Zhang, Y.; Quinones-Grueiro, M.; Zhang, Z.; Wang, Y.; Barbour, W.; Biswas, G.; Work, D. “MARVEL: Bringing Multi-Agent Reinforcement-Learning Based Variable Speed Limit Controllers Closer to Deployment.” IEEE Access, 2024, DOI: 10.1109/ACCESS.2024.3489474.
Variable Speed Limits (VSL) are used worldwide to help manage traffic flow on highways. Most current systems use fixed rules, which can limit their effectiveness in handling different traffic situations. Recent research has explored using advanced machine learning techniques, specifically multi-agent reinforcement learning (MARL), to improve VSL systems. However, existing MARL approaches don’t meet the real-world requirements set by U.S. traffic agencies.
This study introduces a new MARL framework called MARVEL, designed to control VSL on large highway networks while meeting practical deployment needs. MARVEL only uses data from sensors that are commonly available on highways and learns to manage speed limits based on three key traffic goals to ensure it adapts well to different conditions. It shares learned strategies among multiple VSL control points, allowing it to scale across long stretches of road.
The framework was first tested in a detailed traffic simulation with 8 VSL control points over a 7-mile section. Then, it was applied to a larger 17-mile section of Interstate 24 (I-24) near Nashville, Tennessee, involving 34 control points. MARVEL showed significant improvements, increasing traffic safety by 63.4% compared to no VSL control and improving traffic flow by 58.6% compared to the current system used on I-24. The model was also tested using real-world traffic data from I-24, demonstrating its potential for real-world application.
FIGURE 1. We consider a large-scale VSL control problem with multiple gantries evenly distributed along the freeway, where the posted speed limit is identical across lanes for each gantry. Note that there is a traffic sensor collocated with each gantry to provide state input information. We order the VSL agents starting from the most downstream one, i.e., agent 1 manages the most downstream VSL gantry (controller), and agent n manages the most upstream VSL gantry (controller). Interactive Segmentation Model for Placenta Segmentation from 3D Ultrasound Images