Dynamic Traffic Signal Control with Reinforcement Learning: Improving Throughput and Reducing Wait Times in Urban Networks

This paper presents a novel approach to optimizing urban traffic flow using Reinforcement Learning (RL). Traditional traffic control systems, such as fixed-timing signals, often fail to adapt to real-time traffic conditions, leading to congestion, delays, and increased fuel consumption. By leveraging real-time data from the New York City Department of Transportation (NYC DOT) and modeling traffic signal control as an RL problem, we hypothesized that RL could outperform fixed-timing systems. Using the Simulation of Urban Mobility (SUMO) platform, we simulated traffic flows from a reconstructed real series of intersections using NYCDOT data and compared the RL system to a rule-based, fixed-timing system. The RL system reduced average queue lengths by 25%, decreased vehicle wait times by 18%, improved traffic throughput by 12%, and reduced fuel consumption by 15%. These results suggest that RL offers a scalable and adaptive solution for managing urban traffic more efficiently, differing from previous studies that often utilized simulated datasets over single, simple intersections. This study is simulation-based and does not include comparisons with advanced adaptive traffic systems such as SCOOT or SCATS. However, it provides a framework that transportation agencies could integrate to improve traffic flow and reduce operational inefficiencies, leading to shorter travel times, lower fuel consumption, and fewer emissions. Your Paper will be scheduled for publication.