Genetic lateral tuning of membership functions as post-processing for hybrid fuzzy genetics-based machine learning

Abstract

Genetic fuzzy systems (GFS) have been actively studied in the field of fuzzy classifier design. GFS can generate simple and accurate classifiers with a number of fuzzy if-then rules by evolutionary computation (EC). One general concern is how to discretize numerical attributes into fuzzy partitions. Most GFS use homogeneous fuzzy partitions without considering the class distribution of each attribute. This is the simplest idea, but more accurate classifiers could be obtained by optimizing fuzzy partitions. There are three approaches. One is pre-processing where inhomogeneous fuzzy partitions are specified according to the class distribution before applying EC. Another approach is to simultaneously optimize both a set of fuzzy if-then rules and fuzzy partitions by EC. The other is post-processing where fuzzy partitions used in the obtained classifier are optimized afterward. In this paper, we examine the effect of post-processing where fuzzy partition optimization is applied to the obtained classifier by GFS. In computational experiments, we first use our hybrid fuzzy genetics-based machine learning with homogeneous fuzzy partitions for standard data sets and its parallel distributed implementation for large data sets. Then we apply genetic lateral tuning as post-processing to shift the positions of membership functions according to the pattern distribution.