Dynamic Penalty Based GA for Inducing Fuzzy Inference Systems

Fuzzy based models have been used in many areas of research. One issue with these models is that rule bases have the potential for indiscriminant growth. Inference systems with large number of rules can be overspecified, have model comprehension issues and suffer from bad performance. In this research we investigate the use of a genetic algorithm towards the generation of a fuzzy inference system (FIS). We propose using a GA with a dynamic penalty function to manage the rule size of the fuzzy inference system (FIS) while maintaining the exploration of good rules. We apply this method towards the generation of a fuzzy classifier for the search of metabolic pathways. The GA based FIS includes novel mutation and a penalty based fitness scheme which enables the generation of an efficient and compact set of fuzzy rules. Encouraging implementation results are presented for this method as compared with other classification methods. This method should be applicable to a variety of other modelling and classification problems.