A Meta-Evolutionary Algorithm for Co-evolving Genotypes and Genotype / Phenotype Maps
Nathan Gaylinn, Joshua Bongard · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Evolutionary computation (EC) is often used to automatically discover solutions to optimization problems. It is valued because it allows the programmer to intuitively design a search space to fit a task, and because it is a relatively open-ended search process that favors diversity and unanticipated solutions that might be missed with gradient-based methods. Traditionally, the programmer decides on a fixed search strategy a priori, often by designing a specialized mapping from genotype to phenotype (GP map). Unfortunately, this can introduce bias and undermine the open-endedness of EC. Evolved GP maps can mitigate these concerns by automatically discovering efficient search spaces that improve evolvability. However, most research into evolved GP maps emphasizes convergence rate to a fit solution, or rate of recovery after a change in conditions. Here, we frame EC as a search over search strategies rather than a search for fit solutions. We demonstrate that a single meta-evolutionary algorithm with an evolved generative GP map can find better solutions to multiple fitness functions in the domain of 2D cellular automata than a traditional evolutionary algorithm. In the future, we hope these results will further the understanding of evolvability, its relationship to diversity, and the exploratory power of evolved GP maps.