Scholarlink

<p>This paper addresses the challenge of multi-target pursuit–evasion in underwater environments by proposing a hierarchical multi-agent framework that integrates geometry-informed task allocation with enhanced deep reinforcement learning. Existing MARL approaches often suffer from poor team formation, unstable learning, and inefficient target coverage in such domains. To overcome these gaps, this work introduces an Encirclement Advantage (EA) metric that quantifies distance and angular coverage around each evader, enabling interpretable and balanced task assignment via integer linear programming. At the decision-making level, MADDPG is enhanced with dual-critic stabilization, structurally informed replay, and geometry-informed reward shaping. Simulation results in both bounded and coastal map scenarios show that our method achieves capture success rates of up to 98.65%, 20–30% fewer pursuit steps than baseline MARL algorithms, and maintains stable convergence under diverse environmental conditions. These improvements demonstrate the framework's robustness, scalability, and potential for applications, such as marine surveillance, underwater archeology, and autonomous maritime logistics.</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

Encirclement advantage-guided improved multi-agent deep deterministic policy gradient for multi-target underwater pursuit–evasion

  • 期刊论文

  • 发表日期：

    2026年2月7日

查看全文
  • Haoxiang Li,
  • Huaxiang Li,
  • Yalin Xu,
  • Yue Wang,
  • Xiang Cao

摘要

This paper addresses the challenge of multi-target pursuit–evasion in underwater environments by proposing a hierarchical multi-agent framework that integrates geometry-informed task allocation with enhanced deep reinforcement learning. Existing MARL approaches often suffer from poor team formation, unstable learning, and inefficient target coverage in such domains. To overcome these gaps, this work introduces an Encirclement Advantage (EA) metric that quantifies distance and angular coverage around each evader, enabling interpretable and balanced task assignment via integer linear programming. At the decision-making level, MADDPG is enhanced with dual-critic stabilization, structurally informed replay, and geometry-informed reward shaping. Simulation results in both bounded and coastal map scenarios show that our method achieves capture success rates of up to 98.65%, 20–30% fewer pursuit steps than baseline MARL algorithms, and maintains stable convergence under diverse environmental conditions. These improvements demonstrate the framework's robustness, scalability, and potential for applications, such as marine surveillance, underwater archeology, and autonomous maritime logistics.