向量角选择和指标删除的高维多目标进化算法

Many-Objective Evolutionary Algorithm with Vector Angle Selection and Indicator Deletion

针对进化算法求解高维多目标优化问题平衡收敛性和多样性所面临的挑战,提出了向量角选择和指标删除的高维多目标进化算法(MOEA/AS-ID)。该算法在环境选择过程中设计了一种包含两种策略的协作机制逐一删除收敛性和多样性差的解以保留精英个体参与下一代的进化。前者基于向量角的选择策略用于选择一对在目标空间具有相似搜索方向的解,后者基于指标的删除策略采用同时兼顾个体收敛性和分布性的I_(SDE)^(+)指标比较被选择的这一对解,然后删除具有较小指标值的解,进而促使种群朝各个方向收敛到帕累托最优前沿,最终平衡解集的收敛性和多样性。在包含各种特征的3组标准测试系列问题DTLZ、SDTLZ、MaF上,MOEA/AS-ID与近年提出的6个涵盖了当前各种类型的高维多目标进化算法执行了广泛的对比仿真实验和数值结果分析。仿真结果和数值分析表明所提算法MOEA/AS-ID求解各种特征的高维多目标优化问题平衡收敛性和多样性的能力具有较强的竞争力。

Given that the challenge for evolutionary algorithms when solving many-objective optimization problems lies in balancing the convergence and diversity,a many-objective evolutionary algorithm based on vector angle selection and indicator deletion(MOEA/AS-ID),is proposed.In this algorithm,a coordinated mechanism that includes two strategies is designed in the environmental selection process to delete the solutions with poor convergence and diversity one by one,retaining the elitist to participate in the evolution process for the next generation.To be specific,the former strategy based on vector angle selection is used to select a pair of solutions with a similar search direction in the objective space,and the latter indicator-based deletion strategy which uses the I_(SDE)^(+) indicator(indicator shift-based density estimation)that takes into account the convergence and diversity of a single solution,is employed to compare the selected pair of solutions and delete the solution with a smaller indicator value,then encourage the population to converge to the Pareto optimal front toward all directions.Finally,the balance between convergence and diversity of the solution set is achieved.On DTLZ(Deb-Thiele-Laumanns-Zitzler),SDTLZ(scaled DTLZ),and MaF(many-objective function)three benchmark test suites with various characteristics,MOEA/AS-ID and six recently proposed many-objective evolutionary algorithms covering all current types perform extensive comparative simulation experiments and numerical results analysis.Simulation results and numerical analysis show that MOEA/AS-ID has strong competitiveness in balancing the convergence and diversity when solving many-objective optimization problems with various characteristics.