Crossover in Cartesian Genetic Programming: Evaluation of Two Phenotypic Methods

Roman Kalkreuth

Abstract: The tree-based representation model of Genetic Programming is commonly used with subtree crossover as the predominant genetic operator. In contrast, Cartesian Genetic Programming (CGP) is widely used merely with mutation for genetic variation. Compared to comprehensive and advanced knowledge about crossover in the field of tree-based Genetic Programming, the state of knowledge in CGP appears to be generally weak. Even if CGP was officially introduced over twenty years ago, the role of recombination in CGP has been recently considered to be still an open question. Promising steps have been taken in the last years, but the outcomes show that more studies are needed to evaluate crossover operators in CGP on a wider set of benchmark problems. In this work, we compare algorithms that utilize two phenotypic crossover operators for CGP, called subgraph and block crossover, to the traditional mutation-only approach. The results of our experiments on well-known symbolic regression, Boolean function, and images operator design problems demonstrate that the use of crossover-based algorithms can outperform the mutation-only CGP approach on well-known benchmark problems.

External IDs:doi:10.1007/978-3-031-46221-4_3