p(a) = . on the other hand, a rectangular set (figure c) can be used for optimistic representation such that µ(a)= at all points except at the two end points . again the possibility function can be a continuous function (figure e) . but a trapezoidal set (figure d) is seen to be the most practical choice for representing fuzzy knowledge . there are several instances in finite element structural analysis where the knowledge representation becomes fuzzy . the most common one is of determining the dimension of analysis for the structure, in particular, if the structural geometry is irregular and distorted. figure shows how the thickness ratio for a general hexahedral structural domain can be represented using the fuzzy set theory (also see § ). some other instances where fuzzy knowledge representation is required in finite element structural analysis are - the intensity of mesh refinement required in an adaptive finite element approach, the strength of a structural singularity, strength of the inter-laminar bond in a composite structure etc., where the knowledge is often heuristic and verbal . . inference mechanism and expert system architecture an inference mechanism is the part of an expert system that deduces new facts from known facts of the knowledge-base. this mainly constitutes the problem-solving strategy for the required task . problem solving can be considered as a search for solution through a state space by the application of operators, where the space (the possible states in a problem solution) consists of an initial state, a goal state and an intermediate state . the solution path consists of all states that lead from an initial state to a goal state. a number of different problem-solving paradigms are available in the al literature - the describe-and-match paradigm, the goal-reduction paradigm generate-and-test systems, means-ends analysis, rule-based (production) systems etc . (barr & feigenbaum - : cohen & feigenbaum ; ishizuka et al ; o'shea & eisensfadt ; winston ; brownston et al ; charniak & mcdermot t ). many of these approaches overlap or can be used, in conjunction with each other. however, it is the production system which is popularly used for developing most of the knowledge-based expert systems so far . this has additional advantages oper other strategies, being simple to code, with improved clarity, and being suitable for expert systems and finite element analysis h -- h/i ratio as triangular set h/i ratio as rectangular set min . / = mox .(t/i)= . mean / = . min (b/ )= mox(b/ )= mean (b/i)= min (t/b) = . max (t/b) = mean it/b) = a - classical b - order c - iii order (a) d - iii order ( ) e - deiasticity (approximate ranges) template figure . fuzzy sets for the dimension ratios : (a) generation of dimension ratios ; (b) comparison of fuzzy sets. parallel computations and for designing an explanation facility for the system . in particular a rule-based system combined with the goal-reduction paradigm can form a very efficient and powerful model of the human information processing and problem solving abilities . a production system normally consists of three main elements - aset of if-then rules (or knowledge base), a global database or working memory or the context and an inference mechanism . i point x y a b c d e f g h -i ii - - mean dlmn s( ,m) s(j,m) . b s( k,m) t b p naganarayana and g prathap . a inference mechanism - problem solving strategies : the next question is, how are the rules determined from the knowledge base and which ones should be used at a certain state of the solution path . for this one has to choose an inference mechanism or a control strategy which constitutes the heart of a knowledge-based expert system. inference mechanism fires the rules in accordance with its built-in reasoning process . the two major types of reasoning are antecedent reasoning and consequent reasoning. (a) antecedent reasoning (forward-chaining, data-driven control strategy) - in this type of inference mechanism, the rules are scanned until one is found whose antecedents (left-hand sides) match the information for the problem entered in the working memory. then the rule is fired and the working memory is appropriately updated . this process is repeated until the goal state is achieved or no useable rule is found . this strategy is particularly useful if the goal state is not known or if there are too many possible outcomes. complex planning problems can be tackled through this approach . the reasoning involved in the finite element structural analysis predomin- antly falls into this category e .g. determination of the dimension of analysis and the theoretical support required, given the structural topology, material constitution and loading environment, error prediction from finite element analysis and heuristic expertise etc . (b) consequent reasoning (backward-chaining, goal-driven control strategy) -in this type of inference mechanism, rules are scanned and those whose consequent actions can lead to the desired goal are determined . the associated antecedents are tested to see whether they match the working memory information . if they all match, the rules are fired and the problem is solved . if the goal states are known and their number is small then this reasoning process seems to be advantageous . this is popular in the diagnostic type of expert system . sometimes it may be convenient to use backward chaining in finite element structural analysis e .g . selection of appropriate elements and design of the required finite element mesh based on the knowledge of the dimension of analysis and the theoretical supports that are already established, adaptive refinement of the finite element mesh based on error predictions etc . (e) opportunistic reasoning (hybrid or mixed chaining) - this is a combination of both forward and backward chaining . it follows that this becomes mandatory, in particular, for an expert system for finite element structural analysis . this approach can be efficiently utilised through the use of the so-called 'blackboard environment' (nil ). the blackboard will keep track of the simultaneous application of backward and forward reasoning chains which are activated at the most 'opportune' time . . b blackboard architecture: the concept of blackboard architecture was first implemented in hearsay-il, a speech understanding system (reddy et al , pp. - ) . the control of the program can be made flexible and efficient and the restricted accessibility of routines could be eliminated by isolating the routines (or modules), but allowing all the subroutines to make use of the common data structura l the blackboard model of problem solving