Noise Robustness by Using Inverse Mutations

摘要

Recent advances in the theory of evolutionary algorithms have indicated that a hybrid method known as the evolutionary-gradient-search procedure yields superior performance in comparison to contemporary evolution strategies. But the theoretical analysis also indicates a noticeable performance loss in the presence of noise (i.e., noisy fitness evaluations). This paper aims at understanding the reasons for this observable performance loss. It also proposes some modifications, called inverse mutations, to make the process of estimating the gradient direction more noise robust.