基于半监督迁移学习的动态多目标进化算法  

A semi-supervised transfer learning based dynamic multi-objective evolutionary algorithm for dynamic multi-objective optimization

作　　者：刘阚蓉 李岩 谭树彬 刘圆超[1,2] 刘建昌 LIU Kan-rong;LI Yan;TAN Shu-bin;LIU Yuan-chao;LIU Jian-chang(The State Key Laboratory of Synthetical Automation for Process Industries,Northestern University,Shenyang Liaoning 110819,China;College of Information Science and Engineering,Northestern University,Shenyang Liaoning 110819,China;Bensteel Group Corporation Limited,Benxi Liaoning 117000,China)

机构地区：[1]东北大学流程工业综合自动化国家重点实验室,辽宁沈阳110819 [2]东北大学信息科学与工程学院,辽宁沈阳110819 [3]本钢板材股份有限公司,辽宁本溪117000

出　　处：《控制理论与应用》2025年第1期1-12,共12页Control Theory & Applications

基　　金：国家自然科学基金项目(62273080)资助.

摘　　要：动态多目标优化问题中的目标函数随系统运行环境的动态变化而改变,这将导致其Pareto最优前沿发生动态变化.在大多数动态多目标优化问题中,不同环境之间存在一定相关性,也就是说动态多目标优化算法可以利用以往环境信息对动态变化的Pareto最优前沿进行实时追踪.为充分利用环境信息去实时追踪动态变化的Pareto最优前沿,本文提出一种基于半监督迁移学习的动态多目标进化算法(SSTL-DMOEA).SSTL-DMOEA包括两个核心组成部分,首先采用一种半监督知识迁移机制将历史环境有利信息迁移至当前环境,以帮助算法在当前环境生成较好的初始种群,从而可以提高算法在当前环境中的搜寻效率;其次,通过利用历史Pareto最优解集的中心点和新环境的进化信息在目标域中生成一系列样本点,这些点可以帮助算法建立更准确的预测模型.与4种先进的动态多目标优化算法相比,SSTL-DMOEA在处理动态多目标优化问题上具有一定的优越性.In dynamic multi-objective optimization problems,multiple conflicting objectives vary over time,which will lead to the change of Pareto optimal front.In most dynamic multi-objective optimization problems,there exists the correlation between different environments,in other words,the algorithm can use the information from the previous environments to track the dynamically changing Pareto optimal front timely.In order to make full use of environmental information to track the dynamically changing Pareto optimal front,a semi-supervised transfer learning based dynamic multi-objective evolutionary algorithm(SSTL-DMOEA)is proposed in this paper.SSTL-DMOEA consists of two core components.First,it introduces a semi-supervised transfer mechanism to transfer favorable information from the historical environments to the current environment.Thus,the algorithm can generate a good initial population for improving the search efficiency in the current environment.Secondly,a series of sample points are created in the target domain by using the center point of the Pareto optimal solution set from the historical environments and the evolutionary information of the new environment.These points can help the algorithm build a more accurate prediction model.Compared with the four state-of-the-art dynamic multi-objective optimization algorithms,SSTL-DMOEA is competitive in dealing with dynamic multi-objective optimization problems.

关 键 词：动态多目标优化 进化算法 知识迁移 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]