图神经网络引导的演化算法求解约束多目标优化问题  

Graph Neural Network-Guided Evolutionary Algorithm for Solving Constrained Multi-0bjective Optimization Problems

作　　者：张毅芹 韩宗宸 孙靖 赵春亮 ZHANG Yiqin;HAN Zongchen;SUN Jing;ZHAO Chunliang(College of Data Science,Qingdao University of Science and Technology,Qingdao 266061,China;College of Information Science and Technology,Qingdao University of Seience and Technology,Qingdao 26606l,China)

机构地区：[1]青岛科技大学数据科学学院,山东青岛266061 [2]青岛科技大学信息科学技术学院,山东青岛266061

出　　处：《聊城大学学报（自然科学版）》2025年第1期135-146,共12页Journal of Liaocheng University:Natural Science Edition

基　　金：国家自然科学基金项目(62373171);山东省自然科学基金项目(ZR2023QF065);低空飞行智能服务保障山东省工程研究中心开放课题(KF2024SD006)资助

摘　　要：约束多目标优化问题由于其约束复杂性、可行域不规则性和可行解稀疏性,通常存在难以精准刻画约束关系,以及难以找到收敛性好且分布均匀的帕累托非支配解等问题。为此,本文提出了一种图神经网络引导的约束多目标演化算法,该算法包括了学习模块与权向量自适应策略,其中学习模块通过训练图神经网络对解集进行快速评估,权向量自适应策略通过判别准则和更新机制增强种群多样性。实验结果表明,该算法在多个基准测试问题上显著优于现有的五个先进算法,在复杂约束多目标优化问题上表现出色。Constrained multiobjective optimization problems are typically challenging due to the complexi ty of constraints,irregularity of the feasible region,and sparsity of feasible solutions,These factors make it difficult to precisely characterize the constraint relationships and to find Pareto nondominated solutions that are both well-converged and evenly distributed,To address these challenges,this paper proposes a graph neural network guided constrained multi-objective evolutionary algorithm.The algorithm includes a learning module and an adaptive weight vector strategy.The learning module leverages a graph neural net work trained to rapidly evaluate solution sets,while the adaptive weight vector strategy enhances popula tion diversity through a discrimination criterion and an update mechanism,Experimental results show that the proposed algorithm signilicantly outperforms five stateof-theart algorithms on various benchmark test problems and perlorms exceptionally well on complex constrained multi-objective optimization problems.

关 键 词：图神经网络 约束多目标优化问题 约束多目标演化算法 权向量更新 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] O224[自动化与计算机技术—控制科学与工程]