A hybrid fuzzy genetics-based machine learning algorithm: hybridization of Michigan approach and Pittsburgh approach

In this paper, we propose a hybrid genetics-based machine learning algorithm for designing a linguistic classification system that consists of a small number of fuzzy if-then rules with clear linguistic interpretation. Our task is to generate a small number of fuzzy if-then rules from numerical data for a high-dimensional pattern classification problem. We assume that a set of linguistic values is given for each attribute of the pattern classification problem by human experts. Thus our task is described as finding a small number of combinations of linguistic values, each of which is used as the antecedent part of a fuzzy if-then rule. While this task seems to be simple at a glance, it is very difficult especially in the case of high-dimensional problems because the number of possible combinations of antecedent linguistic values exponentially increases with the dimensionality of problems. That is, the search space for high-dimensional problems is terribly huge. In our approach, an individual in genetic algorithms is a set of fuzzy if-then rules. Each rule is coded as a string by its antecedent linguistic values. Thus an individual (i.e., a rule set) is denoted by a concatenated string. The fitness of a rule set, which is defined by its classification performance, is used in a selection operation. New rule sets are generated by a crossover operation from selected rule sets. A mutation operation modifies a part of each rule set generated by the selection and crossover. As a mutation operation, we use a rule generation mechanism of the Michigan approach.