Fair and Efficient Traffic Light Control with Reinforcement Learning

Traditional traffic signal control methods primarily focus on enhancing crossroad efficiency, often overlooking fairness among drivers. Existing solutions for achieving fair traffic light control either rely on oversimplified assumptions and offline optimization or disproportionately prioritize the interests of drivers in the busy lanes. In this paper, we propose a novel reinforcement learning approach to ensuring fairness in traffic light control by minimizing waiting time disparities among drivers while optimizing traffic throughput. We present a comprehensive model of the crossroad intersection and define traffic signal phases that govern permissible driving movements. Our solution utilizes a Deep Q-Network (DQN) agent for continuous-time control and combines offline and online training. Through extensive simulations, we demonstrate the effectiveness of our fairness-driven approach using both classical metrics, such as throughput and average waiting time, and new performance metrics, including waiting time disparities and Jain’s fairness index. Evaluation results show that our approach leads to more equitable and efficient traffic signal control.