doi:10.1016/j.eswa.2006.09.014 www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 551–563 Expert Systems with Applications An expert system for dynamic re-coordination of distributed workflows q William L. Kuechler Jr. a,*, Vijay K. Vaishnavi b a Department of Accounting and Information Systems, University of Nevada at Reno, United States b Department of Computer Information Systems, Georgia State University, United States Abstract A persistent problem in the use of automated workflow management systems for inter-organizational workflows has been the need for manual redefinition of coordination points in the process models when either organization changes its processes. Coordinating commu- nications between production-chain organizations are usually based on the notification of completion of tasks by one party which con- stitute pre-conditions for activity in another organization; autonomous task changes disrupt coordination. This paper describes a workflow re-coordination model and corresponding expert support system based on workflow goals which are more stable than the low-level machine, role and technique dependent activities which implement them. Following development of the model, we describe a set of three elementary disruption cases which span a large number of common workflow changes. Using these cases we demonstrate that common coordination disruption situations can be totally or partially repaired by use of an expert coordination subsystem. � 2006 Elsevier Ltd. All rights reserved. Keywords: Workflow management system; Coordination; Goal-based process modeling 1. Introduction Workflow management is the study and analysis of busi- ness processes with the explicit intent of interfacing, merg- ing, expediting and/or redesigning those processes, when they are considered in their full context. Thus, workflow management differs from traditional industrial engineering process studies primarily in scope. Workflow management intends to cross divisional and traditional single-output- process boundaries in the search for global (organizational) process effectiveness. Workflow management includes anal- ysis, description, design, and augmentation of workflows (Georgakopoulos, Sheth, & Hornick, 1994). As the concep- tion of the enterprise becomes global and virtual (Morita, 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.09.014 q This work is partially supported by NSF Research Grants, IIS- 9810901 and IIS-9811248, and a research grant to the second author from the Robinson College of Business, Georgia State University. * Corresponding author. Tel.: +1 775 784 6910. E-mail addresses: kuechler@unr.edu (W.L. Kuechler Jr.), vvaishna@ gsu.edu (V.K. Vaishnavi). Mukaigaito, & Hayami, 1996; O’Leary, Kuokka, & Plant, 1997) the challenges of semantic ambiguity and dynamic task change and substitution are introduced into inter- organizational workflow management by the necessary autonomy of cooperating organizations (Vaishnavi & Kuechler, 2005). Most existing WFMS are based on ‘manufacturing models’ of work where precise specification of activities and their execution times are (presumed) critical. Yet many work processes, especially knowledge (office) work, neces- sarily incorporate slack time to handle indeterminacy and frequently specify desired results intentionally, that is the process goal(s) are significant, but the details of execution are not (Davenport, Jarvenpaa, & Beers, 1996; Hewitt, 1991). As organizations become more virtual and an increasing number of non-strategic functions are subcon- tracted, traditional production workflows take on the same character of being more readily specified by intention than by concrete task. Fig. 1 shows a macro view of the business environment. Processes in any organization (X, Y or Z) are dependent on mailto:kuechler@unr.edu mailto:vvaishna@ gsu.edu mailto:vvaishna@ gsu.edu Organization X Organization Y Organization Z processes process communications Client Process X Process Request to Y Process Reply from Y Server Process Y Process initiated by X Reply to X (1) Process Reply from Y Reply to X (2) Business Process Communication Client/ Server Process Interaction triggers Fig. 1. Cooperation through Process interaction (adapted from Morita et al., 1996). 552 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 information from concurrent processes in their trading partners. The client–server portion of the diagram illus- trates this dependency. Though the process interactions diagrammed in Fig. 1 may be manual or automated, our research is concerned with situations where the processes are controlled by automated workflow management systems (WFMS), and the process communications are between interoperating WFMS. Each WFMS operates from a model of a work process. A trigger (Joosten, 1994) in a WFMS signals the satisfaction of a set of precon- ditions that cause the enactment engine to perform (or schedule or assign resources to) a pending activity in a sequential process. Recent work by the process modeling and workflow management communities defines the coordination of tasks across semantically heterogeneous environments as a sig- nificant problem area (Andersson, Bider, Johannesson, & Perjons, 2005; Buhler & Vidal, 2005; Khomyakov & Bider, Verify education Check FBI for criminal record Check for in-state criminal record preliminary_ clearance_ complete Duration Activity (original) 1 day (phone) 3 days (fax) 2 weeks Durati 1 day (pho 3 weeks 2 monthsTrigger Fig. 2. WFMS trigger de 2001). We directly address some of the semantic issues in task interpretation that have been identified by these authors and others, for example (Agostini & De Michelis, 2000; Klein, Dellarocas, & Bernstein, 2000; van der Aalst & Jablonski, 2000), as not adequately addressed by prior WFMS and process research. The problem of coordination disruption through auton- omous activity changes in work processes is illustrated in Fig. 2. The figure shows a highly simplified definition of a security clearance process (Vaishnavi, Joosten, & Kuechler, 1996) in which a trigger, preliminary_clearance_complete