Benchmarking IPOP-CMA-ES-TPA and IPOP-CMA-ES-MSR on the BBOB Noiseless Testbed
Asma Atamna · 2015
We benchmark IPOP-CMA-ES, a restart Covariance Matrix Adaptation Evolution Strategy with increasing population size, with two step-size adaptation mechanisms, Two-Point Step-Size Adapation (TPA) and Median Success Rule (MSR), on the BBOB noiseless testbed. We then compare IPOP-CMA-ES-TPA and IPOP-CMA-ES-MSR to IPOP-CMA-ES with the standard step-size adaptation mechanism, Cumulative Step-size Adaptation (CSA). We conduct experiments for a budget of 105 times the dimension of the search space. As expected, the algorithms perform alike on most functions. However, we observe some relevant differences, the most significant being on the attractive sector function where IPOP-CMA-TPA and IPOP-CMA-CSA outperform IPOP-CMA-MSR, and on the Rastrigin function where IPOP-CMA-MSR is the only algorithm to solve the function in all tested dimensions. We also observe that at least one of the three algorithms is comparable to the best BBOB-09 artificial algorithm on 13 functions.