Maintaining the Diversity of Michigan-Style Approaches for Construction Fuzzy Classification System

摘要

Michigan-style genetic algorithms are usually used for learning fuzzy classification rules from numerical examples. In these approaches, each rule is encoded as a chromosome, and then builds up the classification rule set by these chromosomes. So the fitness value can only assign to a single rule rather than a whole rule set. This makes some chromosomes characterized by the minority of instances may be lost from the gene pool, and the approaches can only learn from small subset of the search space. In this paper, we first define the similarity level of one fuzzy rule from another rule using similarity measure. With the similarity level, we then balance the fitness values of different chromosomes by using fitness sharing method, and maintain the diversity of population. So the approaches can not only learn from the major instances, but also to learn from the minor instances. Furthermore, we cache the similarity value of different antecedent fuzzy sets for reducing the computing load when the similarity value are calculated. Finally, experimental results on benchmark classification problems demonstrate that our method is able to efficiently achieve accurate performance.