in a process is described in terms of the completion of activity: check for in state criminal record. The trigger results in train- ing for a new hire in parallel with the conclusion of the secu- rity clearance and having the training begin prior to a complete clearance saves significant time. When the process is redefined, as would likely be the case if the process were shifted to a subcontracting site, the trigger no longer exists Verify education Check interpol for international criminal record Military intelligence investigation for national criminal activity on Activity (redefined) ne) Lost coordination ! The activity defining the trigger has been eliminated. ? Trigger fined as process-state. W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 553 and training is delayed until coordination can be reestab- lished between the contractor and subcontractor. Traditional work process modeling methods such as SADT and trigger modeling (Joosten, 1994) assume a deterministic, fixed element model (Lei & Singh, 1997; Yu, 1995). The system is closed, that is, the activities that make up the various processes are predefined. As indicated above, this assumption is almost certain to be violated when autonomous WFMS seek to interoperate. Autono- mous groups, especially in outsourcing situations, must be free to change process details. Thus we seek some more robust method of specifying trigger conditions than match- ing to fixed patterns of activities. Our prior research in this area (Vaishnavi & Kuechler, 2005; Kuechler, Vaishnavi, & Kuechler, 2001; Vaishnavi et al., 1996) has addressed the issues at a conceptual level and has developed formal models of the contractor–sub- contractor relationship based on the workflow manage- ment consortium (WFMC) workflow enactment model. The research reported in this paper builds on that research to articulate a computable goal-based coordination model to implement the formal models and develops the comput- able model into a working prototype. In the next section of the paper (Section 2) we survey the work of other researchers to introduce flexibility into pro- cess definitions. Section 3 develops the goal-based work- flow coordination model and Section 4 presents the architecture of the corresponding expert system. Section 5 discusses the expert system implementation, introduces three exemplars of coordination disruption as test cases and describes the system response to these cases. In a con- cluding section we discuss some of the ways in which the system and model can be augmented, and our on-going research with the technique. 2. Search for flexibility in workflow and process modeling In our research on coordinating inter-organizational WFMS, we have surveyed literature from multiple fields for approaches to similar problems and alternative app- roaches to coordinating complex systems. Coordination problems are ubiquitous, and in all fields, we find the recognition of the inadequacy of fixed representations for modeling real-world work situations. Cao and Sanderson (1995) find fixed representations of process inadequate even for the constrained field of robotic workcells. Traditional deterministic planning representa- tions are simply not robust enough under real-world condi- tions. Their approach is to expand traditional Petri net models of process by allowing the transitions between states to be expressed as fuzzy variables, interpreted by production rules. While suggesting interesting techniques, their research addresses a sufficiently different problem that their solution is not immediately applicable to workflows. Specifically, their domain requires fixed sequencing of activities, and many workflows, as noted in the literature, are dynamically resequenced in response to changes in the environment. The HFBP model (Morita et al., 1996) of business work- flows is specifically intended to describe work processes between interoperating workflows in autonomous organi- zations. Drawing from research in fault tolerant software systems, they model processes in terms of an attribute grammar. While successful in addressing certain types of error recovery in interoperating WFMS (due to inherent capabilities of attribute grammars) they note that the more general problem of dynamic change in process definitions for interoperating workflows is not sufficiently handled by the attribute grammar approach alone. Much research in office information systems (OIS) describes the complexity of actual work environments, and is acutely aware of the inadequacy of fixed process models for support of real-world work situations (Boland, Maheshwari, Te’eni, Schwartz, & Tenkasi, 1992; Gerson & Star, 1986; Hewitt, 1986, 1991). The following five models come from that area: (i) The partially shared view scheme (Lee & Malone, 1990) is a technique for enhancing the utility of com- puter supported collaborative work systems used between autonomous groups with differing world views and correspondingly different functional requi- rements. The scheme has been implemented in the Object Lens collaborative work support system, and attaches a type hierarchy to each work object the sys- tem supports. The type hierarchy is assumed at least partially shared (especially at its higher levels) by all users of the system. By determining correspondences with a known hierarchy, an object otherwise unknown to a user group can be at least partially recognized and usefully processed. Our workflow model adopts this technique by analogy: the know- ledge in our hierarchy is different, but our partial matching to a hierarchical structure to make infer- ences about an unknown entity is very similar. (ii) The AMS formalism (Ang & Hong, 1994) is both a model and a collaborative work support system implementation. The formalism can be character- ized as ‘‘loosely grammatical’’ in that the activity sequences that make up a workflow are sequentially expanded from a root ‘‘abstract activity’’ by a series of productions. Stressing flexibility, however, it does not attempt to formalize the grammar (and obtain the benefits