Single Camera-enabled Reinforcement Learning Traffic Signal Control System supporting Life-long Assessment

Abstract

Traffic signals, also known as traffic lights, are an essential tool for managing intersections. Improving traffic signal control can enhance the overall performance of transportation systems. Recently, reinforcement learning has shown promise in optimizing traffic signal control by utilizing high-resolution, real-time data from advanced traffic monitoring systems. However, creating a reinforcement learning-based control that is robust to all scenarios is challenging because the real world is a diverse, non-stationarity and open-ended environment. Therefore, a system that can continually monitor and evaluate the traffic signal control’s operation is necessary for deploying such these controls. In this paper, we proposed a novel reinforcement learning-based traffic signal control system that supports near real-time performance assessment.