Migration through Mutation Space: A Means of Accelerating Convergence in Evolutionary Algorithms

Abstract

In this paper a multiple subpopulations technique for evolutionary algorithms is proposed. Each subpopulation is distinguished by the mutation radius and mutation probability assigned to it, with mutation radius being a function of mutation probability. Mutation probabilities across the subpopulations range from 0.005 to 0.75. There is no crossover between subpopulations in the normal course of breeding, and the mechanisms of elite migration from a higher mutation subpopulation to the adjacent subpopulation of lower mutation is used to introduce new genetic material. The evolution of artificial neural networks for solving a variety of problems is demonstrated, with convergence times typically half as long as a standard evolutionary algorithm.