Study and Formalization of a Case-Based Reasoning System with a Description Logic

1 Introduction Keywords: Organization of the case base Case retrieval Case selection description logic Case-Based Reasoning system index concepts automatic concept clas-sication Similarity Least Common Subsumer dis-similarity subsumption relation supervised learning step least common subsumption index concepts classication realization realization In this paper the C-CLASSIC is used to design a with homogeneous, explicit and formal criteria. The case base is organized by means of a taxonomy of. Case retrieval is performed using the of the description logic. between two cases is characterized by the most specic concept which subsumes the two cases (or LCS), and by a concept representing properties which belong to one case but not to the other. A partial order induced by the on these concepts is used to select the most similar cases. Finally, the paper proposes a of index concepts using a learning algorithm whose main operation is LCS computation. In most systems the dierent tasks of Case-Based Reasoning (CBR) are performed using heterogeneous methods which generally rely on ad hoc, application-dependent criteria. What is proposed here is to use a Description Logic (DL) to design a CBR system with homogeneous, formal, application-independent criteria. The ability to automatically build taxonomies of concepts and to link individuals to concepts is an interesting property for the management of a case base. Besides, DLs are restrictions of First Order Logic in which the subsumption relationship has been largely and formally studied. C-CLASSIC (see [2]) is a description logic expressive enough to be of practical use in which the computation of subsumption and are polynomial and in which it is possible to learn with respect to the Valiant paradigm (i.e. C-CLASSIC is PAC-learnable see [12] and [5]). This paper denes homogeneous, explicit criteria in C-CLASSIC to formalize the following tasks of CBR: 1. : the stored cases are represented as individuals and are indexed by concepts organized in a taxonomy. The links between are automatically computed using. The links between cases and index concepts are automatically computed using. 2. : the new case is represented as an individual. The index concepts which the case has to be linked to are found by applying the operation to this individual. Then, the stored cases which share at least one index concept with the new case are retrieved. 3. : each retrieved case is compared with the new case using two criteria, similarity and dissimilarity. Similarity between two cases is characterized by …