Evolution Strategy with Covariance Matrix and decreasing step-size Adaptation (CMDSA-ES)

Abstract

The Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is state-of-the-art among evolution strategies. However, its adaptation process and parameters selection is rather complicated. The simplified version Covariance Matrix Adaptation and Step-size Self Adaptation Evolution Strategy (CMSA-ES) successfully reduces this complexity by turning to the well known step-size self-adaptation method. In the paper, we provide a new strategy to adapt the step size by analyzing the failure of fixed step-sizes and develop the CMDSA-ES, which is then compared with the CMA-ES and the CMSA-ES on different kinds of test functions and able to exhibit competitive performance.