Evolution Consistency Based Decomposition for Coonerative Coevolution

Abstract

Cooperative coevolution has been proven a promising framework for large-scale optimization. However, its performance heavily relies on the problem decomposition strategy which decomposes a high dimensional problem into exclusive smaller sub-problems. Though many decomposition methods have been developed in recent years, they are confronted with limitations in capturing the interdependency among variables and costing a large number of fitness evaluations in the decomposition stage. To alleviate these issues, this paper proposes a data-driven decomposition method, which is called affinity propagation assisted and evolution consistency based decomposition, for cooperative coevolution. Specifically, we take advantage of historical evolutionary data to mine the evolution consistency among variables. Then, based on the mined consistency, we leverage the affinity propagation clustering algorithm to adaptively separate variables into groups with each group as a sub-problem. Particularly, this decomposition method is a dynamic variable grouping strategy, which is executed periodically during the evolution. The most advantageous property of this method is that it does not cost any fitness evaluations in the decomposition stage and could self-adaptively divide variables into groups. Extensive comparison results on two widely used large-scale benchmark sets demonstrate that the proposed decomposition method could assist the cooperative coevolution algorithm to achieve competitive or even better optimization performance than state-of-the-art decomposition methods. Therefore, the proposed decomposition strategy provides a new way to decompose variables into groups.