279

Induction and Reasoning from Cases

a decision tree produced by induction can speed up the consultation by the case-based reasoner. The case-based reasoner can supplement the decision tree when choosing among different conclusions (case-based reasoning is started at the end of the consultation of the tree or during consultation when encountering unknown values). The third level of integration allows the combination of individual modules of the tools (workbench strategy). For instance, the information gain measure module may be used to choose the next attribute to be asked during an interactive CBR consultation. The last level fulfils the final goal of INRECA(seamless integration) by mixing the most relevant parts of the two technologies in a single system. Two critical modules are identified: the information gain computation module for the induction technique, and the similarity computation module for the case-based reasoning technique.

Our main point is that a single system will never meet the needs of everyone. INRECA offers several integration possibilities and must be configured to meet the requirements of a particular application or of a particular category of users. For instance, a naive end-user must be guided step-by-step by the consultation system in a decision-tree like fashion. On the other end, a domain specialist wants to directly supply whatever information he feels is relevant and remain in control of the consultation system. Moreover, what may be viewed as an advantage of a technology in a given context may turn out to be a drawback in another. For instance, incrementality can be seen as an advantage of CBR over induction to maintain the consulation system automatically and keep up with the knowledge that workers learn through their daily experience. On the other end, we are currently working with an equipement manufacturer who distributes the diagnostic system to his customers and who wants to control the advices that are given to the users (let it be for legual reasons). Thus, he prefers a system that does not evolve permanently and that behaves in a predictable way. In that context, the incrementality is a drawback since he wants to compile the case data into an induction tree that is maintained by him periodically. Finally, one technique may be better adapted at a specific stage of the application life cycle (for example, CBR at the begining to enrich the case database) but not at a later stage (for example, induction can compile the case database when it becomes too big and when efficiency becomes a problem). Thus, INRECA provides several options for the four levels of integration and can be configurated by the application developper . In the next section, we present an architecture that deals with the problem of handling unknown values using CBR, but that pre-index the cases using a decision tree for efficiency.