XRouting: Explainable Vehicle Rerouting for Urban Road Congestion Avoidance using Deep Reinforcement Learning

Abstract

Rerouting vehicles for urban congestion avoidance is challenging as the decision has to be undertaken promptly with the consideration of traffic condition changes caused by other vehicles' routing plans. Existing solutions such as the on-board navigation systems (e.g., Google Maps) cannot meet these requirements which is prone to trigger the well-known routing oscillation problem. Though deep reinforcement learning (DRL) approaches are able to provide a high-quality solution and satisfy the real-time requirement, not only do they usually suffer the slow and instability issues for convergence, but the input information, like a picture for each time step, is also teeming with redundant information. In this paper, we propose XRouting model that uses policy-based DRL and the revised Gated Transformer (GTr) architecture to accelerate and stabilize the training convergence in solving dynamic routing problems. Our simulation study validates that compared with existing rerouting solutions, XRouting can achieve higher reductions in travel time, fuel consumption, CO2 emission, and the route length. More importantly, XRouting is capable of determining which features are predominant when vehicles conduct rerouting. This explainable ability of our model can further guide human drivers what features to consider when rerouting manually in real life.