of following a well researched formalism) in the manner of HFBP. The allowed productions by which non-atomic activities can be expanded are (1) an elementary activity, (2) an activity network, and (3) a memory organization packet for activities (MOPA). The MOPA draws from the work of Schank and Abelson (1977, 1995) and others in natu- ral language understanding and is responsible for much of the flexibility of the system. A greatly simpli- fied description of a MOPA is as an abstract script 554 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 that represents an activity sequence in general terms, and is instantiated to actual activities by the run-time context. Though more accommodating to actual work situations than the Petri net models it aug- ments, AMS is essentially a single system model, and has no mechanisms for preserving coordination between interoperating systems which make use of its flexibility. (iii) The Promanand system (Karbe, Ramsperger, & Weiss, 1990) is an office support system using a ‘‘cir- culating folder’’ metaphor. It is unique in that it is the most determined attempt to date to support a deter- ministic view of office work through the exhaustive enumeration of exceptions conditions for a given work environment. The lack of success of the system is viewed by Ellis and Wainer (1994) and others as support for the position that any non-adaptive sys- tem will ultimately prove inadequate to real-world work environments. Further support comes from Hewitt (1991) who has characterized office systems as open systems which cannot be analyzed a priori. (iv) Mahling and King (1999) point out that nearly all of the problems of inflexible computer assisted work- flows were first encountered by researchers in office information systems (OIS) in the 1980s. They recog- nize that workflow models based on goals are more flexible and robust than task descriptions, and their PolyFlow system dynamically expands goal-based process definitions into activity plans for the extremely demanding office work environment. Their approach coordinates changed activity sequences for different agents in a distributed workflow through automated support of human–human negotiation. While acceptable in the single non-repeating project office environment they envision, a wholly automated re-coordination approach seems desirable for repeti- tive, production workflows. Table 1 Flexible workflow models Model Authors Term Primary flexibility en Fuzzy Petri nets Cao and Sanderson (1995) Short Fuzzy transitions to HSBP Morita et al. (1996) Short Attribute grammar d PSV Lee and Malone (1990) Short Documents carry typ AMS Ang and Hong (1994) Both Generalized scripts, w Prominand Karbe et al. (1990) Long Exhaustive enumerat PolyFlow Mahling and King (1999) Long Goal-described proce support for exception AgentWork Muller et al. (2004) Long Rule-based reschedul temporal) are detecte Multiagent workflow enactment Buhler and Vidal (2005) Long Workflows are initiat definition language. T sequences Case-handling van der Aalst et al. (2005) Both Activities are enabled (v) Similar in many respects to PolyFlow is the case-han- dling approach of van der Aalst, Weske, and Grun- bauer (2005). Directed also at office workflows, the distinguishing feature of case-handling, is that it is data driven rather than process driven; activities are enabled based on data availability rather than activ- ity completion preconditions. While allowing for very flexible sequencing of activities, case-handling systems require continuous manual attention and extreme visibility of activity and context data as is unlikely to occur in a contractor–subcontractor pro- duction workflow. Buhler and Vidal (2005) propose that adaptive workflow enactment will, based on historic and functional consider- ations, ultimately be enacted through agent-based systems. In this vision a form of high-level business process descrip- tion language will be used to compose ensembles of agents into an ‘‘initial social order’’ for enactment of the work- flow. As these software agents negotiate with each other and continuously search for web-based alternative services, workflows will be dynamically responsive to environmental changes throughout their enactment. AgentWork (Muller, Greiner, & Rahm, 2004) is a WFMS that seeks to adapt to workflow exception condi- tions by specifying high-level logical rules for approaches for rescheduling activities. The system evolved in a medical environment where timely treatment is vital and so this sys- tem focuses on the detection of temporal exceptions and even projects time overruns before they occur. When excep- tions are detected, a rule-driven scheduling system resched- ules activities and may introduce substitute activities with different execution times or treatment applicability. We note that while the workflow enactment agents of Buhler and Vidal (2005) as well as the rules of the Agent- Work system (Muller et al., 2004) implicitly use goals as the higher level ordering principle for adaptive workflows, hancement augment Petri net models escription of process e/behavior hierarchy allowing partial interpretation hich are situationally instantiated ion (attempted) of all exception conditions for given environment sses dynamically expanded to executable task sequences; negotiation s ing or substitution of activities when workflow exceptions (especially d ed by composing ensembles of software agents with a business process he agents then dynamically direct the enactment of workflow activity based on data availability rather than activity completion preconditions W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 555 neither system explicitly mentions goals. However, actions of the software agents are guided by goals in their search for alternative activity sequences and the rules that guide AgentWork have of necessity been a priori derived from the higher level goals of the processes they enact. Table 1 summarizes the adaptive/flexible work process models discussed in this section. Soffer (2005) distinguishes between two forms of workflow