Researchers have proposed DRG-MAPPO (Hierarchical Dynamic Role-Graph Multi-Agent Proximal Policy Optimization), a multi-agent reinforcement learning framework aimed at improving tactical coordination in cooperative air combat. The work addresses two limitations the authors identify in existing MARL approaches.
The first limitation is the lack of structured relational modeling, which the authors say prevents agents from capturing complex, time-varying interactions among battlefield entities. The second is that conventional flat architectures often cannot explicitly model tactical roles, leading to ambiguous task allocation in highly dynamic environments.
To tackle these issues, DRG-MAPPO builds a graph-based representation of battlefield interactions and applies graph attention mechanisms to extract relational features among allies, enemies, and threats. A high-level policy then uses a dynamic role assignment mechanism to determine tactical responsibilities such as "leader" and "supporter."
Conditioned on those roles and the encoded graph-relational features, a low-level policy executes discrete maneuver actions, allowing tactical strategy and collaborative execution to be optimized jointly. The framework also includes a target-priority auxiliary task intended to encourage behaviors such as focus-fire.
According to the reported experimental results, DRG-MAPPO achieves a state-of-the-art win rate of 87%. The authors suggest this indicates the framework balances relational modeling, interpretability, and optimization stability for cooperative air combat.