Adaptive forecast-driven repositioning for dynamic ride-sharing

In dynamic ride-sharing systems, intelligent repositioning of idle vehicles often improves the overall performance with respect to vehicle utilization, request rejection rates, and customer waiting times. In this work, we present a forecast-driven idle vehicle repositioning algorithm. Our approach takes a demand forecast as well as the current vehicle fleet configuration as inputs and determines suitable repositioning assignments for idle vehicles. The core part of our approach is a mixed-integer programming model that aims to maximize the acceptance rate of anticipated future trip requests while minimizing vehicle travel times for repositioning movements. To account for changes in current trip demand and vehicle supply, our algorithm adapts relevant parameters over time. We embed the repositioning algorithm into a planning service for vehicle dispatching. We evaluate our forecast-driven repositioning approach through extensive simulation studies on real-world datasets from Hamburg, New York City, Manhattan, and Chengdu. The algorithm is tested assuming a perfect demand forecast and applying a naïve forecasting model. These serve as an upper and lower bound on state-of-the-art forecasting methods. As a benchmark algorithm, we utilize a reactive repositioning scheme. Compared to this, our forecast-driven approach reduces trip request rejection rates by an average of 3.5 percentage points and improves customer waiting and ride times.