A Dynamic Global Differential Grouping for Large-Scale Black-Box Optimization

Abstract

Cooperative Co-evolution (CC) framework is an important method to tackle Large Scale Black-Box Optimization (LSBO) problem. One of the main step in CC is grouping for the decision variables, which affects the optimization performance. An ideal grouping result is that the relationship of decision variables in intra-group is stronger as possible and those in inter-groups is weaker as possible. Global Differential Grouping (GDG) is an efficient grouping method based on the idea of partial derivatives of multivariate functions, and it can automatically resolve the problem by maintaining the global information among variables. However, once the grouping result by GDG is determined, it will no longer be updated and will not be automatically adjusted with the evolution of the algorithm, which may affect the optimization performance of the algorithm. Therefore, based on GDG, a Dynamic Global Differential Grouping (DGDG) strategy is proposed for grouping the decision variables in this paper, which can update the grouping results with the evolution processing. DGDG works with Particle Swarm Optimization (PSO) algorithm in this paper, which is termed as CC-DGDG-PSO. The experimental results based on the LSBO benchmark functions from CEC’2010 show that DGDG algorithm can improve the performance of GDG.