基于角度惩罚距离精英选择策略的偏好高维目标优化算法  

Many-Objective Optimization Algorithm with Preference Based on the Angle Penalty Distance Elite Selection Strategy

作　　者：王丽萍[1] 章鸣雷 邱飞岳[2,3] 江波[2,3] 

机构地区：[1]浙江工业大学信息智能与决策优化研究所,杭州310023 [2]浙江工业大学教育科学与技术学院,杭州310023 [3]浙江工业大学现代教育技术研究所,杭州310023

出　　处：《计算机学报》2018年第1期236-253,共18页Chinese Journal of Computers

基　　金：国家自然科学基金(61472366;61379077;61503340);浙江省自然科学基金(LY17F020022;LQ16F030008)资助

摘　　要：基于决策者偏好的高维目标优化算法能有效集中算法资源和减小搜索空间,是处理高维目标优化问题的有效途径之一.现有研究发现,参考点位置选择对算法性能影响显著,位于极端位置的参考点容易引发算法不收敛;同时,算法多样性在种群逼近Pareto前沿的过程中反复遭到破坏.为解决以上问题,该文提出一种基于角度惩罚距离精英选择策略的偏好高维目标优化算法.该算法将决策者偏好信息融入到基于分解的多目标优化算法中,提出偏好向量生成策略,消除算法收敛性对参考点位置的敏感性;同时引入角度惩罚距离(APD)机制,分析该机制在算法搜索后期存在种群退化、收敛放缓等缺陷的基础上,提出APD精英选择策略,通过有效分配算法资源,平衡算法收敛性和多样性.算法性能对比实验中,将该文提出的算法与g-占优、r-占优、双极偏好占优以及MOEA/D-PRE在3至10维DTLZ1-4测试问题上进行性能测试.实验结果表明,该文提出的偏好算法所求解集能够有效反映决策者的偏好信息,并且在高维目标优化问题上,所提算法在偏好区域求得解集的收敛性和均匀性更优.Many-objective optimization problem is the hotspot and difficulty in the field of multi- objective optimization. With the increase of the target dimension, the weakness of traditional multi-objective evolutionary algorithm appears gradually： the proportion of non-dominated solutions in the population increases rapidly, the population size required to cover the entire Pareto frontier grows exponentially, and the balance of convergence and diversity for the algorithm is hard to maintain. The many-objective optimization algorithm based on decision maker＇s preference is promising in solving many-objective optimization problems （MaOP） because of its ability in concentrating computational resources and reducing search space during searching process. Existing work found that the performance of this kind of algorithm is sensitive to the location of reference points. Particularly, it is frequently observed that the non-convergence happened when the reference points are in the extreme region of Pareto front. Besides, the diversity of population is repeatedly destroyed when approximating the Pareto front. To solve these problems, a many- objective optimization algorithm with preference based on the angle penalty distance elite selection strategy is proposed in this work. Firstly, the proposed algorithm incorporates the preference information into the decomposition based multi-objective optimization algorithm, producing a preference vector generation strategy. To reduce influence of the reference point position to algorithm convergence, in the proposed preference vector generation method, the weight vectors are scaled based on the radius parameter of preference region ~ given by the decision maker to produce preference vectors. Secondly, an angle punishment distance （APD） based elite selection strategy is introduced to allocate computational resource more adaptively which aiming at balancing the convergence and diversity, after analyzing the defects of population degradation and slow convergence in the later stag

关 键 词：高维目标优化 偏好向量 角度惩罚距离 精英选择 进化算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]