Genetic programming: the ratio of crossover to mutation as a function of time

Abstract

This article studies the sub-tree operators: mutation and crossover, within the context of Genetic Programming. Two standard problems, symbolic linear regression and a non-linear tree, were presented to the algorithm at each stage. The behaviour of the operators in regard to fitness is first established, followed by an analysis of the most optimal ratio between crossover and mutation. Subsequently, three algorithms are presented as candidates to dynamically learn the most optimal level of this ratio. The results of each algorithm are then compared to each other and the traditional constant ratio.