flexibility: long term adapt- ability sometimes referred to as agility in processes and short term flexibility which is the ability to recover from small, short term changes in a way of working. We have categorized the models in Table 1 according to this criteria for comparison to our model, which is specifically aimed at dynamic recovery from small workflow changes (short term). While all the research overviewed in this section is moti- vated by the understanding that actual work processes are highly situational and require adaptive models, none except HSBP considers the distinct, but closely related problems inherent in modeling multiple, tightly-coupled autonomous production WFMS. HSBP outlines a grammatical model for single systems, but specifically reserves the modeling of interacting systems for later research. In the next section we demonstrate a reconceptualization of workflow coordi- nation that allows both process flexibility and tight cou- pling between trading partner WFMS. 3. Goal-based coordination model framework Enhancing conventional activity-based process models with goal (intentional) information has been suggested as a basis for introducing greater flexibility into automated process enactment since at least the early 1980s (Mahling & King, 1999). In analyses of workflow meta-models (Lei & Singh, 1997), it is suggested that no other technique can provide the directive information required to dynami- cally re-sequence tasks in response to error conditions. However few models or WFMS have been developed that actually operationalize and use goal information. Our workflow coordination model captures goal infor- mation about processes in a formal structure and uses the information straightforwardly to reason about changes to process activity sequences. We have named our model after subpr satifie verify education criminal record check (state) tr satifies_using satifies_using Fig. 3. WFMS trigger reconceptualized as a semantic its most distinguishing feature: a hierarchical overlay of process intentions (HOPI). Unlike agent-based approaches which require dedicated workflow systems (cf. Mahling & King, 1999), HOPI works as an overlay to augment most existing workflow management systems. Several HOPI attributes are unique in workflow and process modeling: first, the goal hierarchy is used not for planning, replanning or evaluating a process but rather for recognition of similarity between processes; second, coordinating communications between cooperating work- flow actors are elevated to the status of first order activities rather than pre or post-condition methods within an activity object. As such they are not bound to any spe- cific activity, but like all activities within HOPI, are linked instead to the goal hierarchy. Indeed, unique to HOPI, coordinating communications are assigned goals indepen- dently of activities so that they can be rescheduled indepen- dently of activities. Fig. 3 shows a high-level object model of the same process modeled in Fig. 2. Note that the trigger, preliminary_clearance_complete is an object at the same semantic level as the other activities in the process. Subject to semantic constraints derived from process goals in addi- tion to timing and other hard constraints, it can be resched- uled and replanned to preserve coordination even under process change or activity substitution. The primary data structure of the HOPI model is a work definition (WD) that specifies a hierarchy of intentions (goals) for the activities of a work process. The WD is an overlay to the activity sequence of the workflow and thus is generally applicable to any process representation and to the many WFMS currently in use. Fig. 4 illustrates the definition of the WD used by the model. The root goal is the overall purpose of the work, and the name by which the work is commonly known in the organization. The layer of subgoals is generated during process design or specification, and represents a process of stepwise decomposition of the root goal into its compo- nents (Simon, 1977). When the subgoals are well enough defined to be implemented, they are linked to the general- ized functionality by which they will be enacted. Incorpo- rating the concept of functions into the WD more closely models the manner in which workers conceive their ocess: security_clearance s_using criminal record check (federal) igger: preliminary _clearance_complete satifies_using ally defined, independent communications event. HOPI (Any conventional process representation) Problem Space (Intentional) Task Space Root goal (R) Sub-goals (S) Generalized functions (F) Activities (A) Trigger (T) Solution Space Legend: Goal-Sub-goal edge (GE) Goal-Function edge (FE) Function-Activity Edge (AE) Predecessor-Successor links Semantics: Sub-goal Contributes-to goal Function Operationalizes sub_goal Activity Enacts function Activity precedes/follows Activity Fig. 4. A hierarchical structure of intentions as an overlay of process representations. 556 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 activities than use of goals alone (Schank & Abelson, 1995). The terminal (leaf) nodes of the WD are the actual work activities containing predecessor/successor infor- mation. Fig. 5 shows the work definition instantiated for one of our three exemplar test cases: a simple activity substitution in the manufacture of a woman’s suit, consisting of a vest and a skirt. Production of the vest is decomposed into sew- ing the vests and storing the vests until final packaging. The function that implements the storage subgoal is load&ship- ViaTruck, which is concretely realized by the activity vest- sToWH3. The interprocess trigger, a communication to a trading partner to order cloth for other manufacturing pro- cesses (not shown) is linked to the completion of vest- sToWH3. The linkage of a coordinating communication to a specific predefined activity state is common in WFMS and renders systems brittle under change. Following work redefinition (K5a ! K5b in Fig. 5), the concrete activity Fig. 5. Original and altered wo has been changed