On the evolutionary behavior of genetic programming with constants optimization

Evolutionary systems are characterized by two seemingly contradictory properties: robustness and evolvability. Robustness is generally defined as an organism's ability to withstand genetic perturbation while maintaining its phenotype. Evolvability, as an organism's ability to produce useful variation. In genetic programming, the relationship between the two, mediated by selection and variation-producing operators (recombination and mutation), makes it difficult to understand the behavior and evolutionary dynamics of the search process. In this paper, we show that a local gradient-based constants optimization step can improve the overall population evolvability by inducing a beneficial structure-preserving bias on selection, which in the long term helps the process maintain diversity and produce better solutions.