动态多目标进化优化研究进展

Research Progress on Dynamic Multi-objective Evolutionary Optimization

动态多目标优化问题在实际生产与生活中广泛存在且问题特性随时间或环境的变化复杂多样.为有效解决该类问题,研究人员先后提出用于跟踪最优解的动态多目标进化优化方法和获得鲁棒Pareto最优解集的鲁棒动态多目标进化优化方法.前者检测到多目标优化问题特性动态改变时,重新触发寻优过程,从而快速、准确地收敛到新环境下优化问题的真实Pareto前沿.因此,环境变化检测机制和环境响应策略是核心.本文从环境变化检测方法和多样性保持、记忆策略、预测机制、迁移学习等环境响应策略入手,对已有研究进行分类归纳和总结.为有效降低解的切换代价,在有限时间内为用户提供可行且满意的较优解,已有的鲁棒动态多目标进化方法通过构建新型鲁棒多目标优化模型,寻找鲁棒Pareto最优解,不仅在当前环境下具有最优的收敛性能,还能以一定的满意程度,逼近未来多个相邻动态环境下优化问题的真实Pareto前沿.为合理评价上述方法的性能,本文列出常用的性能评价指标.最后,基于对已有方法局限性的分析,探讨了动态多目标进化优化存在的难点和挑战.

Dynamic multi-objective optimization problems widely occur in real-world production and life.Their characteristics change over time or vary with environments.To solve these problems effectively,researchers proposed two kinds of problem solvers,namely,dynamic multi-objective evolutionary optimization methods to track optimal solutions and robust dynamic multi-objective evolutionary optimization methods to search for robust Pareto-optimal solutions.In the first kind,optimization is retriggered after dynamic changes in the characteristics of multi-objective optimization problems are detected.This process aims to converge to real Pareto fronts of optimization problems in a new environment quickly and accurately.Thus,the mechanisms of detecting an environmental change and the corresponding response strategies are essential.In this paper,we classify and summarize previous studies on environmental detection methods and response strategies,including diversity preservation,memory,prediction mechanism,and transfer learning.To reduce the switching cost among solutions effectively and provide feasible and satisfactory optimal solutions for users in a limited time,we construct a new robust multi-objective optimization model through robust dynamic multi-objective evolutionary optimization and develop robust Pareto-optimal solutions.Results reveal that the proposed model has an optimal convergence performance under the current environment and approaches the true Pareto fronts of optimization problems in several subsequent future dynamic environment with the satisfied threshold.We also provide a series of commonly used indicators to evaluate two kinds of above-mentioned algorithm performence.Lastly,we emphasize the difficulties and challenges in dynamic multi-objective evolutionary optimization through an analysis of the limitations of existing methods.