A Deep Reinforcement Learning Agent for Traffic Intersection Control Optimization

Abstract

The efficiency of traffic flows in urban areas largely depends on signal operation. The state-of-the-art traffic signal control strategies are not able to efficiently deal with varying or over-saturated conditions. To optimize the performance of existing traffic signal infrastructure, we present an end-to-end autonomous intersection control agent, based on Deep Reinforcement Learning (DRL). In the recent years, DRL has emerged as a powerful tool, solving control problems involving sequential decision making and demonstrating unprecedented success in complex settings. Our DRL traffic intersection control agent configures the traffic signal regimes based solely on live photo-realistic camera footage. We demonstrate that our agent consistently, significantly outperforms state-of-the-art fixed (pre-defined) and adaptive (induction loop-based) signal control methods under a wide range of ambient conditions, by increasing the traffic throughput and decreasing the intersection traversal time for individual vehicles.