Simple modifications on heuristic rule generation and rule evaluation in Michigan-style fuzzy genetics-based machine learning

Fuzzy genetics-based machine learning (FGBML) is one of the representative approaches to obtain a set of fuzzy if-then rules by evolutionary computation. A number of FGBML methods have been proposed so far. Among them, Michigan-style approaches are popular thanks to thier lower computational cost than Pittsburgh approaches. In this study, we introduce two simple modifications for our Michigan-style FGBML. One is related to heuristic rule generation. In the original FGBML, each rule in an initial population is generated from a randomly-selected training pattern in a heuristic manner. The heuristic rule generation also performs during evolution where each rule is generated from a misclassified pattern. As its modification, we propose the use of multiple patterns to generate each fuzzy if-then rule. The other is related to the fitness calculation. In the original FGBML, the fitness of each rule is calculated as the number of correctly classified training patterns, while the number of misclassified patterns is ignored. As its modification, we incorporate a penalty term into the fitness function. Through computational experiments using 20 benchmark data sets, we examine the effects of these two modifications on the search ability of our Michigan-style FGBML.