Scholarlink

The success of maritime search and rescue (SAR) missions relies heavily on efficient perception and intelligent decision-making in dynamic and uncertain environments. This paper addresses the UAV-based maritime target search path planning problem by proposing a reinforcement learning framework that integrates Long Short-Term Memory (LSTM) networks with Deep Q-Networks (DQN). First, a dynamic probabilistic field model is constructed, incorporating time-evolving target position distributions (based on a Markov motion model), no-fly zone constraints, and a grid-based representation of the search area. Second, the UAV’s state (position, velocity, heading) and action space (move forward, turn left, turn right) are modeled. On this basis, a novel LSTM-DQN algorithm architecture is designed: the LSTM captures temporal dependencies in the search process to address partial observability; an attention mechanism is introduced to focus on critical local environmental features, enhancing decision efficiency and generalization capability; a representation network is built to extract high-level state features; and a reward-punishment function is designed to guide efficient search strategies, including first-visit rewards and revisit penalties. Simulation results demonstrate that the proposed algorithm outperforms traditional DQN and PPO algorithms in convergence speed, final search performance, and stability, validating its effectiveness in dynamic and uncertain maritime search scenarios.

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

A Reinforcement Learning Framework for Maritime UAV Search Based on LSTM-DQN Under Partial Observability

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zihan Xu,
  • Yihan Li,
  • Zhiwei Yang,
  • Shiqi Zhang

摘要

The success of maritime search and rescue (SAR) missions relies heavily on efficient perception and intelligent decision-making in dynamic and uncertain environments. This paper addresses the UAV-based maritime target search path planning problem by proposing a reinforcement learning framework that integrates Long Short-Term Memory (LSTM) networks with Deep Q-Networks (DQN). First, a dynamic probabilistic field model is constructed, incorporating time-evolving target position distributions (based on a Markov motion model), no-fly zone constraints, and a grid-based representation of the search area. Second, the UAV’s state (position, velocity, heading) and action space (move forward, turn left, turn right) are modeled. On this basis, a novel LSTM-DQN algorithm architecture is designed: the LSTM captures temporal dependencies in the search process to address partial observability; an attention mechanism is introduced to focus on critical local environmental features, enhancing decision efficiency and generalization capability; a representation network is built to extract high-level state features; and a reward-punishment function is designed to guide efficient search strategies, including first-visit rewards and revisit penalties. Simulation results demonstrate that the proposed algorithm outperforms traditional DQN and PPO algorithms in convergence speed, final search performance, and stability, validating its effectiveness in dynamic and uncertain maritime search scenarios.