DRG-MAPPO Uses Dynamic Role Graphs for Cooperative Air Combat

Summary

The paper introduces DRG-MAPPO, a hierarchical multi-agent reinforcement learning framework for cooperative air-combat decision-making. It targets two problems in existing approaches: weak modeling of time-varying relations among battlefield entities and the lack of explicit tactical roles in flat policies. The method represents allies, enemies, and threats as a graph, then uses graph attention to extract relational features. A high-level policy dynamically assigns roles such as leader and supporter, while a low-level policy uses those roles and graph features to choose discrete maneuvers. The framework also adds a target-priority auxiliary task intended to encourage coordinated behaviors such as focus-fire. In experiments, DRG-MAPPO reports an 87% win rate, described by the authors as state of the art, and is presented as balancing relational modeling, interpretability, and optimization stability.