CTRL: Cooperative Traffic Tolling via Reinforcement Learning

Abstract

People have been working long to tackle the traffic congestion problem. Among the different measures, traffic tolling has been recognized as an effective way to mitigate citywide congestion. However, traditional tolling methods can not deal with the dynamic traffic flow in cities. Meanwhile, thanks to the development of traffic sensing technology, how to set appropriate dynamic tolling according to real time traffic observations has attracted research attention in recent years. In this paper, we put the dynamic tolling problem in a reinforcement learning setting and try to tackle the three key challenges of complex state representation, pricing action credit assignment, and route price relative competition. We propose a soft actor-critic method with (1) a route-level state attention, (2) an interpretable and provable reward design, and (3) a competition-aware Q attention. Extensive experiments on real datasets have shown the superior performance of our proposed method. In addition, interesting analysis on pricing actions and vehicle routes have demonstrated why the proposed method can outperform baselines.