关键词: 区域防空反导, 多智能体强化学习, 火力分配, 可扩展决策网络, 时序推理

Abstract: To address the challenges of rapid evolution of the battlefield situation and dynamic changes in the number of incoming targets caused by the complex coupled factors in regional air defense and anti-missile operations, this study proposes a weapon-target assignment method based on the QMIX with Dynamic extension and Spatiotemporal reasoning (QMIX-DS). The proposed method considers each fire unit as an agent, and constructs a decision network that generates weapon-target assignment strategies for each agent. The core improvements of the proposed method are: A dynamically expandable feature encoding module is designed for the decision network of each agent and can adaptively process various targets. Contrastive learning is introduced to highlight target category properties, forming differentiated feature representations. The two-layer multi-head self-attention mechanism is utilized to capture the interdependencies among different categories of targets, enabling rapid reasoning of situational changes during the mission and optimizing weapon-target assignment strategies. Simulation results on different scale scenarios conducted on MOZI platform demonstrate that the proposed weapon-target assignment method can generate effective air defense and anti-missile strategies under dynamically changing battlefield conditions. Compared with standard QMIX and other mainstream algorithms, the proposed algorithm showed advantages in indicators such as target interception rate, battlefield survival rate, and missile consumption, and exhibited high scalability and generalization across different scenarios.

Key words: regional air defense and anti-missile, multi-agent reinforcement learning, weapon-target assignment, scalable decision network, spatiotemporal reasoning