Enhancing multi-intersection coordination with dynamic role allocation in reinforcement learning

Coordinated control of urban intersections is crucial for alleviating traffic congestion. Traditional multi-agent reinforcement learning methods often suffer from rigid policies that struggle to adapt to dynamic traffic situations, leading to inefficient coordination. To address this, this paper proposes a dynamic role allocation reinforcement learning framework. The framework dynamically assigns cooperative roles to autonomous agents by identifying macroscopic traffic regimes, such as free-flow and downstream congestion. Each agent learns a set of heterogeneous policies tied to these roles and executes specific tasks based on its current role. This mechanism guides agents to shift from pursuing local optima to promoting global coordination. Simulation experiments in SUMO demonstrate that the proposed framework can effectively suppress congestion propagation and significantly reduce average vehicle travel time, showcasing high-performance coordinated control.