关键词: 多机协同空战, 多智能体强化学习, 时空信息融合, 图注意力, 交叉注意力

Abstract: Multi-agent reinforcement learning is currently one of the most promising methods for realizing multi-UAV autonomous cooperative air combat. However, existing methods are constrained by the end-to-end network architecture, facing critical issues such as poor multi-UAV coordination and difficulty in reflecting decision-making motivation in air combat. To address this, this paper proposes a multi-UAV cooperative air combat decision-making method based on spatial-temporal information fusion to improve the cooperation and interpretability of multi-aircraft air combat. First, a spatial information fusion method based on graph attention mechanism is designed to aggregate local observations of agents and form global situation assessment, solving the problem of weak dynamic adaptability of fully connected critic networks. Second, a spatial-temporal information fusion method combining cross-attention and gated recurrent unit is developed to aggregate information of enemy and friendly units and temporal information, fusing coordination features for the policy network. Finally, a spatial-temporal information fusion-based multi-UAV cooperative air combat decision-making algorithm is constructed by integrating reinforcement learning and validated in a high-fidelity air combat environment. Exper-imental results show that the proposed method exhibits strong coordination and interpretability of decision-making motivation.

Key words: multi-UAV cooperative air combat, multi-agent reinforcement learning, spatial-temporal information fusion, graph attention, cross-attention