Improving Evaluation of Recombination-based Cartesian Genetic Programming

Researchers demonstrate that recombination-based operators in Cartesian Genetic Programming can achieve competitive performance when combined with proper hyperparameter optimization, challenging the long-held assumption that mutation-only approaches are superior for symbolic regression tasks.

This research addresses a fundamental question in evolutionary computation: whether recombination-based genetic operators can enhance Cartesian Genetic Programming (CGP) performance. Historically, mutation has dominated CGP implementations while recombination approaches were largely abandoned due to perceived ineffectiveness. The study evaluates two recombination strategies—subgraph crossover and discrete phenotypic recombination—using the SRBench benchmarking platform and TinyverseGP framework, finding that systematic hyperparameter optimization unlocks performance improvements previously thought impossible.

The significance lies in challenging established orthodoxy within the genetic programming community. CGP has evolved as a specialized technique for symbolic regression, yet its design principles have remained relatively static. By revisiting dismissed approaches with modern optimization techniques, researchers reveal that prior conclusions may have resulted from inadequate parameter tuning rather than fundamental algorithmic limitations. This finding suggests that other abandoned evolutionary strategies deserve reconsideration under optimized conditions.

For practitioners and researchers in machine learning and symbolic regression, this work validates the importance of thorough hyperparameter exploration when evaluating new operators. The implications extend beyond academic interest—improved recombination methods could enhance CGP's effectiveness for real-world symbolic regression problems in physics simulation, financial modeling, and automated scientific discovery. The research demonstrates that conventional wisdom in evolutionary computation requires periodic reevaluation as optimization methodologies advance.

Future directions include expanding these findings to other problem domains, comparing computational efficiency between operators, and investigating whether other historically dismissed genetic operators might benefit from similar systematic optimization approaches.

• →Recombination-based Cartesian Genetic Programming operators achieve competitive performance when hyperparameters are properly optimized
• →Prior dismissals of recombination approaches may have resulted from inadequate parameter tuning rather than fundamental algorithmic flaws
• →Systematic hyperparameter optimization can unlock previously overlooked performance gains in evolutionary algorithms
• →The study validates the importance of reevaluating rejected techniques under modern optimization conditions
• →Findings have potential applications in symbolic regression, physics simulation, and automated scientific discovery