A criterion of comparison between two case bases

This paper presents a criterion of comparison between two case bases for a case-based system for which the retrieval process is done thanks to a similarity metric. Such a criterion can be useful for at least two things. First, it allows to define what a better case base of a given size can be. Second it enables us to build a “forgetting criterion” which aims at answering the question “What are the p cases that can be lost from the base that conduct to a minimal loss of performance?” The definition of case-based reasoning given in this paper stresses the fact that the goal of the similarity measure is to give an o priori estimation of the performance of the adaptation process. With an introducing example -the “locksmith's problem”- and then more generally, the criterion is defined thanks to a number associated with each case base; this number is characteristic of the mean performance of the system working with this base. It seems that very few results can be demonstrated without hypothesis on the representation of problems. For this reason, we have performed a study for two different types of representation of these problems (or on their associated indices). Finally we solve the locksmith's problem.