Approximate reasoning to learn classification rules

Abstract

In this paper, we propose an original use of approximate reasoning not only as a mode of inference but also as a means to refine a learning process. This work is done within the framework of the supervised learning method SUCRAGE which is based on automatic generation of classification rules. Production rules whose conclusions are accompanied by belief degrees, are obtained by supervised learning from a training set. These rules are then exploited by a basic inference engine: it fires only the rules with which the new observation to classify matches exactly. To introduce more flexibility, this engine was extended to an approximate inference which allows to fire rules not too far from the new observation. In this paper, we propose to use approximate reasoning to generate new rules with widened premises: thus imprecision of the observations are taken into account and problems due to the discretization of continuous attributes are eased. The objective is then to exploit the new base of rules by a basic inference engine, easier to interpret. The proposed method was implemented and experimental tests were carried out.