Data-enabled predictive control for dynamic traffic routing

Abstract

In the vision of wide adoption of autonomous vehicles (AVs), this paper proposes a dynamic traffic routing algorithm in the framework of data-driven predictive control (DeePC) for traffic congestion mitigation. Different from existing model-based approaches, DeePC does not require a parametric model to describe the traffic dynamics but directly predicts the network throughput given the route assignment based on historical observations. Compared to learning-based methods, DeePC is also more flexible in incorporating additional constraints. Yet, two challenges emerge when applying DeePC to dynamic traffic routing. First, DeePC is designed for linear time-invariant systems, whereas the behaviors of traffic networks are highly nonlinear. To address system nonlinearity, we adopt a relaxed formulation of DeePC and propose customized data collection strategies. Secondly, DeePC can hardly be centrally solved on large networks due to the great number of feasible routes and flow conservation constraints. To tackle this scalability issue, we develop a distributed optimization algorithm that decomposes the original DeePC problem into local problems and solves them in a coordinative way. Our numerical experiments on both small and medium-sized networks demonstrate the performance of DeePC in optimizing the network throughput and reducing traffic congestion. Further, DeePC archives reasonable fairness and shows a higher robustness towards the prediction error of traffic demand, which implies great potential in real practice.