A Simple Yet Effective Resampling Rule in Noisy Evolutionary Optimization

Jialin Liu, Olivier Teytaud

Abstract: Noisy optimization refers to the optimization of objective functions corrupted by noise, which happens in many real-world optimization problems. Resampling has been widely used in evolutionary algorithms for noisy optimization. It has been theoretically proved that evolutionary algorithms with resampling can achieve a “log-log convergence” slope of - 21when optimizing functions corrupted by unbiased additive noise [1]. Various dynamic resampling rules have been proposed in the literature. However, determining their optimal hyperparameter values for reaching the optimal slope is hard. In this work, we reach this slope using resampling rules optimized numerically though automatic parameter tuning. We have found a parameter-free yet effective new resampling rule depending on the iteration number and the problem dimension. This simple parameter-free resampling rule is compared to several state-of-the-art rules and achieved superior performance on functions corrupted by asymmetric additive noise or in case of very high noise levels.