A niching scheme for steady state GA-P and its application to fuzzy rule based classifiers induction

Abstract

A new method for applying grammar based Genetic Programming to learn fuzzy rule based classifiers from examples is proposed. It will produce linguistically understandable, rule based definitions in which not all features are sent in the antecedents. A feature selection is implicit in the algorithm. Since both surface and deep structure will be learned, standard grammar based GP is not applicable to this problem. We have adapted GA-P algorithms, a method formerly defined as an hybrid between GA and GP, that is able to perform a more effective search in the parameters space than canonical GP do. Our version of GA-P supports a grammatical description of the genotype, a syntax tree based codification (which is more efficient than parse tree based representations) and a niching scheme which improves the convergence properties of this algorithm when applied to this problem