This project implements a hierarchical traffic signal control framework for urban corridors, combining model-based optimization and reinforcement learning in a SUMO–Ray RLlib environment.
Architecture:
- High-Level Coordinator (HLC) – selects Max-Flow Coordination (MFC) or Green-Wave Coordination (GWC) based on traffic demand.
- Corridor Coordinator – converts the selected strategy into signal phase constraints.
- Hybrid Signal Agents (HSAs) – RL agents control intersections with action masking for feasibility.
Training:
- Low-level HSAs: MFC-aware, GWC-aware, or Pure Agent Control (PAC).
- High-level HLC: dynamically switches strategies to balance corridor-level and network-wide performance.
- Hierarchical multi-agent control
- Multiple signal coordination strategies (MFC, GWC, PAC)
- Multi-agent PPO training with RLlib
- SUMO integration (GUI or headless)
- agent_based_control/ – RL controllers (HLC, HSAs)
- coordination/ – Model-based coordination (GWC, MFC)
- high_demand/ – Heavy traffic configs
- medium_demand/ – Medium traffic configs
- low_demand/ – Light traffic configs
- high_level_sim/ – High-level sim configs
- environment.py – RLlib environment
- network.py – Interaction with SUMO
- train.py – Training script
- utils.py – Helpers (specs, mapping)
# Train a low-level GWC-aware agent
python train.py --train-policy low_level_GWC --use-gui
# Train the high-level coordinator
python train.py --train-policy high_level