Graph Topography-Aware Reinforcement Learning for Intelligent Traffic Signal Control

Abstract

Scheduling of extensive traffic signal plays an essential role in relieving the traffic pressure. Massive-scale adaptive traffic-signal control involves dealing with combinatorial state and action spaces. Previous work attempts to address this challenge by distributing control to independent agents who only use its local state information. Considering the unstructured nature of transportation network data, we design a graph topography-aware signal control framework base on deep reinforcement learning, which explicitly utilizes the distribution of traffic intersections to extract global information. To further simulate the real-world diversified road network and its vehicle physical properties, we develop a flexible and realistic traffic intersection environment based on Unreal Engine [1]. Our graph adaptive signal controller can significantly improve traffic performance and considerably reduce traffic congestion delay compared to the traditional baseline. Additionally, our environment can lay a foundation for the subsequent optimization of transportation. Code and video are available at https://github.com/emigmo/InterSim.git.