Artificial neural network based crossover for evolutionary algorithms

摘要

Recombination is a powerful way of generating new solutions in Evolutionary Algorithms. There are many ways to implement recombination. Traditional recombination operators do not use information about parents, evolutionary process, or models for variable interaction in order to find better ways to recombine solutions. Some modern recombination operators use information about parents and models for variable interaction, but they cannot always be efficiently applied. We propose to use an artificial neural network to compute the recombination mask, given two parents. Here, a radial basis function network (RBFN) is trained online using past successful recombination cases obtained during the optimization performed by the evolutionary algorithm. The RBFN crossover (RBFNX) is used together with other recombination operators (here, uniform crossover is employed). Applying RBFNX has O(N) time complexity, where N is the dimension of the optimization problem. Results of experiments with genetic algorithms, applied to two binary optimization problems, and evolution strategies, applied to continuous optimization test problems, indicate that RBFNX is generally able to improve the successful recombination rates. (C) 2020 Elsevier B.V. All rights reserved.