A Data-Driven Dynamic Stochastic Programming Framework for Ride-Sharing Rebalancing Problem under Demand Uncertainty

摘要

To make a reasonable decision on ride-sharing platform utilizing time-series historical data, we propose a deep learning-based stochastic programming framework for open driver guidance and rebalancing to further reduce the rider average wait time. The approach proactively guides open drivers to the designated regions by integrating an innovative deep neural network named LSTM-MDN with a two-stage stochastic programming model, which is capable of yielding high quality guidance solutions by leveraging the rider demand information predicted by LSTM-MDN from time-series historical data. To validate the performance of the proposed framework, we conduct a group of numerical experiments based on the New York taxi trip data sets. The results show that our proposed framework is capable of reducing driver rebalancing distance significantly, which implies that the riders' wait time can be decreased effectively. Most importantly, it turns out that by average, riders' wait time with guidance using our approach is 97% lower than the myopic batch matching algorithm without guidance.