Opposites attract: Complementary phenotype selection for crossover in genetic programming

Abstract

Standard crossover in genetic programming (GP) selects two parents independently, based on fitness, and swaps randomly chosen portions of genetic material (subtrees). The mechanism by which the crossover operator achieves success in GP, and even whether crossover does in fact exhibit relative success compared to other operators such as mutation, is anything but clear [14]. An intuitive explanation for successful crossover would be that the operator produces fit offspring by combining the “strengths” of each parent. However, standard selection schemes choose each parent independently of the other, and with regard to overall fitness rather than more specific phenotypic traits. We present an algorithm for choosing parents which have complementary performance on a set of fitness cases, with an eye toward enabling the crossover operator to produce offspring which combine the distinct strengths of each parent. We test Complementary Phenotype Selection in three genetic programming domains: Boolean 6-Multiplexer, Intertwined Spirals Classification, and Sunspot Prediction. We demonstrate significant performance gains over the control methods in all of them and present a preliminary analysis of these results.