Attention-based Spatio-Temporal Graph Convolutional Networks-enhanced deep reinforcement learning for adaptive traffic signal control in urban traffic networks.

Abstract

Urban traffic congestion remains a critical challenge due to the inflexibility of conventional fixed-time or rule-based traffic signal control systems, which cannot adapt to real-time dynamic traffic flows. Existing deep reinforcement learning (DRL)-based adaptive signal control methods often neglect complex spatio-temporal dependencies in urban road networks, and graph-based spatio-temporal models (e.g., ASTGCN) are mostly limited to traffic flow prediction rather than signal control optimization. To address these gaps, this study proposes a novel adaptive traffic signal control framework by integrating Attention-based Spatio-Temporal Graph Convolutional Networks (ASTGCN) with Multi-Agent Deep Deterministic Policy Gradient (MADDPG). The ASTGCN module effectively captures spatial topological correlations and temporal dynamic patterns of traffic flow via graph attention and temporal convolution, while the MADDPG framework realizes centralized training and decentralized execution to solve multi-intersection cooperative control under partial observability. Experiments are conducted on the SUMO simulation platform with a four-intersection road network, using both real-world and stochastic traffic flows for validation. Comparative results with fixed-time control, Max-Pressure, centralized/distributed DRL, and vanilla MADDPG baselines show that the proposed method reduces traffic congestion by 20%-25% and achieves superior performance in convergence speed, generalization ability, extreme traffic load resilience, and non-uniform traffic adaptation. Ablation studies further verify the effectiveness of the ASTGCN-based spatio-temporal feature extraction component. This framework provides a scalable and data-driven solution for intelligent traffic signal control in urban traffic networks, supporting the development of smart mobility systems.