Differential evolution with fusion of local and global search strategies

Abstract

Differential Evolution (DE) is a simple yet powerful evolutionary algorithm for solving both single-objective and many-objective optimizations. The search within DE can be either local or global. The local search (LS) can significantly improve the accuracy of the solution while the global search (GS) has the capability of jumping out from local optima. Most existing algorithms focus more on the exploration capability. This paper proposes an objective-oriented local segmenting (OLS) mechanism to improve the exploitability of algorithms. OLS mechanism uses different strategies to divide population based on different characteristics of objective space. To further balance exploitation and exploration, an effective segment strategy for the fusion of OLS and GS is proposed. In the proposed strategy, the evolution process is divided into non-overlap segments with OLS and GS alternately assigned using pre-defined generation intervals. The proposed method was incorporated into several popular DE variants and further tested on a set of 30 benchmark functions for single-objective optimization (SOP). The experimental results show that the proposed strategy significantly enhances the considered algorithms. For many-objective optimization (MOP), a many-objective optimizer, named many-objective local global DE (MOLG-DE) was constructed with the proposed mechansims and shows competitive performance against several state-of-the-art many-objective evolutionary algorithms on widely used benchmark functions.