A new approach to genetics based machine learning in fuzzy controller design

This paper proposes an evolutionary approach to fuzzy controller design based on the "Pittsburgh" style classifier system in which whole rule-sets are the unit of credit assignment and selection. We present a description of a system based on this idea, together with experimental results using the system to learn function identification. The representation used allows the genetic algorithm to vary both membership functions (centres and widths) and fuzzy relations. We introduce a new crossover operator which employs crosspoints in the input space and demonstrate its efficacy. Finally, we present results which show that the classifier system is capable of self-organisation of membership functions and fuzzy relations simultaneously.<>