Comparison of different fitness functions in genetic fuzzy rule selection
Masakazu Yamane, Akihito Ueda, Naoshi Tadokoro, Yusuke Nojima, Hisao Ishibuchi · 2012
Genetic fuzzy rule selection is a well-known fuzzy rule-based classifier design method. It searches for a simple but accurate fuzzy rule-based classifier by simultaneously minimizing the error rate and some complexity measures such as the number of rules and the total rule length. In this paper, we propose the modification of a fitness function to improve the generalization ability of fuzzy classifiers on test patterns. Two additional terms of the fitness function are defined using the margin in the winner rule selection criterion between the target class and the other classes for each training pattern. It is expected that the decision boundary will be moved towards the center region between different classes by the added terms. It is also expected that the proposed modification can make the fitness landscape smoother, which will improve the search performance of genetic fuzzy rule selection. We examine these expected effects of the proposed modification through computational experiments.