Effects of the Use of Multiple Fuzzy Partitions on the Search Ability of Multiobjective Fuzzy Genetics-Based Machine Learning

Abstract

An important issue in the design of fuzzy rule-based systems is to find a good accuracy-complexity tradeoff. While simple fuzzy systems with high interpretability are usually not accurate, complicated fuzzy systems with high accuracy are usually not interpretable. Recently evolutionary multiobjective optimization (EMO) algorithms have been used to search for simple and accurate fuzzy systems. The main advantage of EMO-based approaches over single-objective techniques is that a number of alternative fuzzy systems with different accuracy-complexity tradeoffs can be obtained by their single run. We have already proposed a multiobjective fuzzy genetics-based machine learning (GBML) algorithm for pattern classification problems. In our GBML algorithm, multiple fuzzy partitions with different granularities are simultaneously used. This is because we usually do not know an appropriate fuzzy partition for each input variable. However, the use of multiple fuzzy partitions significantly increases the size of the search space. In this paper, we examine the effect of the use of multiple fuzzy partitions on the search ability of our multiobjective fuzzy GBML algorithms through computational experiments.