A Spatial-Temporal Deep Reinforcement Learning Model for Large-Scale Centralized Traffic Signal Control

Deep Reinforcement Learning (DRL) has achieved great success in traffic signal control. Most DRL-based methods regard intersections as agents, which cooperate in a decentralized way. There are two main issues with the decentralized way: cooperation and stability. To overcome these issues, we propose a novel centralized control method with a global agent to control the whole network. To mitigate the curse of dimensionality problem, we use three techniques: first, a decomposition mechanism is proposed to decompose the high dimensional state-action space; second, an action-feedback technique is introduced to learn the temporal pattern from the historical decisions so as to improve the decision-making; third, a GAT model is applied to learn the spatial feature of surrounding intersection to effectively estimate the future rewards. By using the three techniques, our model can easily tackle the large-scale traffic network. We conduct extensive experiments on both synthetic and real-world data. The experiment results demonstrate that our model outperforms the traditional and state-of-the-art DRL-based control methods.