A new gradient stochastic ranking-based multi-indicator algorithm for many-objective optimization

作     者：Chen, Ye Yuan, Xiaoping Cang, Xiaohui 

作者机构：China Univ Min & Technol Sch Informat & Control Engn Xuzhou 221008 Jiangsu Peoples R China Zhejiang Univ Sch Med Zhejiang Childrens Hosp Div Med Genet & Genom Hangzhou 310058 Zhejiang Peoples R China Zhejiang Univ Sch Med Inst Genet Hangzhou 310058 Zhejiang Peoples R China 

出 版 物：《SOFT COMPUTING》 (Soft Comput.)

年 卷 期：2019年第23卷第21期

页      面：10911-10929页

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China National Key Technology Support Program of China [2013BAK06B08] Scientific Research Fund of Zhejiang Provincial Education Department (China) [Y201432207] Natural Science Fund of Jiangsu Province (China) [BK20130187] 

主　　题：Many-objective evolutionary algorithm Multi-indicator Gaussian niche preservation operation Gradient stochastic ranking 

摘      要：In this paper, we propose a gradient stochastic ranking-based multi-indicator algorithm (GSRA) to guide the direction of Pareto front selection pressure. The proposed algorithm primarily aims to enhance the relationship among the different indicators in indicator-based MOEAs. In GSRA, we assigned the Gaussian niche-preservation operation to evaluate the perpendicular distance of every niche member, we also chose two indicators with different biases to balance the convergence and diversity. In environmental selection, we used a two-tier gradient stochastic ranking method to carry out the offspring selection. Seven state-of-the-art EMO algorithms are selected as the peer algorithms to validate GSRA. A series of extensive experiments is conducted on 15 test problems taken from MaF test suites, these functions in the suites are frequently used in many-objective optimization such as DTLZ and WFC problems. The experimental results revealed that the GSRA can achieve both the desired convergence and distribution properties. The solution set obtained by GrEA can achieve a better coverage of the Pareto front than that obtained by other algorithms on most of the tested problems.