基于决策变量交互识别的多目标优化算法  

Multi-objective optimization algorithm based on interactive identification of decision variables

作　　者：王丽萍 林豪[2,3] 潘笑天 俞维 WANG Liping;LIN Hao;PAN Xiaotian;YU Wei(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China;Institute for Information Intelligence and Decision Optimization,Zhejiang University of Technology,Hangzhou 310023,China;School of Management,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,浙江杭州310023 [2]浙江工业大学信息智能与决策优化研究所,浙江杭州310023 [3]浙江工业大学管理学院,浙江杭州310023

出　　处：《浙江工业大学学报》2021年第4期355-367,共13页Journal of Zhejiang University of Technology

基　　金：国家自然科学基金资助项目(61472366,61379077);浙江省自然科学基金资助项目(LQ20F020014,LY17F020022);浙江省科技发展计划重点项目(2018C01080)。

摘　　要：当前求解大规模优化问题的关键在于决策变量的有效分组。随着决策变量规模的增大,决策变量间以及决策变量与目标函数间的关系愈加复杂。在总的适应度评价次数给定的条件下,决策变量分组所消耗的适应度评价次数越多,种群进化过程中所剩适应度评价次数越少,从而影响算法收敛能力,导致解集质量下降。为解决以上问题,首先,提出了一种决策变量交互识别策略,该策略能够以较少的适应度评价次数,识别决策变量的潜在交互结构并形成子组件,实现每个子组件间关联性最小;其次,在决策空间中根据个体间角度来划分每个子组件的邻域范围;最后,结合MOEA/D算法框架,提出了MOEA/D-IRG(基于决策变量交互识别的多目标优化)算法分别独立优化各个子组件。仿真实验结果表明:在LSMOP1-4测试问题上,随着决策变量规模的增加,MOEA/D-IRG算法性能明显优于NSGA-II、MOEA/D和S3-CMA-ES算法,所获解集质量更高。Grouping decision variables is the key to solving large-scale multi-objective optimization problems.With the increasing in the scale of the decision variables,the relationship between decision variables and between decision variables and the objective functions is becoming more complex.Under the condition of a certain number of the total fitness evaluation,the more fitness evaluation times consumed by the grouping of decision variables,the fewer fitness evaluation times remaining in the population evolution process.It will affect the ability of the algorithm to convergence and leadto decrese in the quality of the solution set.This paper proposes a strategy for interactive identification of decision variables.This strategy can identify the potential interaction structure of decision variables and form sub-components with a small number of fitness evaluations.It can minimize the correlation between each sub-component.Then the neighborhood scope of each subcomponent is divided according to the individual angle in the decision space.Finally,combined with the MOEA/D algorithm framework,the MOEA/D-IRG algorithm is proposed to optimize each sub-component independently.The simulation results show that the performance of MOEA/D-IRG algorithm is obviously better than that of NSGA-II,MOEA/D and S3-CMA-ES in LSMOP1-4 test problem with the increase of the scale of decision variables.The solution set is of higher quality.

关 键 词：大规模变量 变量交互识别 决策变量分组 适应度评价 多目标优化 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]