DRG-MAPPO: Hierarchical Dynamic Role-Graph MARL for Cooperative Air Combat
Original title:DRG-MAPPO: Hierarchical Dynamic Role-Graph Multi-Agent Reinforcement Learning for Cooperative Air Combat
Cooperative aerial engagement poses persistent coordination challenges for multi-agent reinforcement learning, where flat policy architectures often struggle to assign distinct tactical responsibilities. DRG-MAPPO addresses this by combining graph attention networks with a two-tier hierarchical policy. A high-level controller dynamically designates roles such as leader and supporter over an evolving relational graph of threats and allies, while low-level policies guide tactical maneuvers with an auxiliary focus-fire objective. In simulated trials, the framework achieved an 87% win rate, bringing structural interpretability to autonomous team combat.
Why it's worth reading
It offers a concrete framework for integrating dynamic role allocation and relational graph modeling into multi-agent systems, balancing coordination clarity with competitive performance.