文献类型：期刊文献

中文题名：基于决策变量时域变化特征分类的动态多目标进化算法

英文题名：Dynamic Multi-objective Evolutionary Algorithm Based on Classification of Decision Variable Temporal Change Characteristics

作者：闵芬[1,2];董文波[1];丁炜超[1,2]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]上海市计算机软件评测重点实验室,上海201112

年份：2024

卷号：50

期号：11

起止页码：2154

中文期刊名：自动化学报

外文期刊名：Acta Automatica Sinica

收录：CSTPCD;;EI(收录号：20244917487985);Scopus;北大核心:【北大核心2023】；CSCD:【CSCD2023_2024】；PubMed;

基金：上海市基础研究特区计划(22TQ1400100-16);上海市自然科学基金(23ZR1414900);上海市计算机软件评测重点实验室开放课题(SSTL2023_03)资助。

语种：中文

中文关键词：傅里叶变换;动态多目标优化问题;决策变量分类;动态多目标进化算法;预测策略

外文关键词：Fourier transform;dynamic multi-objective optimization problems(DMOPs);decision variable classification;dynamic multi-objective evolutionary algorithms(DMOEAs);prediction strategies

摘要：动态多目标优化问题(Dynamic multi-objective optimization problems,DMOPs)广泛存在于科学研究和工程实践中,其主要考虑在动态环境下同时联合优化多个冲突目标.现有方法往往关注于目标空间的时域特征,忽视了对单个决策变量变化特性的探索与利用,从而在处理更复杂的问题时不能有效引导种群收敛.为此,提出一种基于决策变量时域变化特征分类的动态多目标进化算法(Dynamic multi-objective evolutionary algorithm based on classification of decision variable temporal change characteristics,FT-DMOEA).所提算法在环境动态变化时,首先基于决策变量时域变化特征分类方法将当前时刻决策变量划分为线性变化和非线性变化两种类型;然后分别采用拉格朗日外插法和傅里叶预测模型对线性和非线性变化决策变量进行下一时刻的初始化操作.为了更有效地识别非线性决策变量变化模式,傅里叶预测模型通过傅里叶变换将历史种群数据从时域转换到频域,在分析周期性频率特征后,使用自回归模型进行频谱估计后再反变换至时域.在多个基准数据集上和其他算法进行对比,实验结果表明,所提算法是有效的.
Dynamic multi-objective optimization problems(DMOPs)are widely encountered in scientific research and engineering practice,where the main focus is on jointly optimizing multiple conflicting objectives in dynamic environments.Existing methods often emphasize the temporal characteristics of the objective space,neglecting the exploration and utilization of the characteristics of individual decision variable changes,thus failing to effectively guide population convergence when dealing with more complex problems.To address this issue,a dynamic multi-objective evolutionary algorithm based on classification of decision variable temporal change characteristics(FTDMOEA)is proposed.When the environment undergoes dynamic changes,the algorithm first classifies the decision variables at the current time into two types:Linear change and nonlinear change,based on the decision variable temporal change feature classification method.Subsequently,the algorithm uses Lagrange interpolation and Fourier prediction models to initialize the linear and nonlinear change decision variables for the next time step,respectively.In order to more effectively identify patterns of nonlinear decision variable changes,the Fourier prediction model transforms historical population data from the time domain to the frequency domain using Fourier transformation.After analyzing the periodic frequency features,an autoregressive model is used for spectral estimation before transforming back to the time domain.The experimental results indicate that the proposed algorithm is effective when compared with other algorithms on multiple benchmark datasets.