to VestsTo9thStWH, resulting in coordi- nation disruption and the failure to order the cloth in a timely manner. The dynamic workflow re-coordination model depends on similarity inferences between an original process defini- tion and an altered definition, both modeled with HOPI. Fig. 6 illustrates the similarity concepts used to interpret the intentionally specified work definitions. The WD is a tree structure in which each element is a node. The path from the root goal to any activity node is termed the inten- tional context (IC) of the node, since tracing the path yields the semantics of the node. When ICs of activity nodes in different WDs are compared beginning at the root, the low- est identical node of the ICs is called the lowest common point (LCP) of the ICs. For example, if nodes 1 through 4 of WD’s A and B in Fig. 6 are identical, but node 5 differs in A and B then node 4 is the LCP of the ICs. The depth of the LCP is a measure of IC similarity. Activity nodes also rk definitions for test case. Fig. 6. Similarity inferences based on the work definition. W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 557 have a functional context (FC), which is loosely defined as the set of activities adjacent to a given activity node. Func- tional context similarity between activity nodes i and x in two WDs is measured by an index ranging from 0 (no sim- ilarity) to 1 (identical). The index is computed by adding a weight to the index for each activity node position relative to i and x for which the nodes are identically defined. Weights diminish geometrically with distance from i and x, as shown in Fig. 6. When information on the expected duration of the activities is available, a further measure of functional context similarity is functional context dura- tion, the ratio of the times from the start of the process to the trigger activity and from the trigger activity to the end of the process in the original and the altered process expressed as a percentage. Using the example of Fig. 5 to illustrate the use of the similarity measures, note that both activity hvests to WH3i in process K5a and activity hvests to 9th StWHi in process K5b have identical intentional contexts; the path down the intentional tree from the root goal hmakeSuiti is the same in both cases. Further, the activities that precede and follow activity hvests to WH3i in process K5a are iden- tical to those that precede and follow activity hvests to 9th StWHi in process K5b and those activities in both pro- cesses implement the same function hload&shipViaTrucki. Thus the functional contexts are also identical. Together these measures strongly indicate that the change to the workflow activity is nominal and the trigger may be attached to the new activity with no loss of information. This is precisely the determination of the expert system as shown later in Fig. 13. The similarity heuristics are based on Barsalou’s (1983) work on ad hoc categories and the work of Schank and Abelson (1977, 1995) on the relation of scripts and goals to understanding. The key concept is that when activities and their goals are structured in a conceptual hierarchy as in our WDs, each higher level node serves as a cognitively organizing principle for the nodes beneath it, enabling inten- sional classification and recognition of lower level nodes. In the state-oriented view of processes (Khomyakov & Bider, 2001) a business process is a trajectory in a multidi- mensional state space where each change of state (motion in the state space) takes the process closer to one or more goal states. Under this view, two processes are considered similar if (1) there is an isomorphic mapping from one state space to another (2) the process spaces have similar goal states (3) the processes have the same valid movements in the state space toward the goal (Andersson et al., 2005). Although one is derived from a cognitive model of process similarity perception and the other from a state-space for- malism, HOPI and the Khomyakov–Bider process views are mutually reinforcing; our re-coordination model oper- ationalizes all three Khomyakov–Bider similarity metrics. Coordinating through goals rather than specific activi- ties is closely analogous to the software engineering proce- dures by which subroutines are made loosely coupled (less dependent) in order that the system as a whole should be more robust (less fragile). Prior to discussion of the operation of these principles within the expert system, a brief discussion of the usage sce- nario envisioned for the model will provide a high-level context (see Fig. 7). Two business entities, a contractor who requests work and a subcontractor who provides the work are involved in each use of the model. Both entities manage their processes with compatible WFMS and at some initial point share a common process definition for the subcontracted work. At a later time the subcontractor unilaterally changes the definition of the work being requested by the contractor; however the contractor requires notice of some intermediate point in the process (a trigger) to enable scheduling efficiencies. The coordina- tion model attempts to determine the appropriate point Fig. 7. A usage model for the coordination subsystem. 558 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 in the altered definition to schedule the trigger, eliminating the user communication and reprogramming that would otherwise be required. This usage model corresponds to a reference process- based multi-enterprise process (Schuster, Georgakopoulos, Cichocki, & Baker, 2000) in which cooperating partners construct relatively long-lived virtual enterprises. Also referred to as outsourced workflows (Schulz & Orlowska, 2004), such inter-organizational processes are normally tightly coupled through work state dependencies which limit flexibility. However, the HOPI subsystem enables autonomous process redefinition, an adaptiveness normally found only in service-based workflows while retaining close, state-based coupling. 