is highly structured and essentially a special case of opportunistic reasoning . the knowledge necessary for solving the problem is divided into independent groups of rules called knowledge sources . a blackboar d plays the part of a communication vehicle among knowledge sources and keeps track of the incremental changes made in the problem state until a solution is found . in the present context of expert systems for finite element structural analysis such the an architecture appears to be the most suitable since all the information and are associated rules necessary to model the structural domain geometry, boun expert systems and finite element analysis conditions, material constitution and loading conditions (acquired/generated as mentioned in the previous section) can be stored in the respective independent knowledge sources. figure shows a schematic diagram of the blackboard model for the system which acts as a global database in controlling the above-mentioned logical knowledge sources as well as the computational knowledge sources. . c user interface: the primary motivation for expert-system development is to make the expert knowledge, in a specific area of interest, available to the amateur user of the system. thus inclusion of an exhaustive user interface in the system becomes mandatory. this module generally includes advice/help, explanation and learning facilities. this monitors functions like modelling the problem (e .g. structural topology and material constitution and the loading environment), selection of appropriate tools (e.g . finite elements, mesh size, type of analysis and solution routines etc.) through a live conversation with the user . at any state in the problem space the system should be able to advise the user with proper verbal/graphic support, if asked for, about taking decisions appropriately depending upon the problem environment. also capability of self-justification is desirable so that the user should finally know why and how the system arrived at the suggested goal/advice. in fact the main purpose of an expert system, that transfer of the experts' knowledge to the novice users of the system be simple and flexible, cannot be fulfilled without such an explanation facility . for example, if the system decides that -d analysis is enough for a structure, it should be able to educate the user about why such a decision was taken and what are the preceding factors that lead to such a decision, thus enabling him to examine the reasoning process . finally, the system should update its knowledge, as its human counterpart would normally do, from time to time . for example, if the user creates a new geometrical shape which is unknown to the system and if he feels that it invites frequent usage, the system should be able to 'learn' the necessary details and store them in its library with the help of the user . a more powerful example is of self learning . for example, if the system reaches the same goal (e .g. nonlinear analysis for a structure) during several runs, appropriate chunks of knowledge (e .g. geometry, material, boundary conditions and loading) involved in all such runs may be compared to generate, probably, a new heuristic rule which can be used in future under similar conditions without doing the actual computations required . building such automated learning systems is quite a difficult challenge and the expected efficiency and effectiveness are yet to be achieved . . a case study - an expert system for finite element analysis of structural problems using -dimensional elements a project on developing an expert system for control of modelling of the structural problems using finite elements (only -d elements at present) is under progress at the national aeronautical laboratory in bangalore (prathap & naganarayana ) . mesh refinement is automated using an h-type strategy based on strain energy density error norms. in this section, we take this up as a case study to demonstrate some of the issues involved in such developmental work . the inexact representation of certain problem knowledge using fuzzy sets is first discussed . later some rules related to the present system are briefly listed below (the knowledge-base, however, is built for using 'b, -d as well as -d elements) . b p naganarayana and g prathap . inexact representation of knowledge the major parameters influencing the selection of the theoretical bases to be used, as we already discussed, are the domain dimensions, material constitution and presence of singularities. usually it is not possible to represent the knowledge about these conditions in an exact sense . the fuzzy set representation of these can streamline the problem-solving process for elegance and efficiency . . a fuzzy sets for ratios of characteristic structural dimensions -. and represent the fuzzy sets for the accuracy of laminar stress recovery demanded by the user and the interlaminar bond weakness, respectively. using sl as a weighting factor, the resultant fuzzy set for lamination effects (c) can be constructed . . some typical rules some of the typical rules that can be used for modelling a structure using finite-elements and for remeshing the mesh adaptively, are listed here. . a rules for -d analysis: (i) global -d analysis if, {number of the binding surfaces _< ) or { or more adjacent surfaces are fully or partially supported/loaded) or {both (d,,) and 's are high or very high and are very high) . ( ) local -d analysis around ith singularity if {when and/or is very high) . . b rules for -d analysis ( ) -d analysis if, {any of 's, are moderate or low or very low but not (d)'s as well as are very low) . ( ) elementary plate/shell theory if, {all are very low). geometric basis: now construct the mid-surface in the -b curvilinear 'plane' by considering the midpoints of the edges in the h-direction and find out the curvatures kl and k in the -h and b-h planes respectively . this is done if standard domain shapes are not chosen . find the gaussian curvature k = kl x k . b p naganarayana and g prathap ( ) plates if, {kl =k = } . ( ) plane stress model if { ( ) and ( ) or ( ) are satisfied) . ( ) plane strain model if, {both 's are high or very high and the h-, b- dimensions, loads and boundary conditions do not vary with the i-dimension) . ( ) zero gaussian curvature shells if, {either of kl and k are non-zero but k = } . ( ) positive gaussian curvature shells if, {k > } . ( ) negative gaussian curvature shells if, {k < } . ( ) shallow shells if, kl and k are very small. . c rules for -d analysis : ( ) -d analysis if, { (dhb ) is very high but is not high or very high) . ( ) higher order beam theory if, {