GlobalLight: Exploring global influence in multi-agent deep reinforcement learning for large-scale traffic signal control

Abstract

By treating each intersection as an intelligent agent, multi-agent deep reinforcement learning (MADRL) offers a promising solution to adaptive traffic signal control (ATSC) in complex urban environments. However, existing approaches often emphasize the interactions between adjacent intersections while overlooking the global influence of distant relationships. This oversight limits their scalability to small-scale traffic networks, reducing their effectiveness in real-world urban transportation systems. In this paper, we propose GlobalLight, a novel MADRL-based traffic signal control method that addresses these challenges by exploring and exploiting global influence in traffic networks. We first propose a multidimensional feature extraction module via a multi-head graph attention network, which captures the mutual influence among locally adjacent intersections. Then we design a similarity mining module with two loss functions to analyze node embeddings in the representation space, uncovering latent relationships across distant intersections in the global traffic network. Finally, GlobalLight enables similar intersections to share policy parameters for decision-making within an effective MADRL framework. Our method simultaneously considers local dependencies between adjacent intersections and global traffic flow influence, enhancing scalability and decision efficiency for ATSC in city-level larger-scale traffic systems. Experimental evaluations on both synthetic and real-world traffic networks, encompassing up to 1000 of intersections, demonstrate that our method significantly outperforms SOTA approaches across multiple performance metrics, particularly in large-scale traffic scenarios.