Reinforcement Learning of Traffic Light Controllers Adapting to Traffic Congestion

Due to the increasing amounts of traffic in and around urban areas there is a growing need for intelligent traffic lights that optimize traffic flow. In this paper we describe the optimization of traffic light controllers using a multi-agent, model-based reinforcement learning or approximate real-time dynamic programming approach. Our methods optimize individual traffic lights locally, but the optimization takes into account traffic congestion at neighboring traffic lights, such that there is implicit cooperation between them. We show, using experiments performed with a traffic simulator, that this approach outperforms existing methods.