Scholarlink

While biological metaphors have long guided metaheuristic research, cognitive science frameworks offer a promising but underutilized alternative. This paper introduces the Reinforcement Learning Agent (RLA), which models global optimization as a sequential navigation task within a Markov Decision Process (MDP). By pairing a Q-learning engine with a multi-tiered memory system, RLA autonomously coordinates seven search operators, ranging from Lévy flights and chaotic maps to adaptive differential evolution. Our experiments on the CEC 2020 benchmark suite, spanning dimensions 5 through 20, provide strong evidence of the algorithm’s robustness. RLA achieved the top Friedman rank of 1.585, ahead of AGSK at 2.716. The agent reached zero error on five functions ( \(\text {f}_1\) , \(\text {f}_2\) , \(\text {f}_5\) , \(\text {f}_6\) , \(\text {f}_7\) ) while remaining competitive on the challenging composition problems. Statistical analysis via Wilcoxon tests shows RLA significantly outperforms L-SHADE variants for dimensions \(D \ge 10\) ( \(p<0.05\) ). These results, achieved under fixed evaluation budgets, demonstrate that RLA is a stable and scalable choice for complex optimization, reinforcing the value of learning-based architectures in modern heuristic design.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Learning to Search: A Reinforcement Learning Agent for Global Optimization

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Boukthir Haddar,
  • Mohamed Ali Elleuch,
  • Chedli Damak

摘要

While biological metaphors have long guided metaheuristic research, cognitive science frameworks offer a promising but underutilized alternative. This paper introduces the Reinforcement Learning Agent (RLA), which models global optimization as a sequential navigation task within a Markov Decision Process (MDP). By pairing a Q-learning engine with a multi-tiered memory system, RLA autonomously coordinates seven search operators, ranging from Lévy flights and chaotic maps to adaptive differential evolution. Our experiments on the CEC 2020 benchmark suite, spanning dimensions 5 through 20, provide strong evidence of the algorithm’s robustness. RLA achieved the top Friedman rank of 1.585, ahead of AGSK at 2.716. The agent reached zero error on five functions ( \(\text {f}_1\) , \(\text {f}_2\) , \(\text {f}_5\) , \(\text {f}_6\) , \(\text {f}_7\) ) while remaining competitive on the challenging composition problems. Statistical analysis via Wilcoxon tests shows RLA significantly outperforms L-SHADE variants for dimensions \(D \ge 10\) ( \(p<0.05\) ). These results, achieved under fixed evaluation budgets, demonstrate that RLA is a stable and scalable choice for complex optimization, reinforcing the value of learning-based architectures in modern heuristic design.