Dynamic two-level optimization for ride-sourcing vehicle dispatch

Abstract

While a ride-sourcing platform can dynamically dispatch a nearby driver to serve the nearest request, in the long-term, it is valuable to perform batch matching and additionally leverage spatial and temporal demand knowledge. We propose a two-level framework that incorporates future spatial demand and supply information to improve performance. The upper level predicts the aggregated demand and supply in the near future, and solves a flow maximization problem to optimize request and vehicle flows. The lower level solves an Integer Quadratic Programming problem to minimize the total cost of vehicle dispatch decisions at a shorter interval, where deviations from upper layer flow targets penalized in the objective function. This paper reports performances with varying upper-level prediction horizons and lower-level deviation penalties. Results demonstrate that during high-demand periods, the proposed two-level optimization consistently improves the service level compared to a dispatch policy of matching with the nearest vehicle.