Active Update of Mutation Matrix Adaptation for Variable Metric Evolution Strategy
Huilin Cheng, Zhenhua Li, Detian Yang, You Shu, Xinye Cai · 2021 17th International Conference on Computational Intelligence and Security (CIS) · 2021
Covariance matrix adaptation evolution strategy (CMA-ES) is one of the most successful evolutionary algorithms. To improve the efficiency of CMA-ES and make sufficient use of the sampled points, we propose an active update scheme for the mutation matrix which approximates the factorization of the covariance matrix. The active update effectively reduces the variance along undesired search directions with negative weights. We design a set of symmetric weights for all the sampled points as well as a larger learning rate. We conduct extensive experiments and show that the proposed active update scheme effectively accelerates the convergence to the optimum.