RL-Based Cooperative Optimization of Channelization and Ramp Metering
Authors: Diantao Deng, Bo Yu, Duo Xu, Yuren Chen, You Kong
Journal: Journal of Advanced Transportation, 2023
DOI: 10.1155/2023/4771946
Impact Factor: 2.3 | Category: SCI Q3
Motivation
Urban expressway weaving areas are notorious for congestion. When vehicles need to merge or diverge across multiple lanes in a short distance, conflicts arise β and conventional single-strategy controls (either lane markings or ramp signals, never both together) typically fail to handle them effectively.
The key insight of this paper: channelization (how lanes are physically divided) and ramp metering (how vehicles are admitted from on-ramps) are not independent problems. Optimizing them jointly β rather than in isolation β can unlock significant performance gains.
Method
The proposed framework uses a Q-learning agent to dynamically coordinate both strategies:
- Channelization strategies β two types of lane-marking configurations that guide how vehicles merge/diverge
- Ramp metering β adaptive signal control at the on-ramp to regulate inflow
- Cooperative mode β Q-learning decides the optimal combination of both in real time
The environment is built in SUMO (Simulation of Urban Mobility), with real traffic data collected via UAV aerial surveys used to calibrate and validate the simulation.
Results
The cooperative method significantly outperforms all alternatives. Lane-3 β the most heavily impacted by merge conflicts β sees a dramatic 37% improvement in average vehicle speed:
- Lane-1: +14.51% average speed increase
- Lane-2: +14.81% average speed increase
- Lane-3: +37.03% average speed increase
Key Takeaways
- Joint optimization beats isolated strategies. Traffic control is a systems problem; treating it as such pays dividends.
- Q-learning is viable for traffic signal control even without a full dynamics model β the agent learns the optimal policy purely from reward signals in simulation.
- SUMO + Python co-simulation provides a practical platform for developing and testing RL-based traffic controllers before real-world deployment.
- UAV-based data collection offers a scalable way to obtain ground-truth traffic data for simulation calibration.
Related Work
This paper draws on prior SUMO simulation research from the broader traffic engineering community, and sits alongside other RL-based signal control work in the literature. The SUMO-Python co-simulation pipeline developed here became the foundation for the Simulation Platform project referenced in my About page.
Full paper available at: https://doi.org/10.1155/2023/4771946