An Approximate Dynamic Programming Approach to Vehicle Dispatching and Relocation Using Time-Dependent Travel Times.

The imbalance between vehicle supply and on-demand customers has been a long-standing challenge for central ride-sourcing platforms. Current literature usually bases the design of dispatching and relocation strategies on the time-independent traffic condition (speed) assumption to reduce the problem dimension while uncertain demand and travel time subject to traffic congestion can significantly affect the optimal solutions. Therefore, we first propose a network-level traffic state estimation algorithm using functional data analysis. Then a multi-stage decision model is proposed to address the matching and repositioning of a centralized platform controlling a fleet of vehicles. Further, the customer spatial-temporal uncertainty is considered under the formulation of a stochastic programming problem. Then, an Approximate Dynamic Programming (ADP) based approach is developed for solving the multi-stage decisions efficiently. Our algorithm is evaluated in a designed simulator based on NYC yellow taxi data and the Manhattan road network. Simulation results show that the total profit can be enhanced compared with traditional time-independent traffic assumption strategies and several decision strategies.