Co-evolving an effective fitness sample

Abstract

We investigate two techniques for co-evolving and sampling from a population of fitness cases, and compare these with a random sampling technique. We design three symbolic regression problems on which to test these techniques, and also measure their relative performance on a modular robot control problem. The methods have varying relative performance, but in all of our experiments, at least one of the co-evolutionary methods outperforms the random sampling method by guiding evolution, with substantially fewer fitness evaluations, toward solutions that generalize best on an out-of-sample test set.