Abstract
Rebalancing vacant vehicles is a critical strategy in ride-hailing operations, aimed at reducing empty miles and customer wait times by matching supply with uncertain demand. This study introduces a predictive prescription framework that combines the benefits of predictive and data-driven optimization. Built upon the matching-integrated vehicle rebalancing (MIVR) model, the framework uses machine learning to handle demand uncertainty. Evaluations using New York City ride-hailing data demonstrate that predictive prescription models outperform traditional point-prediction-driven models in volatile scenarios, achieving competitive performance with robust optimization while maintaining operational efficiency.
Introduction
In modern ride-hailing systems, drivers often spend a significant portion of their time—up to 40%—cruising without passengers. Vehicle rebalancing aims to mitigate this by proactively moving empty vehicles to areas of anticipated high demand. However, the primary challenge lies in demand uncertainty, which is influenced by factors like weather, time of day, and local events.
Traditionally, operators have relied on two main paradigms:
- Point-Prediction-Driven Optimization: Forecasting exact demand and optimizing vehicle movement accordingly.
- Data-Driven Optimization: Utilizing historical averages to guide rebalancing.
Neither approach is perfect. A highly accurate demand prediction does not always yield the best operational outcome, as overestimating demand in remote areas can lead to systemic inefficiencies and increased wait times across the network.
Methodology
The researchers propose a Predictive Prescription framework integrated with the Matching-Integrated Vehicle Rebalancing (MIVR) model. Instead of predicting a single demand value, this approach uses auxiliary data (e.g., weather, time, historical trends) to assign weights to various historical demand scenarios.
The MIVR model functions within a rolling horizon framework, allowing the system to adjust decisions as new data becomes available.
Two primary machine learning algorithms were used to generate these prescriptive weights:
- k-Nearest-Neighbors (KNN): Selecting historical days that most closely match current conditions.
- Optimal Regression Trees (ORT): Partitioning data to identify the most relevant historical contexts for decision-making.
Experimental Setup
The framework was tested using a massive real-world dataset from New York City, covering 63 taxi zones in Manhattan during June 2019. The study integrated:
- Ride-hailing trip data: High-volume data from platforms like Uber and Lyft.
- Uber Movement: Travel speed data for calculating inter-zone travel times.
- Auxiliary Data: Hourly weather conditions, Points of Interest (POIs), and public transit accessibility metrics.
The models were evaluated across four distinct demand scenarios: Morning Off-Peak, Morning Peak, Mid-day, and Evening Rush.
Results and Discussion
The study revealed several critical insights:
- Operational Success vs. Prediction Accuracy: Models with the lowest Mean Squared Error (MSE) in demand prediction did not necessarily produce the shortest wait times.
- The Benefit of Underestimation: Conservative rebalancing, often a byproduct of underestimating demand, can be advantageous. It keeps vehicles in central, high-connectivity areas, providing a natural buffer against demand volatility in the periphery. This "conservativeness" prevents the system from over-committing resources to uncertain, remote demand.
- Performance in Volatile Scenarios: In high-uncertainty periods, such as the Morning Peak and Evening Rush, the KNN-5 prescriptive model consistently achieved shorter wait times than advanced LSTM-based prediction models. The prescriptive approach also proved competitive with robust optimization models while being more computationally efficient.
Conclusion
Predictive prescriptions offer a powerful alternative to the traditional "predict-then-optimize" workflow. By leveraging contextual historical data, these models provide robust rebalancing strategies that excel in volatile urban environments. Future research will focus on integrating these prescriptions with robust optimization and applying them to autonomous mobility-on-demand (AMoD) systems where the platform has total control over every vehicle.
Summary updated