Curriculum Learning for Adaptive Multi-UAV Swarm Target Search in Obstacle Environment

Previous studies have paid limited attention to the critical role of unmanned aerial vehicle (UAV) altitude variation in determining three-dimensional (3D) search efficiency. To address this gap, this paper introduces a deep reinforcement learning framework integrated with adaptive matrix optimization in multi-UAV target search. The proposed framework explicitply models altitude variations' impacts on observation range and detection accuracy. To balance flight safety, coverage, and detection precision, deep Q-network (DQN)-altitude matrix optimization (AMO) discretizes 3D trajectories into horizontal paths and vertical altitude decisions, effectively reducing the problem dimensionality. Furthermore, a curriculum learning approach is employed to decompose the search process into phased subtasks, each with customized decision rules and reward mechanisms. This hierarchical strategy accelerates agent learning and enhances performance in complex scenarios. Comprehensive experimental evaluations in simulated environments demonstrate that the proposed DQN-AMO outperforms benchmark methods in both robustness and generalization.