Managing city traffic with boundary flow control
Boundary flow control is potentially one of the most effective approaches for city traffic management. Under the control scheme, a city is partitioned into multiple, properly-sized neighborhoods and the transfer flows crossing neighborhood boundaries are metered. In this work, we divide boundary flow control into two stages and address them separately. In the first stage, the total metering rate of each boundary is determined. In the second, the boundary metering rate is distributed among individual road links.
The first stage is handled by a model-predictive controller embedded within a macro- scopic traffic model. The most commonly used macroscopic traffic model, which we refer to as Neighborhood Transmission Model (NTM), uses Macroscopic Fundamental Diagrams (MFDs) in combination with flow conservation laws to model the dynamics of neighborhood vehicle accumulations. We propose an enhanced NTM, which resolves multiple concerns of previous models. Unlike its predecessors, our model accounts for the constraining effects that outbound queues at the boundary can impose on a neighborhood’s circulating traffic. It also differentiates between saturated and under-saturated metering operations. With the proposed NTM, we use a model-predictive control algorithm to generate control actions at small time steps based on their forecasted impacts. Our scalable numerical method can do this for an arbitrary number of neighborhoods within a city. Computer simulations show that these enhancements improve the predictions of trip completion rates in a neighbor- hood, and of the rates at which vehicles cross metered boundaries. Optimal control actions generated as a result are similarly shown to do a better job in reducing the vehicle-hours traveled (VHT) in a city. VHT reductions stemming from the proposed model and from its predecessors differed by as much as 18%.
The second stage of control distributes boundary metering rates onto individual road links along the controlled boundary, which we refer to as the metering distribution. An effective model-free approach for determining the metering distribution is proposed in this work. To do this, we represent the street network as a directed graph and formulate the metering distribution problem as a Reinforcement Learning (RL) problem, which is solved with an actor-critic method. In extensive numerical experiments, the RL-based metering distribution policy shows consistently better performance in terms of VHT reduction than a baseline policy.