Scholarlink

<p>Evolutionary Reinforcement Learning (ERL) combines the global exploration capability of evolutionary algorithms with gradient-based reinforcement learning optimization, showing promising results on complex tasks. However, traditional ERL methods often fail to capture the underlying environmental dynamics, leading to slow convergence in early training and suboptimal policy learning. A novel Evolutionary Reinforcement Learning framework based on State–Action Co-Representation Embedding (SAR-ERL) is proposed, in which a state–action co-representation encoder is designed to facilitate policy optimization and to strengthen the modeling of environment dynamics. The encoder models fine-grained interactions between states and actions, enabling effective representation learning from raw observations. During encoder training, a reward prediction module is designed to enhance representation quality and stability through joint optimization. The joint representations are embedded in both the value and policy networks. This enhances action-value estimation using latent dynamics and accelerates policy convergence through structured representations. Theoretical analysis further shows that, under mild assumptions, the augmented value and policy functions satisfy Lipschitz continuity, ensuring numerical stability and convergence. Experiments on MuJoCo benchmark tasks demonstrate that SAR-ERL improves average performance by 17.3% over EvoRainbow.</p>

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

SAR-ERL: an evolutionary reinforcement learning optimization method based on state–action co-representation embedding

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月29日

查看全文
  • Jinchao Li,
  • Guowei Shen,
  • Chun Guo,
  • Yi Chen,
  • Yunhe Cui

摘要

Evolutionary Reinforcement Learning (ERL) combines the global exploration capability of evolutionary algorithms with gradient-based reinforcement learning optimization, showing promising results on complex tasks. However, traditional ERL methods often fail to capture the underlying environmental dynamics, leading to slow convergence in early training and suboptimal policy learning. A novel Evolutionary Reinforcement Learning framework based on State–Action Co-Representation Embedding (SAR-ERL) is proposed, in which a state–action co-representation encoder is designed to facilitate policy optimization and to strengthen the modeling of environment dynamics. The encoder models fine-grained interactions between states and actions, enabling effective representation learning from raw observations. During encoder training, a reward prediction module is designed to enhance representation quality and stability through joint optimization. The joint representations are embedded in both the value and policy networks. This enhances action-value estimation using latent dynamics and accelerates policy convergence through structured representations. Theoretical analysis further shows that, under mild assumptions, the augmented value and policy functions satisfy Lipschitz continuity, ensuring numerical stability and convergence. Experiments on MuJoCo benchmark tasks demonstrate that SAR-ERL improves average performance by 17.3% over EvoRainbow.