Evolution strategies with adaptively rescaled mutation vectors

Rescaled mutations have been seen to have the potential to significantly improve the performance of evolution strategies in the presence of noise. However, to make use of that potential, the rescaling factor that determines the ratio of the lengths of the trial and search steps needs to be set appropriately. Good settings depend on a multitude of parameters and may vary over time. In this paper, an adaptive approach to generating rescaling factors is proposed. In experiments involving fitness-proportionate noise on several ellipsoidal test functions is it seen that robust and nearly optimal performance is achieved across a range of noise strengths.