Enhanced Delta-tolling: Traffic Optimization via Policy Gradient Reinforcement Learning

In the micro-tolling paradigm, a centralized system manager sets different toll values for each link in a given traffic network with the objective of optimizing the system's performance. A recently proposed micro-tolling scheme, denoted Δ-tolling, was shown to yield up to 32% reduction in total travel time when compared to a no-toll scheme. Δ-tolling, computes a toll value for each link in a given network based on two global parameters: β which is a proportional parameter and R which controls the rate of toll change over time. In this paper, we propose to generalize Δ-tolling such that it would consider different R and β parameters for each link. a policy gradient reinforcement learning algorithm is used in order to tune this high-dimensional optimization problem. The results show that such a variant of Δ-tolling far surpasses the original Δ-tolling scheme, yielding up to 38% reduced system travel time compared to the original Δ-tolling scheme.