关键词: 多目标搜索, 区域侦察, 集群协同, 深度强化学习, 广域搜索

Abstract: In order to efficiently guide a drone swarm in searching multiple unknown dynamic targets within a designated area, a deep reinforcement learning-based predictive-driven collaborative search decision method (DRL-P-CSDM) is proposed. Using a grid-based approach, the method integrates environmental and historical search information to construct environmental information maps and information determinacy maps. A time-decaying window is designed to generate state variables, guiding drones to revisit regions and counteract the targets' active evasion, thereby improving search efficiency. A neural network architecture is designed, incorporating target location prediction, survival decision-making, and action judgment functionalities. This structure enables autonomous environmental prediction, eliminating the issue of poor adaptability commonly associated with manually designed models. A reward function is formulated based on reinforcement learning, where capture probability is introduced into a dense reward framework to accelerate the convergence process. A distributed architecture is employed, capable of accommodating arbitrary numbers of drones, and can still complete the task despite communication distance limitations and information update delays.Through algorithm comparison, robustness analysis, and semi-physical simulation, the effectiveness of the method is validated. Simulation results show that the DRL-P-CSDM improves target detection rate by 13.72%, reduces mission completion time by 48.65%, and increases drone survival probability by 14.86%. This method demonstrates strong comprehensiveness, robustness and versatility, capable of stable operation in multi-scale complex environments without limitations on swarm size. It holds broad engineering application potential in fields such as security monitoring, battlefield reconnaissance, forest area inspection, and post-disaster rescue operations.

Key words: multi-target search, area reconnaissance, swarm collaboration, deep reinforcement learning, wide-area search