偏好多目标进化算法研究综述  

Survey on Preference-Based Multi-Objective Evolutionary Algorithms

作　　者：王丽萍 丰美玲[2] 邱启仓 章鸣雷 邱飞岳 WANG Li-Ping;FENG Mei-Ling;QIU Qi-Cang;ZHANG Ming-Lei;QIU Fei-Yue(College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou 310023;Institute of Information Intelligence and Decision Optimization, Zhejiang University of Technology, Hangzhou 310023;Zhejiang Lab, Hangzhou 311100;Zhejiang University of Technology, Hangzhou 310023)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]浙江工业大学信息智能与决策优化研究所,杭州310023 [3]之江实验室,杭州311100 [4]浙江工业大学,杭州310023

出　　处：《计算机学报》2019年第6期1289-1315,共27页Chinese Journal of Computers

基　　金：国家自然科学基金项目(61472366,61379077);浙江省自然科学基金(LY17F020022);浙江省重点研发计划项目(2018C01080)资助

摘　　要：多目标优化需要同时优化若干相互冲突的目标,其目的是获得均匀分布于整个Pareto前沿上的最优解集.然而在实际多目标优化问题中,决策者通常只对目标空间中部分区域内的Pareto最优解感兴趣,因此将决策者的偏好信息与多目标优化方法相结合成为进化计算领域的研究热点.偏好多目标进化算法通过引入决策者的偏好信息,将算法的搜索集中在决策者感兴趣的偏好区域,有效利用算法的计算资源,提高算法的求解效率,降低计算复杂度,同时有利于决策者高效地做出最终决策.本文从偏好的设置方法和算法性能两个角度介绍偏好多目标进化算法.在偏好的设置上,从占优关系、角度关系、权重向量和偏好集四个方面综述融入偏好信息的多目标进化算法;在算法性能上,从上述四类偏好的设置方法中各选取两种偏好算法进行仿真实验,从偏好策略的有效性、解集的整体性以及算法的复杂度三个方面进行实验对比并深入分析其优缺点.最后,总结了偏好多目标进化算法的未来发展趋势.Multi-objective optimization needs to optimize several conflicting objective simultaneously. The aim of the optimization algorithms is to obtain the optimal solution sets which uniformly distributed on the whole Pareto front. However, decision makers usually are only interested in the Pareto optimal solutions in some regions of the objective space in practical multi-objective optimization problems. Therefore, combining decision maker’s preference information with multi-objective optimization methods is gradually becoming a hot survey topic in the research field of evolutionary computing. Combining the decision maker’s preference information with the multi-objective evolutionary algorithms, the search process of the algorithms can be concentrated in the preference region. This can not only use the computational resources effectively and improve the efficiency of the algorithms to solve the optimization problems, but also can reduce the computational complexity and help decision makers make the final decision efficiently. In this survey, our research team systematically describe the preference multi-objective evolutionary algorithms from two aspects: the preference setting methods and the performance of the algorithms. In the setting of preferences, we summarize the multi-objective evolutionary algorithms incorporating preference information from the four aspects:dominance relationship, angle relationship, weight vector and preference set. Firstly, we summarize the preference multi-objective evolutionary algorithms based on the dominance relationship. The core idea behind it is to modify the Pareto dominance relationship, enhance the selection pressure of the algorithms and guide the algorithms to quickly converge to the preference region. Secondly, we summarize the preference multi-objective evolutionary algorithms based on angle relationship. The core idea is to use the angle relationship between the solutions to guide the evolution of the population, while using the angle to control the range of the preferen

关 键 词：多目标优化 偏好设置 占优关系 角度关系 权重向量 偏好集 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]