Abstract
Ride-hailing platforms face a persistent challenge in balancing demand and supply across urban environments. Traditional vehicle repositioning techniques often treat drivers as homogeneous agents and assume full compliance with system-generated instructions. In this paper, we introduce i-Rebalance, a personalized vehicle repositioning framework that accounts for unique driver cruising preferences and autonomous decision-making. Utilizing a dual-agent Deep Reinforcement Learning (DRL) approach, i-Rebalance estimates driver acceptance through an on-field study of 99 real-world drivers. The system employs a sequential strategy: a Grid Agent determines the repositioning order of idle vehicles, while a Vehicle Agent provides personalized recommendations. Evaluation using real-world trajectory data demonstrates that i-Rebalance improves driver acceptance rates by 38.07% and increases total driver income by 9.97%.
Introduction
In the ride-hailing industry, the spatial-temporal imbalance between passenger demand and vehicle supply is a primary cause of inefficiency. While platforms like Uber, Lyft, and Didi use repositioning algorithms to move idle drivers to high-demand areas, these systems often fail because they ignore the human element. Most existing models operate under a "command-and-control" paradigm, assuming drivers will obediently follow any suggested route.
In reality, drivers are autonomous agents with distinct "cruising preferences"—some prefer familiar neighborhoods, while others prioritize high-traffic hubs. When a recommendation conflicts with these preferences, drivers often reject it, causing the system's planned balance to collapse. i-Rebalance addresses this by treating driver preference not as a secondary perk, but as a fundamental constraint for system optimization.
Methodology
i-Rebalance operates through a two-phase process involving behavioral modeling and a sequential dual-agent DRL architecture.
1. Driver Behavior Modeling
To predict how drivers will react to recommendations, the framework utilizes two distinct models:
- Cruising Preference: A lightweight Long Short-Term Memory (LSTM) network analyzes historical trajectories to predict where a driver is naturally inclined to go next.
- Acceptance Probability: Based on an on-field user study with 99 professional drivers, the authors developed a binomial logistic regression model. This model predicts the likelihood of a driver accepting a recommendation based on three factors: destination appeal, expected income, and the driver’s inherent "obedience" level.
2. Sequential Dual-Agent DRL
To avoid the computational "curse of dimensionality" associated with moving thousands of cars simultaneously, i-Rebalance splits the task between two agents:
- Grid Agent: This agent determines the order in which drivers in a specific area receive recommendations. Strategic ordering is vital; by addressing "picky" drivers early, the system can satisfy their specific preferences before the most appealing options are taken.
- Vehicle Agent: Following the sequence established by the Grid Agent, this agent provides specific, personalized destination recommendations to each driver.
This sequential approach allows the agents to operate within a smaller, more efficient action space while achieving superior coordination compared to traditional joint-action methods.
Experimental Evaluation
Datasets
The framework was validated using the Chengdu Taxi Trajectory Dataset, which includes data from approximately 14,800 taxis and over 10 million passenger trips in Chengdu, China. This was supplemented by OpenStreetMap data for realistic travel time calculations and the custom on-field user study data to model human decision-making.
Key Results
i-Rebalance was compared against several baselines, including "Demand Greedy" and "Collective Preference" models. The results highlight the efficacy of personalization:
- Acceptance Rate: i-Rebalance achieved an 84.43% acceptance rate, a significant increase over the 66.48% achieved by the next best baseline.
- Total Driver Income (TDI): The framework boosted total income by 9.97% compared to a "No Reposition" scenario.
- System Efficiency: The model maintained a 78.35% Request Response Rate, proving that it can balance individual preferences without sacrificing overall platform performance.
Conclusion and Future Work
i-Rebalance demonstrates that respecting driver autonomy is not just a matter of fairness, but a requirement for algorithmic efficiency. By predicting acceptance and personalizing recommendations, the system ensures that its repositioning strategies actually manifest in the real world.
Future research will focus on incorporating explicit monetary incentives (bonuses) to encourage drivers to take less popular but necessary routes. Additionally, the team aims to address the engineering challenges of scaling individual preference models to support hundreds of thousands of drivers in real-time environments.
Summary updated