4. An architectural view of the coordination subsystem The subsystem is divided into two major components, a process architecture and a knowledge architecture. This divi- sion is typical of knowledge engineering, where a Process and an Expertise analysis model are defined for every sys- tem (Chandrasekaran, 1986; Giarrantano & Riley, 2005; Tansley & Hayball, 1993). The knowledge architecture, the intelligence used in maintaining coordination, is an articulation of the HOPI intentional description of the workflow rationale as described in Section 3. The process architecture, which illustrates the relationship of the major functional modules required for dynamically maintaining coordination between loosely coupled (autonomous) workflows, is shown in Fig. 8. The basic structure is similar to well known architectures for analogous tasks: plan- ning, understanding from plans, and constraint-based scheduling. 4.1. Process overview Moving from left to right in Fig. 8: the original work- flow definition (activities, their ordering, etc. as described in Section 3) and a modification of that original definition are available to the model. The comparison module deter- mines differences and similarities between the definitions using the concepts of functional context and intentional con- text as set out in the previous section. Note that compari- son is restricted to those portions of a Work Definition that contain triggers. The model does not concern itself with changes in any portions of a process that do not constitute conceptual or functional contexts for triggers. After comparison, difference data is passed to the inter- pretation module. The interpretation module attempts to ‘make sense’ of the differences, and reconcile them to known elements in the original process definition. If this activity is successful, as judged by evaluation heuristics, the ‘recognized’ activity set is passed to the scheduler. That module attempts to reschedule coordinating triggers in a manner that satisfies both ‘hard’ constraints, such as fixed times for events, and fixed predecessor/successor relations and the semantics attached to the coordinating triggers. The scheduler applies domain specific and domain inde- pendent heuristics to the scheduling of coordinating com- munications. This is required, for example, when changes Difference data orand 2 Recognition/ interpretation 1 comparison Modified workflow definition Original workflow definition Reconciled workflow definition ERROR No acceptable definition found Eval Data flow Control flow Principal function Knowledge Process 3 scheduling Activity similarity rankings, original and modified WD’s Fig. 8. Knowledge and process architecture for the dynamic WFMS coordination subsystem. W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 559 to the workflow definition have resulted in the elaboration of a single task in the original definition into multiple activ- ities in the changed definition. Heuristics may be used to determine a point within the elaborated activity at which to communicate state completion to the modified workflow. A process and a modification of it can always be recon- ciled at some level of abstraction. In general, very rough confidence factors can be assigned to a reconciliation (trig- ger rescheduling) based on measures of intentional and functional context. These factors along with the scheduling actions taken by the rescheduling algorithm are posted to a log file by the system for operator review. 5. Coordination system implementation The coordination subsystem prototype has been imple- mented in CLIPS, the widely used NASA developed production system language (Riley, 2005). The HOPI workflow description hierarchy (described in Section 3) Table 2 CLIPS data structures and instantiations for a HOPI work definition HOPI node structure definition P (deftemplate goalNode ( (slot name) (slot parent) (slot workProcess) ) (deftemplate activity ( (slot name) (slot parent) (slot scheduledDuration) (slot predecessor) (slot successor) (slot workProcess) ) has been modeled as a tree structure of deftemplates nodes linked by predecessor/successor pointers. The architecture modules, comparison, recognition/interpretation and sched- uling (described in Section 4) have been modeled by functions within different CLIPS modules (namespaces). We have chosen to use the CLIPS 6.23 Windows inter- face, even though it is primarily text-based since at this stage of the project only the developers are using the system. Each case – a set of CLIPS deffacts representing a HOPI model of two workflows, original and altered – is stored in a text file maintained with any text editor. In keeping with the usage model of Fig. 7, in the WD’s the original workflow definition is termed internal (to the contracting company) and the altered definition in use at the sub- contracting site is termed external. A partial listing of HOPI data structures definitions and data instances is shown in Table 2. Pseudo-code for one of the heuristic rules that reschedule triggers within the altered workflow definition is shown in Fig. 9. artial HOPI instantiation deffacts case1ExtGt ‘‘goal tree for external process, case 1X’’ (goalNode (name XverifyReferences) (parent xVerifyPersonalInfo) (workProcess XsecurityClearance) (goalNode. . . deffacts case1Acts ‘‘activity sequence for case 1X’’ (activity (name XphoneFormerEmployers) (parent Xvalidate2) (scheduledDuration 48) (predecessor XfaxSchool) (successor XfaxToFBI) (workProcess XsecurityClearance) (activity. . . Fig. 10. Simple activity substitution (equivalent function, same level). Fig. 9. A trigger rescheduling heuristic. 560 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 6. System run with test cases 6.1. Coordination disruption exemplars (test cases) Any workflow activity change1 of arbitrary complexity can be resolved into a repeated, sequential application of the elementary change patterns diagramed below in Figs. 10–12: (1) simple activity substitution involving a termino- logical change (identical function, different terminology), (2) multiple activities are subsumed in a new, more encom- passing process step, (3) a single process step is articulated into multiple activities. The cases we describe are patterns as defined in Andersson et al. (2005) that provide the isomor- phic mappings from an original process to an altered process. 