Graph Attention Network-Based Multi-Agent Reinforcement Learning for Traffic Signal Priority Control

Reinforcement learning (RL) is widely applied to the problem of adaptive traffic signal control (ATSC). This paper presents an efficient RL algorithm to address ATSC issues. We adopt the decentralized multi-agent advantage actor-critic (A2C) algorithm, where distributed control methods introduce new challenges: the environment becomes partially observable for each local agent. To tackle this issue, we incorporate spatial infor-mation from neighboring agents using graph attention networks (GAT) to learn collaborative control strategies. Additionally, to better adapt to real-world road environments, we propose a hybrid reward function model. This model ensures the traffic efficiency for social vehicles while emphasizing the priority of emergency vehicles (EMV), thereby reducing their travel time. We numerically evaluate the proposed method in a simulated environment containing 25 intersections. Experimental results demonstrate that the approach presented in this paper has significant advantages compared to similar algorithms applied to multi-agent ATSC.

I. Introduction

In recent decades, with the continuous acceleration of urban-ization and the sharp increase in the number of vehicles, traffic congestion has become an increasingly prominent issue and a common challenge faced by cities worldwide. Controlling traf-fic signals is an effective method to improve road efficiency, al-leviate traffic congestion, and reduce environmental pollution. Early ATSC algorithms such as SCOOT [1] and SCATS [2] relied on manually designed plans. However, in the face of the increasingly complex and dynamic traffic environment, these empirical observation-based methods are no longer favored. In recent years, the rapid advancement of artificial intelligence has significantly propelled the development of ATSC. ATSC has incorporated various interdisciplinary technologies, such as fuzzy logic algorithms [3], genetic algorithms [4], and the most widely used RL [5]. RL is a trial-and-error learning approach that models the ATSC problem as a markov decision process, selecting actions based on state information collected from the environment to maximize rewards.