Scholarlink

<p>To address the challenges of multi-unmanned surface vehicles (multi-USV) interception tasks in complex marine environments, including low interception efficiency, incomplete observations, and training convergence difficulties under dynamic disturbances, this paper proposes a Multi-Agent Proximal Policy Optimization (MAPPO) method enhanced by adaptive curriculum learning (ACL) and long short-term memory (LSTM), termed ACL-LSTM-MAPPO. First, a complex simulation environment incorporating coastlines, a harbor, obstacles, and dynamic ocean-current disturbances is constructed. Second, to mitigate decision biases caused by incomplete information under partial observability, a Long Short-Term Memory network is embedded into the Actor network, enabling USVs to utilize historical observations for decision-making and thereby enhancing temporal modeling capability and environmental adaptability. Furthermore, to address convergence difficulties caused by dynamic disturbances and sparse rewards, an adaptive curriculum learning mechanism is introduced to adjust environmental difficulty based on training performance. A dense reward design is also used to strengthen the learning signal and improve convergence efficiency and stability. Simulation results demonstrate that the proposed method effectively improves task success rates, reduces the average task completion time, and exhibits strong robustness and adaptability in complex environments.</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

Multi-USV interception decision-making method based on deep reinforcement learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月25日

查看全文
  • Bo Xiang,
  • Zheng Chen,
  • Xuanzhi Zhao,
  • Jingmin Tang,
  • Shaowei Rong,
  • Zengli Liu

摘要

To address the challenges of multi-unmanned surface vehicles (multi-USV) interception tasks in complex marine environments, including low interception efficiency, incomplete observations, and training convergence difficulties under dynamic disturbances, this paper proposes a Multi-Agent Proximal Policy Optimization (MAPPO) method enhanced by adaptive curriculum learning (ACL) and long short-term memory (LSTM), termed ACL-LSTM-MAPPO. First, a complex simulation environment incorporating coastlines, a harbor, obstacles, and dynamic ocean-current disturbances is constructed. Second, to mitigate decision biases caused by incomplete information under partial observability, a Long Short-Term Memory network is embedded into the Actor network, enabling USVs to utilize historical observations for decision-making and thereby enhancing temporal modeling capability and environmental adaptability. Furthermore, to address convergence difficulties caused by dynamic disturbances and sparse rewards, an adaptive curriculum learning mechanism is introduced to adjust environmental difficulty based on training performance. A dense reward design is also used to strengthen the learning signal and improve convergence efficiency and stability. Simulation results demonstrate that the proposed method effectively improves task success rates, reduces the average task completion time, and exhibits strong robustness and adaptability in complex environments.