DRG-MAPPO: Hierarchical Dynamic Role-Graph Multi-Agent Reinforcement Learning for Cooperative Air Combat
The paper reports an 87% win rate for a MARL system that assigns tactical roles inside a graph-modeled air-combat environment.
DRG-MAPPO represents battlefield entities as a dynamic interaction graph, then uses graph attention to capture relationships among allies, enemies, and threats. A high-level policy assigns roles such as “leader” and “supporter,” while a lower-level policy chooses maneuver actions based on those roles and graph features. The authors also add a target-priority auxiliary task intended to encourage focus-fire behavior. Source: HF Daily Papers' note.
DRG-MAPPO represents battlefield entities as a dynamic interaction graph, then uses graph attention to capture relationships among allies, enemies, and threats. A high-level policy assigns roles such as “leader” and “supporter,” while a lower-level policy chooses maneuver actions based on those roles and graph features. The authors also add a target-priority auxiliary task intended to encourage focus-fire behavior. Source: HF Daily Papers' note.
score 4