6.2. System performance under test cases Running the prototype with test cases is straightfor- ward. The coordination subsystem code (HOPI.clp) is 1 Process redesign (substantial, deliberate activity resequencing) is not discussed in this paper nor addressed by the coordination subsystem in its current form. loaded into CLIPS using the (load) command. When run, the system asks for the name of a case file of facts defining an original and altered workflow as described above. Out- put from a run of the system with the elementary activity substitution example described in Section 1 (manufacture of a women’s suit) is shown in Fig. 13 (cf. discussion of context similarity and intentional similarity in Section 3). If attached to a working WFMS, the subsystem could transmit the new, reconciled workflow – the altered work- flow with the trigger activity from the original workflow rescheduled within it – directly to the subcontractor’s workflow enactment engine as described in the usage model of Fig. 7. Alternatively, the reconciled workflow could be transmitted to supervisory personnel for review prior to transmission to the workflow enactment engine. 7. Summary and future work The functionality of the coordination subsystem in its present form is useful in many real-world situations since many alterations to work processes conform to our elemen- tary change types. More complex changes typically evolve from the execution of multiple elementary changes over Fig. 11. New activity subsumes original activity (goal equivalent; level shifted up in the abstraction hierarchy). Fig. 12. Articulation of original activity (goal equivalent; level shifted down in the abstraction hierarchy). W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 561 extended periods of time. Encouraged by the performance of the coordination model and prototype to date we have projects either planned or in process to extend this work in two directions. First, we wish to validate and extend the heuristics used in the similarity computations and rescheduling rules of the model. We have already performed a pilot study to validate the instruments that we will use in a full concurrent verbal protocol exploration of expert process coordination prob- lem solving. Analysis of the protocols should yield greater understanding of how experts reason when modifying coordinated inter-organizational processes. This experi- ment will validate the cognitive model on which HOPI is based. A second research direction seeks to make HOPI and its associated re-coordination system more practical by par- tially automating the construction of the goal hierarchy through the machine analysis of textual documentation on a process. Most work processes in large organizations are accompanied by process manuals, design documents, memoranda to personnel involved in the process and other textual material. Using many of the same natural language processing (NLP) techniques that are currently used to monitor large bodies of e-mail and other textual mate- rial and for conceptual clustering of www inquiry results, we hope to construct tools for extracting process goals and functions from textual material ancillary to a work process. Fig. 13. Re-coordination system output. 562 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 The same NLP techniques can be used to automatically construct a lexicon (process ontology) that describes the process, its activities, its performers and their relationships. Using a lexicon, distributed artificial intelligence inference techniques (Vincenzo & Sessa, 2001) can be used to signif- icantly extend the reasoning capabilities of the coor- dination subsystem. For example, a change from one functionally equivalent process step to another involving only a change of tool could be identified with much greater confidence using a lexicon. Inferences as to the similarity of new goals to existing goal/subgoal chains would be likewise enhanced. An additional project which we hope to launch as soon as practical is to embed the coordination subsystem in an actual working inter-departmental workflow system pro- posed for the university of the first author. The information gained from real-world experience with the system would be impossible to gather any other way, and we look for- ward to the challenges this project holds. References Agostini, A., & De Michelis, G. (2000). In W. v. d. Aalst, J. Desel, & A. Oberweis (Eds.), Improving flexibility of workflow management sys- tems.. Business process management: Models, techniques and empirical studies (1806 of LNCS, pp. 218–234). Berlin: Springer Verlag. Andersson, B., Bider, I., Johannesson, P., & Perjons, E. (2005). Towards a formal definition of goal-oriented business processes patterns. Business Process Management Journal, 11(6), 1–14. Available fromhttp:// www.ibissoft.se/publications/publications.htm. Ang, J., & Hong, J. (1994). AMS formalism: an approach to office modeling and OIS development. DataBase, 25(4), 25–38. Barsalou, L. W. (1983). Ad hoc categories. Memory and Cognition, 11, 211–217. Boland, R., Maheshwari, A., Te’eni, D., Schwartz, D., & Tenkasi, R. (1992). Sharing perspectives in distributed decision making. In Proceedings of CSCW ‘92, November, 1992 (pp. 306–313). Buhler, P., & Vidal, J. (2005). Towards adaptive workflow enactment using multiagent systems. Information Technology and Management, 6, 61–87. Cao, T., & Sanderson, A. C. (1995). Task sequence planning using fuzzy Petri nets. IEEE Transactions on Systems, Man and Cybernetics, 25(5), 755–768. Chandrasekaran, B. (1986). Generic tasks in knowledge based reasoning: high level building blocks for expert system design. IEEE Expert, 3, 23–30. Davenport, T., Jarvenpaa, S., & Beers, M. (1996). Improving knowledge work processes. Sloan Management Review(Summer), 53–65. Ellis, C., & Wainer, J. (1994). Goal-based models of