ciency 0 duction Systems based uth Maintenance

Expert systems in complex domains require rich knowledge representation formalisms and problem solving paradigms. A typical framework may involve a blackboard architecture and a Reason Maintenance System (RMS) to guarantee the consistency of the links between the blackboard nodes. However, in order to satisfy computational feasibility and become operational, the resulting expert system must often be rewritten using less expressive tools. We propose an architecture integrating efficiently an OPS-like inference engine and an Assumption based Truth Maintenance System (AT&IS). These paradigms have been separately investigated and extended. Roles distribution between an ATMS and an inference engine integrated in a single framework is one of the major issues to obtain good overall performance. Two architectures will be studied : loose coupling, where the ATMS and the inference engine are clearly separated, and tight coupling where the ATMS is intimately integrated with the match phase of a RETEbased inference engine. The advantages and drawbacks of both solutions are described in details. Finally, future work is discussed. Expert systems in complex domains require rich knowledge representation formalisms and problem solving paradigms. Commercially available expert system shells provide some compromise between expressiveness and tractability. A forward chaining engine with an OPS-like rule language is one of the key components of such shells. Its operation involves a match-select-act cycle: 1. Match : The condition part of each rule is compared to the content of the fact base (or working memory). If a set of facts conjointly satisfy all the Thii work has been supported in part by the DRET (French DARPA) under grant number 89/568. 268 EXTENSIONS OF TRUTH MAINTENANCE conditions, the rule is said to be instantiated. One or many rule instantiations may thus be found, and are queued in a list of executable operators, called the conflict set or the agenda. Select : One or more rule instantiations are selected from the agenda for future execution of their action part. Selection is done according to some user defined conflict resolution strategy. Predefined strategies usually include FIFO, LIFO, highest priority, and more. Act : The right-hand-side actions of the selected rule, or rules, are executed. These actions may modify the fact base, which will possibly instantiate new rules. Having the possibility to retract facts from the working memory is necessary in many applications. When allowing this, one should be aware that the conclusions derived from the removed facts are not necessarily valid anymore. And when there are contradictions in the fact base, the system may not be able to pursue its reasoning process. To avoid handling these problems manually, Reason Maintenance Systems @MS) have been developed. Expert systems using a RMS store justifications : a justification is a link between a fact created on the righthand-side of a rule and the facts which instantiated this rule. Let us illustrate this through an example in an OPSlike syntax : Rule base; (Rule birds-fly (Bird ?x) + (assert (fly ?x))) Fact base: (Bird Tweety) In this example, the justification (Bird Tweety) + (Fly Tweefy) will be created. When retracting a fact, the system follows the links established by the justifications, to retract not only the desired fact, but also all the facts it enabled to derive. Among the different RMSs, the ATMS (Assumption based Truth Maintenance System) became very popular in the last few years. ATMSs are a convenient way of exploring many choices in parallel when solving a problem. From: AAAI-91 Proceedings. Copyright ©1991, AAAI (www.aaai.org). All rights reserved.