Abstract
Fuzzy rule-based classifiers are well-known as having high accuracy and interpretability for human users. There are several techniques to design a fuzzy rule-based classifier. Fuzzy genetics-based machine learning (FGBML) is one of them. It is, however, difficult to design the best classifier with respect to both accuracy and interpretability because of their tradeoff. Furthermore, FGBML takes heavy computational load if applied to large data sets. To address these problems, we proposed multiobjective FGBML (MoFGBML) and its parallel distributed implementation in the previous study. MoFGBML can design a number of classifiers along with the tradeoff and parallel distributed implementation can reduce the computation time. Through computational experiments, we found that the computation time became much shorter by applying parallel distributed implementation to our MoFGBML, but the number of the obtained non-dominated classifiers became small. Moreover, classifiers with high accuracy were not obtained for some data sets. In this paper, we propose a simple idea to bias the search direction of our MoFGBML to obtain classifiers with high accuracy. Our idea is to rotate one or two objective functions in our MoFGBML. This rotation changes the domination relation in multiobjective optimization. We examine the rotation effects of the objective functions on the search ability of our MoFGBML through computational experiments using large data sets.
