The goal of this paper is to present an end-to-end, data-driven framework tocontrol Autonomous Mobility-on-Demand systems (AMoD, i.e. fleets ofself-driving vehicles). We first model the AMoD system using a time-expandednetwork, and present a formulation that computes the optimal rebalancingstrategy (i.e., preemptive repositioning) and the minimum feasible fleet sizefor a given travel demand. Then, we adapt this formulation to devise a ModelPredictive Control (MPC) algorithm that leverages short-term demand forecastsbased on historical data to compute rebalancing strategies. We test theend-to-end performance of this controller with a state-of-the-art LSTM neuralnetwork to predict customer demand and real customer data from DiDi Chuxing: weshow that this approach scales very well for large systems (indeed, thecomputational complexity of the MPC algorithm does not depend on the number ofcustomers and of vehicles in the system) and outperforms state-of-the-artrebalancing strategies by reducing the mean customer wait time by up to to89.6%.
展开▼
