A Comprehensive Study on Subgraph Crossover in Cartesian Genetic Programming

Roman Kalkreuth

Abstract

While tree-based Genetic Programming is often used with crossover, Cartesian Genetic Programming (CGP) is mostly used only with mutation as the sole genetic operator. In contrast to comprehensive and fundamental knowledge about crossover in tree-based GP, the state of knowledge in CGP appears to be still ambiguous and ambivalent. Two decades after CGP was officially introduced, the role of recombination in CGP is still considered to be an open and remaining question. Although some promising steps have been taken in the last years, comprehensive studies are needed to evaluate the role of crossover in CGP on a large set of problems. In this paper, we take a step forward on the crossover issue by comparing algorithms that utilize the subgraph crossover technique which has been proposed for CGP to the traditional mutation-only CGP. Experiments on well-known symbolic regression and Boolean function problems demonstrate that the use of algorithms that utilize the subgraph crossover outperform the mutation-only CGP on well-known benchmark problems.

Download


Paper Citation