TrafficWise: Leveraging World Models for Generalized and Interpretable Traffic Control

This article proposes an optimization method for traffic signal control strategies based on world models to address the issue of robust control in urban road networks. The data-driven world model utilizes an image stream to dynamically describe traffic flows across the network. It comprises three primary components: an encoder, implicit reasoning based on encoded space, and a decoder. These elements allow the model to simulate and extrapolate the state changes of the network under various traffic control schemes, thereby uncovering the dynamic evolution patterns of complex traffic scenarios. During the implicit reasoning process of the constructed world model, the robustness and generalizability of networked control strategies are enhanced through an actor–critic architecture coupled with a multiagent cooperative strategy optimization. In the burgeoning urban landscape of Xiong’an, China, we validate the proposed cognitive traffic control method within its road network. Experimental findings reveal that the proposed approach surpasses classical traffic control methods and model-free reinforcement learning techniques. Furthermore, the method’s interpretability is enhanced through the visualization of the world model, providing decision-making insights. Such robust and explainable characteristics serve to advance the deployment of trustworthy traffic control.