Systematic Analyses of Reinforcement Learning Controllers in Signalized Urban Corridors

A research team from the University of Bristol has published a systematic analysis demonstrating how AI can revolutionize urban traffic management. The study, "Systematic Analyses of Reinforcement Learning Controllers in Signalized Urban Corridors," trained and compared three types of AI controllers: centralized, fully decentralized, and a novel parameter-sharing decentralized model. These were benchmarked against the classical MaxPressure algorithm. The key finding is that the parameter-sharing decentralized controller performed on par with the more complex centralized AI and the established baseline, proving that efficient, scalable AI control is feasible.

The most significant breakthrough is the controller's ability to generalize. When the parameter-sharing AI, trained only on a small corridor network, was deployed on a larger, unseen network, it induced emergent coordination. Traffic spontaneously organized into 'green waves'—where platoons of vehicles encounter consecutive green lights—even though the AI controlling each junction had no explicit communication or coordination protocol with its neighbors. This suggests a path toward deploying lightweight, robust AI traffic systems that can scale across entire cities from limited training data, potentially reducing Average Travel Times (ATT) and expanding network capacity region.