On the anytime behavior of IPOP-CMA-ES

Abstract

Anytime algorithms aim to produce a high-quality solution for any termination criterion. A recent proposal is to improve automatically the anytime behavior of single-objective optimization algorithms by incorporating the hypervolume, a well-known quality measure in multi-objective optimization, into an automatic configuration tool. In this paper, we show that the anytime behavior of IPOP-CMA-ES can be significantly improved with respect to its default parameters by applying this method. We also show that tuning IPOP-CMA-ES with respect to the final quality obtained after a large termination criterion leads to better results at that particular termination criterion, but worsens the performance of IPOP-CMA-ES when stopped earlier. The main conclusion is that IPOP-CMA-ES should be tuned with respect to the anytime behavior if the exact termination criterion is not known in advance. © 2012 Springer-Verlag.