Learning Control Information in Rule-Based Systems: a Weak Method

The authors present a weak method for learning control information in rule-based systems. The method proposes establishing 'connections' between rules each time the system successfully performs a task. Connections have associated strengths which code the history of the success/failure of sequences of rule firings during problem solving. Connection strength is used to provide guidance to the inference engine in two important ways: selection of the next rule to consider for matching, and a basis for a conflict resolution scheme. The connection strength is also used in guiding the generation of rules through composition. It insures that composition will only occur on sequences that have established their utility to the system through problem solving experience. The authors assert that the method proposed will, in general, result in significantly decreasing the matching and search efforts. The method has been implemented in a program called STAC and some favorable results have been obtained in the test domain of Euclidean geometry proofs.<>