Network-wide traffic signal control based on the discovery of critical nodes and deep reinforcement learning

Abstract

To improve the traffic efficiency of city-wide road networks, we propose a traffic signal control framework that prioritizes the optimal control policies on critical nodes in road networks. In this framework, we first use a data-driven approach to discover the critical nodes. Critical nodes are identified as nodes that would cause a dramatic reduction in the traffic efficiency of the road network if they were to fail. This approach models the dynamic of road networks using a tripartite graph based on the vehicle trajectories and can accurately identify the city-wide critical nodes from a global perspective. Second, for the discovered critical nodes, we introduce a novel traffic signal control approach based on deep reinforcement learning; this approach can learn the optimal policy via constantly interacting with the road network in an iterative mode. We conduct several experiments with a transportation simulator; the results of experiments show that the proposed framework reduces the average delay and travel time compared to the baseline methods.

Keywords:

• Critical node
• deep reinforcement learning
• traffic signal controller
• tripartite graph

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China under Grant No. U1509205 and No. 51408237, and Toyota Research Center.