强化学习
数学优化
计算机科学
进化算法
帕累托原理
水准点(测量)
趋同(经济学)
约束(计算机辅助设计)
集合(抽象数据类型)
进化计算
多目标优化
人工智能
动作选择
最优化问题
约束优化
代表(政治)
秩(图论)
缩小
选择(遗传算法)
约束满足
动作(物理)
功能(生物学)
约束满足问题
自适应系统
桥(图论)
作者
Kejian Pang,Zhiyuan Cai,Qiqi Xie,Yanglin Chen,Haoyu Luo,Chaoda Peng
出处
期刊:
日期:2025-12-12
卷期号:: 1-6
标识
DOI:10.1109/cis69366.2025.11433920
摘要
Balancing constraint compliance with objective optimization represents a substantial challenge in constrained multiobjective optimization problems (CMOPs). Most constrainthandling methods rely on fixed strategies that struggle to adapt to the changing preferences for feasibility and optimality at different evolutionary stages. This work introduces an adaptive epsilonconstraint method that integrates deep reinforcement learning with evolutionary algorithms. Our approach employs an agent to learn a policy that dynamically selects an action from a set of discrete epsilon-adjustment rules. This selection is based on a state representation that captures the population's convergence and feasibility status. A composite reward function is designed to quantify the improvement in the population's Pareto dominance and its proximity to the feasible region. Based on this reward signal, the agent autonomously balances the trade-off between feasibility and optimality across different evolutionary stages. Experimental results on 11 MW benchmark problems show that the proposed method achieves the best IGD values on 9 out of 11 test instances and maintains an average Friedman rank of 1.36, outperforming five competing algorithms. These results demonstrate that autonomous policy learning can effectively adapt constraint handling to problem-specific characteristics in CMOPs.
科研通智能强力驱动
Strongly Powered by AbleSci AI