Global optimization of multimodal deceptive functions

Abstract

Local search algorithms operating in high-dimensional and multimodal search spaces often suffer from getting trapped in a local optima, therefore requiring many restarts. Even with multiple restarts, their search efficiency critically depends on the choice of the neighborhood structure. In this paper we propose an approach in which the need for the restarts is exploited to improve the neighborhood definitions. Namely, a graph clustering based linkage detection method is used to mine the information from several runs, in order to extract variable dependencies and update the neighborhood structure, variation operators accordingly. We show that the adaptive neighborhood structure approach enables the efficient solving of challenging global optimization problems that are both deceptive and multimodal.