A Traffic Light Control System Based on Reinforcement Learning and Adaptive Timing

Intelligent traffic light control is a key approach to improve the efficiency of transportation system. However, existing intelligent traffic light control methods usually only adjust phase with fixed duration or just adjust duration in a fixed phase circle. In actual scenarios with complicated and dynamic traffic flow, these methods cannot give the optimal phase and duration corresponding to the current situation because of their restricted traffic control mode, which limits the potential to further improve the efficiency of traffic transportation. For this sake, we propose a novel traffic light control system that achieves completely dynamic control. The system is able to efficiently adjust both phase and duration via deep reinforcement learning and adaptive timing. Among them, the reinforcement learning model is specially used for phase decision and the adaptive timing algorithm used for duration decision is designed for effective utilization of green time in each phase. We test our system in different traffic flows and explore the relationship between optimal duration and traffic flow. We also verify the superb performance of our traffic light control system in a whole-day traffic scene.