Optimised Traffic Light Management through Reinforcement Learning: Traffic State Agnostic Agent vs. Holistic Agent with Current V2I Traffic State Knowledge

Traffic light control falls into two main categories: Agnostic systems that do not exploit knowledge of the current traffic state, e.g., the positions and velocities of vehicles approaching intersections, and holistic systems that exploit knowledge of the current traffic state. Emerging fifth generation (5G) wireless networks enable Vehicle-to-Infrastructure (V2I) communication to reliably and quickly collect the current traffic state. However, to the best of our knowledge, the optimized traffic light management without and with current traffic state information has not been compared in detail. This study fills this gap in the literature by designing representative Deep Reinforcement Learning (DRL) agents that learn the control of multiple traffic lights without and with current traffic state information. Our agnostic agent considers mainly the current phase of all traffic lights and the expired times since the last change. In addition, our holistic agent considers the positions and velocities of the vehicles approaching the intersections. We compare the agnostic and holistic agents for simulated traffic scenarios, including a road network from Barcelona, Spain. We find that the holistic system substantially increases average vehicle velocities and flow rates, while reducing CO2 emissions, average wait and trip times, as well as a driver stress metric.