Improving population size adapting CMA-ES algorithm on step-size blow-up in weakly-structured multimodal functions

TL;DR

This paper improves the PSA-CMA-ES algorithm by reformulating its step-size correction to prevent blow-ups in weakly-structured multimodal functions, leading to better convergence.

Contribution

The paper introduces a reformulated step-size correction strategy for PSA-CMA-ES, addressing uncontrolled step-size blow-ups in weakly-structured multimodal problems.

Findings

Successfully mitigates step-size blow-up in experiments

Enhances convergence near the global optimum

Outperforms the original algorithm on benchmark problems

Abstract

Multimodal optimization requires both exploration and exploitation. Exploration identifies promising attraction basins, while exploitation finds the best solutions within these basins. The balance between exploration and exploitation can be maintained by adjusting parameter settings. The population size adaptation covariance matrix adaption evolutionary strategy algorithm (PSA-CMA-ES) achieves this balance by dynamically adjusting population size. PSA-CMA-ES performs well on well-structured multimodal benchmark problems. In weakly structured multimodal problems, however, the algorithm struggles to effectively manage step-size increases, resulting in uncontrolled step-size blow-ups that impede convergence near the global optimum. In this study, we reformulated the step-size correction strategy to overcome this limitation. We analytically identified the cause of the step-size blow-up and demonstrate the existence of a significance level for population size change guiding a safe passage to step-size correction. These insights were incorporated to form the reformulation. Through computer experiments on two weakly structured multimodal benchmark problems, we evaluated the performance of the new approach and compared the results with the state-of-the-art algorithm. The improved algorithm successfully mitigates step-size blow-up, enabling a better balance between exploration and exploitation near the global optimum enhancing convergence.