Strategy Parameter Variety In Self-adaptation Of Mutation Rates

Self-adaptation has been widely used in Evolution Strategies (ES) and Evolutionary Programming (EP), where it has proved useful in varying the mutation step size for continuous objective variables. To date, relatively little work has been performed on applying selfadaptation to the canonical Genetic Algorithm (GA). This research applies a simple discrete model of self-adaptation to test functions with differing characteristics. We show that the discrete model is able to provide more reliable problem solving than the classical lognormal self-adaptation scheme on the test problems examined. We find that although self-adaptation parameter choices representing conventional thinking perform best for unimodal functions, very different parameter settings are required for optimum performance on multimodal functions. These results are discussed in terms of the strategy parameter variety needed for selfadaptation to work effectively and we outline a self-adaptation mechanism designed to capitalize on these findings.