collaboration. Collaborative Computing, 1, 61–86. Georgakopoulos, D., Sheth, A., & Hornick, M. (1994). An overview of workflow management: from process modeling to workflow automa- tion infrastructure. International Journal of Distributed and Parallel Databases, 3(2). Gerson, E., & Star, S. (1986). Analyzing due process in the workplace. ACM Transactions on Office Information Systems, 4(3), 257–270. Giarrantano, J., & Riley, G. (2005). Expert systems: principles and programming. Boston, MA: Thomson. Hewitt, C. (1986). Offices are open systems. ACM Transactions on Office Information Systems, 4(3), 271–287. Hewitt, C. (1991). Open information systems semantics for distributed artificial intelligence. Artificial Intelligence., 47(79), 79–106. Joosten, S. (1994). Trigger modeling for workflow analysis. Proceedings of CON ’94 (Workflow Mgt. Div.). Vienna: R. Oldenburg. Karbe, B., Ramsperger, N., & Weiss, P. (1990). Support of cooperative work by electronic circulation folders. In Proceedings of the 1990 ACM COIS ‘90 (pp. 109–117). Khomyakov, M., & Bider, I. (2001). Achieving workflow flexibility through taming the chaos. Journal of Conceptual Modeling, 21, 1–14. Klein, M., Dellarocas, C., & Bernstein, A. (Eds.). (2000). Adaptive workflow systems [Special issue]. Journal of Computer Supported Cooperative Work, 9, 3–4. Kuechler, W., Vaishnavi, V., & Kuechler, D. (2001). Supporting optimi- zation of business to business e-commerce relationships. Journal of Decision Support Systems, (31), 363–377. Lee, J., & Malone, T. (1990). Partially shared views: a scheme for communicating among groups that use different type hierarchies. ACM Transactions on Information Systems, 8(1), 1–26. http://www.ibissoft.se/publications/publications.htm http://www.ibissoft.se/publications/publications.htm W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 563 Lei, Y., & Singh, M. (1997). A comparison of workflow metamodels. In Stephen W. Liddle (Ed.), Proceedings of the ER’97 workshop on behavioral models and design transformations: Issues and opportunities in conceptual modeling. Available from http://osm7.cs.byu.edu/er97/ workshop4/ls.html. Mahling, D., & King, R. (1999). A goal-based workflow system for multiagent task coordination. Journal of Organizational Computing and Electronic Commerce, 9(1), 57–82. Morita, M., Mukaigaito, T., & Hayami, H., (1996). HFBP: a hierarchal and functional business process model for describing distributed autonomous business environments. In Proceedings of the NSF workshop on workflow and process automation in information systems, Athens, GA, UGA, 1996 (pp. 91–99). Muller, R., Greiner, U., & Rahm, E. (2004). AgentWork: a workflow system supporting rule-based workflow adaptation. Data and Knowl- edge Engineering, 51, 223–256. O’Leary, D., Kuokka, D., & Plant, R. (1997). Artificial intelligence and virtual organizations. Communications of the ACM, 40(1), 52–59. Riley, G. (2005). Clips reference manual: Volume 1 – Basic programming guide, June, 2005. Available from http://www.ghg.net/clips/clips.html. Schank, R., & Abelson, R. (1977). Scripts, plans, goals and understanding: an inquiry into human knowledge. Hillsdale, NJ: Lawrence Erlbaum. Schank, R., & Abelson, R. (1995). Knowledge and memory: the real story. Hillsdale, NJ: Lawrence Erlbaum. Schulz, K., & Orlowska, M. (2004). Facilitating cross-organizational workflows with a workflow view approach. Data and Knowledge Engineering, 51, 109–147. Schuster, H., Georgakopoulos, D., Cichocki, A., & Baker, D. (2000). Modeling and Composing Service-Based and Reference Process-Based Multi-enterprise Processes. In B. Wangler & L. Bergman (Eds.). Proceedings of advanced information systems engineering: 12th interna- tional conference, CAiSE 2000, Stockholm, Sweden, June 2000 (LNCS 1789/2000). Berlin: Springer Verlag. Simon, H. (1977). The new science of management decision (Revised Edition). Englewood Cliffs, NJ: Prentice Hall. Soffer, P. (2005). On the notion of flexibility in business processes. In Proceedings of the 6th workshop on business process modeling, devel- opment and support, Porto, Portugal. Available from http://lam- swww.epfl.ch/conference/bpmds05/program/soffer_4.pdf. Tansley, D., & Hayball, E. (1993). Knowledge-based systems analysis and design: a KADS developers handbook. New York: Prentice Hall. Vaishnavi, V., Joosten, S., & Kuechler, W. (1996). Representing workflow management systems with smart objects. In Proceedings of NSF workshop on workflow and process automation in information systems, Athens GA. Vaishnavi, V., & Kuechler, W. (2005). Universal enterprise integration: challenges of and approaches to web-enabled virtual organizations. Information Technology and Management, 6(1), 5–16. van der Aalst, W., & Jablonski, S. (2000). Dealing with workflow change: identification of issues and solutions. International Journal of Computer Systems, Science and Engineering, 15(5), 267–276. van der Aalst, W., Weske, M., & Grunbauer, D. (2005). Case handling: a new paradigm for business process support. Data and Knowledge Engineering, 53, 129–162. Vincenzo, L., & Sessa, S. (Eds.). (2001). Soft computing agents: new trends for designing autonomous systems. Heidelberg [Germany], New York: Physica-Verlag. Yu, E. (1995). Models for supporting the redesign of organizational work. In Proceedings of the Conference on Organizational Computing Systems (pp. 225–236). Milpitas, CA: ACM Press. http://osm7.cs.byu.edu/er97/workshop4/ls.html http://osm7.cs.byu.edu/er97/workshop4/ls.html http://www.ghg.net/clips/clips.html http://lamswww.epfl.ch/conference/bpmds05/program/soffer_4.pdf http://lamswww.epfl.ch/conference/bpmds05/program/soffer_4.pdf An expert system for dynamic re-coordination of distributed workflows Introduction Search for flexibility in workflow and process modeling Goal-based coordination model framework An architectural view of the coordination subsystem Process overview Coordination system implementation System run with test cases Coordination disruption exemplars (test cases) System performance under test cases Summary and future work References