首页|基于分解和聚类的昂贵高维多目标进化算法

基于分解和聚类的昂贵高维多目标进化算法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
使用进化算法解决昂贵高维多目标优化问题时，因目标维数较高，导致收敛性和多样性平衡困难，并且消耗成本过高，使得计算资源有限时难以收敛。为此，提出一种基于分解和聚类的昂贵高维多目标进化算法(DC-EMEA)，使用克里金模型近似目标函数，减少昂贵函数的评价次数。在优化器对模型的最优解集搜索时，借助参考向量分解目标空间，有利于收敛性和多样性的平衡，同时采取两轮选择的方式，保证后代种群规模与父代相同，为填充准则选择真实评价的个体时，提供更多选择，提升搜索效率。同时，提出一种自适应填充准则，首先使用K均值算法将种群划分为k个子种群。通过划分邻域，将子种群自适应地分成不同类型，根据子种群的类型选择个体，提升计算资源的利用率。在选择个体时，侧重于对收敛性压力的维持，提升收敛速度。将选出的个体用于更新模型和档案。实验结果表明，DC-EMEA能够很好地平衡收敛性和多样性，同时具有较强的收敛能力。
Decomposition and cluster based expensive many-objective evolutionary algorithm
When using evolutionary algorithms to solve expensive many-objective optimization problems,the many-objective leads to difficulties in balancing convergence and diversity and makes convergence difficult when computational resources are limited due to high consumption costs.Therefore,this paper proposes a decomposition and cluster based expensive many-objective evolutionary algorithm(DC-EMEA),which uses the Kriging model to approximate the objective function and reduces the number of evaluations of real expensive functions.When the optimizer searches for the optimal solution set of the model,the objective space is decomposed with the help of the reference vector,which is conducive to the balance of convergence and diversity.At the same time,two rounds of selection are adopted to ensure that the offspring population size is the same as that of the parents,providing more options for the selection of individuals for real evaluation by the infill criterion and improving the search efficiency.Meanwhile,an adaptive infill criterion is proposed to firstly divide the population into k subpopulations using the K-means algorithm.Then,by dividing the neighborhood,the subpopulations are adaptively divided into different types,and individuals are selected according to the types of subpopulations to improve the utilization of computational resources.In the selection of individuals,the focus is on the maintenance of convergence pressure to improve the convergence speed.Finally,the selected individuals are used to update the model and the archive.The experiments show that the DC-EMEA can balance convergence and diversity well and has a strong convergence ability.

Kriging modelexpensive many-objective optimizationinfill criterionreference vectorevolutionary algorithmcluster algorithm

徐三水、李军华、李凌、黎明

展开 >

南昌航空大学江西省图像处理与模式识别重点实验室,南昌 330063

克里金模型 高维昂贵优化 填充准则 参考向量 进化算法 聚类算法

国家自然科学基金项目国家自然科学基金项目国家自然科学基金项目江西省研究生创新基金项目

620660316186602561866026YC2021-S678

2024

10.13195/j.kzyjc.2022.0744

控制与决策
东北大学

控制与决策

CSTPCD北大核心
影响因子：1.227
ISSN：1001-0920
年,卷(期)：2024.39(2)
  • 25