Paper: RL-Based Cooperative Optimization of Channelization and Ramp Metering in Weaving Areas

A first-author SCI Q3 paper introducing a reinforcement learning approach to coordinate channelization design and ramp metering for urban expressway weaving areas.

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:

  1. Channelization strategies β€” two types of lane-marking configurations that guide how vehicles merge/diverge
  2. Ramp metering β€” adaptive signal control at the on-ramp to regulate inflow
  3. 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:

Key Takeaways

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