Poster: Can Traffic Lights and CAV Work Together using Deep Reinforcement Learning?

The optimal control of traffic lights and vehicle speed are two common ways to improve urban road traffic. Adaptive traffic signal control systems (ATSC) can adjust traffic light signal plans to maximize the intersection throughput globally, while connected autonomous vehicles (CAV) can proactively change their speed to stabilize traffic flow locally. Recent studies apply deep reinforcement learning (DRL) to achieve better control of either ATSC or CAV, respectively, thanks to the rise of big data. However, as it is difficult to train two agent types in an ever-changing environment, the joint optimization of ATSC and CAV still remains traditional transportation methods (e.g., mixed-integer linear programming). We propose a proximal policy optimization (PPO) based DRL model to simultaneously control traffic lights and CAV, relying on the vehicle to infrastructure (V2I) communications. Preliminary results under a 2×3 urban grid map show the effectiveness of our new DRL model in reducing fuel consumption, CO
2
emissions, and travel time.