Explainable reinforcement learning for improved traffic signal control

Reinforcement learning has become a popular approach for traffic signal control, but its complexity often hinders practical application. To improve this, we propose a deep Q-network framework with an attention mechanism inspired by how traffic police manage intersections. This model aggregates vehicle data directly, avoiding the need for lane-based structures, and shows state-of-the-art performance across various scenarios. On a public dataset, it outperformed previous methods, achieving a 44% reduction in travel time and over 50% fewer queue lengths, compared to fixed-time control. Additionally, on a self-collected dataset, it improved the average travel time by 17.3% and queue length by 12% during peak hours. The method balances travel time, delay, and throughput while adapting to various traffic patterns, enhancing its effectiveness for traffic signal optimization. By visualizing attention weights on key vehicles, we provide transparency in decision-making, fostering trust among traffic agencies and the public.