Neighborhood evolutionary sampling with dynamic repulsion for expensive multimodal optimization

Abstract

Many real-world applications can be categorized as expensive multimodal optimization problems which not only have multiple global optima, but also their objective functions are time-consuming. To address them, the article proposed a neighborhood evolutionary sampling framework with dynamic repulsion, termed DR-NESO, which consists of two main parts. On one hand, DR-NESO uses neighborhood evolutionary sampling strategies to generate candidate solutions for real fitness evaluation, balance exploration and exploitation, and accelerate population rapidly converge to multiple global optima. On the other hand, when potential global optimal solutions are identified and archived during optimization, DR-NESO dynamically sets taboo areas based on them. Individuals located in taboo areas are moved to other promising areas by a surrogate-assisted elite restart strategy. Based on this framework, expensive multimodal optimization problems can be solved well in a limited real fitness evaluation budget. We compared the approach with state-of-the-art expensive multimodal and multimodal algorithms on the CEC2013 benchmark. The result shows DR-NESO has the best performance.