Rule Exclusion Mechanism in Evolutionary Fuzzy Systems

Abstract

This paper aims to propose a rule exclusion system and, consequently, the model simplification in evolutionary fuzzy systems. Such simplification has some benefits, being highlighted, for example, the task of labelling the rules by an expert in unsupervised systems and the explanation of the rules obtained. For the execution of the work, it was considered an algorithm present in the literature, ALMNo, with the addition of the proposed exclusion mechanism. The proposed mechanism uses the distance between the centers of the membership functions of the rules, normalized by the standard deviation of a sliding window with the last 10 data analyzed. The normalization is intended to detect a change in the context of the data, and once it is detected, provides greater generalizability to the system. This is due to the fact that data belonging to another region of space generates a larger standard deviation. The results were analyzed by comparing the original ALMNo algorithm with that without the exclusion mechanism. Numerical results show that the proposed mechanism is promising in terms of reducing the number of rules and maintaining a competitive level of accuracy. Furthermore, test results indicate that setting the necessary parameters is not decisive for the success of the algorithm.