1 2 3 4 5 6 7



Michel MANAGO (1), Klaus-Dieter ALTHOFF (2), Eric AURIOL(1), Ralph TRAPHÖNER (3),
Stefan WESS
(2), Noël CONRUYT (1), Frank MAURER(2)



We present the INRECA european project (ESPRIT 6322) on integration of induction and case- based reasoning (CBR) technologies for solving diagnostic tasks. A key distinction between case-based reasoning and induction is given in [1]: "In case-based methods, a new problem is solved by recognising its similarities to a specific known problem then transferring the solution of the known problem to new one (...) In contrast, other methods of problem solving derive a solution either from a general characterisation of a group of problems or by search through a still more general body of knowledge". In this paper, we distinguish between a pure inductive approach and a case-based one on the basis that induction first computes an abstraction of the case database (ex: a decision tree or a set of rules) and then uses this general knowledge for problem solving. During the problem solving stage, the system does not access the cases.


INRECA's inductive and case-based approaches

Induction is a technology that automatically extracts general knowledge from training cases. KATE is the inductive component of INRECA. It builds a decision tree from the cases by using the same search strategy, hill-climbing, and same preference criteria that is based on Shannon's entropy as ID3[2]. Unlike most induction algorithms, KATE can handle complex domains where cases are represented as structured objects with relations and it can use background knowledge. At each node, KATE generates the set of relevant attributes of objects for the current context and selects the one that yields the highest informationgain. For instance, an attributes such as "pregnant" for a patient whose sex is known to be "male" further up in the decision tree is eliminated before the information gain computation. Background domain knowledge and class descriptions allow to constrain the search space during induction [3].

Case-based reasoning is a technology that makes direct use of past experiences to solve a new problem by recognising its similarity with a specific known problem and by applying the known solution to the new problem. PATDEX is the case-based component of INRECA. It consists of two case-based reasoning subcomponents for classification and test selection. A procedure that dynamically partitions the case base enables an efficient computation and updating of the similarity measures used by the CBR subcomponents. For the classification subcomponent, the applied similarity measures are dynamic. The underlying evaluation

IMAGE imgs/Annexe601.gif