Multi-Agent Deep Reinforcement Learning for Regional Traffic Signal Control: A Topology-Aware Approach

Existing deep reinforcement learning-based traffic signal control methods seldom focus on the influence of adjacent intersections and are limited in their ability to extract latent features. To address these challenges, this paper proposes an Advantage Actor-Critic (A2C) algorithm that integrates regional road network topology information. Built upon the A2C framework, the proposed method employs a Multi-Layer Perceptron (MLP) to encode observed traffic state features at intersections and leverages a Graph Convolutional Network (GCN) to extract spatial relationships between roads. Additionally, a multi-head attention mechanism is introduced to capture interactions between agents, and a novel reward mechanism incorporates traffic flow information from adjacent intersections to enhance coordinated control. Experiments conducted in the SUMO simulation environment demonstrate that, compared to baseline algorithms, the proposed method improves average travel speed by 18.9%, significantly reduces average delay time and stop frequency, and enhances overall traffic efficiency.