Many-objective optimization algorithm based on adaptive reference vector

Article type: Research Article

Authors: Hu, Ziyu^{a; c} | Ma, Xuemin^{a; c} | Sun, Hao^{a; c; *} | Yang, Jingming^{a; c} | Zhao, Zhiwei^b

Affiliations: [a] Department of Electrical and Engineering, Yanshan University, Qinhuangdao, PR China | [b] Department of Computer Science and Technology, Tangshan College, Tangshan, PR China | [c] Key Laboratory of Industrial Computer Control Engineering of Hebei Province, Yanshan University, Qinhuangdao, China

Correspondence: [*] Corresponding author. Hao Sun, E-mail: [email protected].

Abstract: When dealing with multi-objective optimization, the proportion of non-dominated solutions increase rapidly with the increase of optimization objective. Pareto-dominance-based algorithms suffer the low selection pressure towards the true Pareto front. Decomposition-based algorithms may fail to solve the problems with highly irregular Pareto front. Based on the analysis of the two selection mechanism, a dynamic reference-vector-based many-objective evolutionary algorithm(RMaEA) is proposed. Adaptive-adjusted reference vector is used to improve the distribution of the algorithm in global area, and the improved non-dominated relationship is used to improve the convergence in a certain local area. Compared with four state-of-art algorithms on DTLZ benchmark with 5-, 10- and 15-objective, the proposed algorithm obtains 13 minimum mean IGD values and 8 minimum standard deviations among 15 test problem.

Keywords: Many-objective optimization, evolutionary algorithm, Gaussian mixture model, selection mechanism

DOI: 10.3233/JIFS-192124

Journal: Journal of Intelligent & Fuzzy Systems, vol. 40, no. 1, pp. 449-461, 2021