Evolving Mean-Update Selection Methods for CMA-ES

This paper details an investigation of the extent to which performance can be improved for the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) by tuning the selection of individuals used for the mean-update algorithm. A hyper-heuristic is employed to explore the space of algorithms which select individuals from the population. We show the increase in performance obtained with a tuned selection algorithm, versus the unmodified CMA-ES mean-update algorithm. Specifically, we measure performance on instances from several real-valued benchmark function classes to demonstrate generalization of the improved performance