Reinforcement Learning for Adaptive Traffic Signal Control at Urban Intersections

Abstract

Fixed-time traffic signals cannot respond to fluctuating demand, causing long queues during peak hours. We train a deep reinforcement learning agent to choose signal phases at a four-arm intersection, using queue lengths and waiting times as the state and cumulative delay as the reward. The agent was trained and evaluated in a microscopic traffic simulation calibrated with two weeks of loop-detector counts. Compared with the existing fixed-time plan, the agent reduced average vehicle delay by 23% in the morning peak and 17% across the full day, without increasing pedestrian waiting time beyond the regulatory limit. Robustness tests with 20% higher demand still showed a 14% reduction in delay. The study outlines the steps needed before a field trial.