Reinforcement Learning Controlled Variable Speed Limits in Urban Expressway Mixed Traffic

Abstract

Increase in Rapid urbanization has intensified traffic congestion, delays, and emissions on urban expressways, revealing the limitations of conventional control strategies. This study aims to develops a Reinforcement Learning (RL) based Variable Speed Limit (VSL) control framework using a Deep QNetwork (DQN) implemented in the Simulation of Urban Mobility (SUMO) environment to enhance traffic efficiency at an expressway merging section. Traffic data used was collected using a video camera recorder and radar speed gun, with vehicle trajectories extracted through the Traffic Data Extractor developed by IIT Bombay. The model implemented was calibrated and validated using field observations from two sitesone representing uninterrupted flow and the other in an onramp merging area. The simulation compared three configurations: A baseline case without control, a conventional rule-based VSL controller, and the proposed DQN-based VSL approach and the findings reveal that the DQN agent achieved a 18.8% reduction in total travel time compared to the baseline, while the rule-based VSL controller worsened performance by 21.7%. The learning-based controller effectively mitigated congestion, reduced shockwave formation, maintained higher average speeds, and improved travel time reliability under dynamic and stochastic traffic conditions demonstrating that a reinforcement learning-driven VSL system can significantly enhance both traffic flow efficiency and user-level reliability, outperforming traditional heuristic control strategies on urban expressways. © 2026 IEEE.