A memetic evolution system with statistical variable classification for large-scale many-objective optimization

Abstract

Although the optimization algorithms have been widely studied, the large-scale many-objective optimization problems (LSMaOPs) remain challenging. Due to the existence of a large number of decision variables, it is necessary to carry out decision variable analysis. However, it is often difficult to discriminate the diversity-related and convergence-related variables when the problem has complex characteristics. Meanwhile, as the number of decision variables and the number of objectives increase, many algorithms will suffer from the convergence challenge. To overcome these challenges, this paper proposes a Memetic Evolution System with Statistical Variable Classification (MES-SVC). A statistical variable classification method is proposed to discriminate the convergence-related and the diversity-related variables. A memetic evolution system, which includes a memetic exploitation and exploration module, and a memetic elite imitation module, is proposed to make information guidance during the evolution, thereby promote convergence. The performance of MES-SVC is compared with the state-of-the-art algorithms on 50 test instances with 3 to 10 objectives and 300 to 1000 decision variables. Experimental studies demonstrate the promising performance of the proposed MES-SVC in terms of both diversity and convergence of solutions.