Traffic Congestion Management Using Deep Reinforcement Learning and Decentralized Routing

Traffic congestion management and routing of vehicles are interdependent problems. The performance of the problems depends on how well these problems learn about the behaviours of one another. In this paper, we propose centralized traffic congestion management where a grid agent recommends speed limits for the road segments based on the measured traffic volume whereas the vehicles perform decentralized routing and speed selection based on recommendations. The objective of the congestion management problem is to maximize the expected sum utility of the heterogeneous vehicles which traverse the links managed by the grid agent. The problem of congestion management is formulated as a deep reinforcement learning problem which is solved using the deep deterministic policy gradient algorithm. The objective of any vehicle is to select the fastest path suitable for its travel requirement. Using real-world traffic datasets, we analyze the performance of the grid agent and the vehicles in terms of the expected total utility of the vehicles and the mean link congestion. The results demonstrate that the proposed method outperforms a greedy method and a random method.