Shrinking Mechanism as an Alternative to Cumulative Step-size Adaptation for Evolution Strategy
Bo Zhang, Lin Wang, Bo Yang · 2023
In this paper, an adaptive step-size decay mechanism based on the required accuracy of the problem is proposed to replace the traditional cumulative step-size adaptation (CSA). By simply setting the accuracy requirement of the target problem, our method achieves better results than the existing CSA methods. Compared with CSA, our decay mechanism pays more attention to exploration and avoids falling into local optima. In the mid and late stages, we smooth the decay step to converge to the threshold set by the expert, thus progressively refining the search process. We validate this on the last 10 complex multimodal test problems of the COCO dataset.