Tuning parameters across mixed dimensional instances
Tianjun Liao, Marco A. Montes de, Thomas Stützle · 2011
Sep-G-CMA-ES is a variant of G-CMA-ES with lower time complexity. In this paper, we evaluate the impact that various ways of tuning have on the performance of Sep-G-CMA-ES on scalable continuous benchmark functions. We have extracted seven parameters from Sep-G-CMA-ES and tuned them across training functions with different features using an automatic algorithm configuration tool called Iterated F-Race. The best performance of Sep-G-CMA-ES was obtained when it was tuned using functions of different dimensionality (a strategy that we call mixed dimensional). Our comparative study on scalable benchmark functions also shows that the default Sep-G-CMA-ES outperforms G-CMA-ES. Moreover, the tuned version of Sep-G-CMA-ES significantly improves over both G-CMA-ES and default Sep-G-CMA-ES.