Generation of fuzzy models via evolutionary strategies

This paper presents a framework for studying the effectiveness of evolutionary strategies for generating fuzzy rule bases and function approximations from training data. To facilitate the evolutionary operations that modify the elements of the population, a fuzzy rule base is represented as a real-valued matrix. A comparison of the training data with the function approximation associated with a fuzzy rule base provides a measure of agreement of the rule base with the training data. The analysis of training data provides the ability to generate both global and local fitness assessments. The effectiveness of incorporating local information into the evolutionary search is demonstrated by comparing the generation of rule consequences using the global and local strategies.