Hyperparameter Tuning Unlocks Recombination in Cartesian Genetic Programming

Cartesian Genetic Programming (CGP) has historically relied on mutation as its primary genetic operator, with recombination-based approaches largely avoided due to poor performance. A new study by Tran et al. challenges this assumption by applying hyperparameter optimization to two recently proposed recombination operators: subgraph crossover and discrete phenotypic recombination. Using the TinyverseGP framework and the SRBench symbolic regression benchmark, the authors systematically tuned hyperparameters for each representation, revealing that optimized settings can yield significant performance improvements over mutation-only baselines. The work was accepted as a workshop paper at the GECCO 2026 conference.

The findings mark a turning point for evolutionary computation. By demonstrating that hyperparameter optimization is the key to making recombination viable in CGP, the study opens the door to richer genetic diversity and faster convergence in symbolic regression and other domains. Practitioners can now leverage these operators to solve complex problems where mutation-only approaches stagnate, potentially reducing time-to-solution and improving model accuracy. This research underscores the importance of careful hyperparameter tuning in genetic programming and suggests that other traditionally avoided operators may deserve a second look.