i-Rebalance: Personalized Vehicle Repositioning for Supply Demand Balance

Haoyang Chen, Peiyan Sun, Qiyuan Song, Wanyuan Wang, Weiwei Wu, Wencan Zhang, Guanyu Gao, Yan Lyu

TL;DR

Existing ride-hailing repositioning algorithms assume drivers blindly follow centralized orders, leading to high rejection rates in the real world. i-Rebalance introduces a driver-centric deep reinforcement learning (DRL) framework that models individual driver cruising preferences and acceptance behaviors. By decoupling repositioning into a dual-agent sequential process—determining order first, then destination—it achieves a 38% boost in driver acceptance and increases total driver income by nearly 10%.

What Problem Does This Paper Address?

In on-demand ride-hailing platforms, passenger demand and vehicle supply are rarely balanced across a city. While algorithms can recommend that idle drivers relocate to high-demand areas, standard approaches treat drivers as identical, compliant units.

In reality, drivers have personal cruising habits (e.g., sticking near familiar neighborhoods or specific amenities) and can freely reject platform suggestions. When drivers reject algorithmic directives, supply-demand imbalances persist, and forced compliance risks driver churn. The central challenge is designing a personalized repositioning strategy that coordinates drivers to fix supply shortages while respecting their preferences—all without suffering from the combinatorial action space explosion typical of multi-vehicle joint optimization.

Methods

The authors structure i-Rebalance into two primary phases:

[Driver Behavior Modeling] 
 ├── LSTM Cruising Predictor  ──> Predicts driver-preferred cruising routes
 └── Logistic Regression Model ──> Predicts acceptance probability (from 99-driver study)
             │
             ▼
[Sequential Dual-Agent DRL]
 ├── Grid Agent    ──> Observes local gap & driver pool; outputs priority ranking (order)
 └── Vehicle Agent ──> Recommends 3×3 neighbor grid per driver using updated real-time gap
  1. Driver Behavior Modeling:
  2. Cruising Preference Network: A two-layer LSTM predicts each driver’s probability of visiting neighboring $3 \times 3$ grids based on historical trajectories, current driver earnings/shift time, nearby POIs, traffic, and local supply-demand metrics.
  3. Empirical Acceptance Model: Based on an on-field user study of 99 drivers, the authors fit a binomial logistic regression model to predict the probability of a driver accepting a recommendation using three key features: recommendation preference rank, expected trip income, and general driver obedience.
  4. Sequential Dual-Agent DRL (A2C Architecture):
  5. Grid Agent: Observes the projected supply-demand gap $\Delta_{t+1}$ and the preference vectors of idle drivers in a grid, outputting priority scores to define a recommendation order.
  6. Vehicle Agent: Follows this order to assign destination grids one by one. It receives an updated local supply-demand gap and rewards based on both gap reduction (Balance Reward) and driver preference satisfaction (Preference Reward).
  7. Theoretical Guarantee: The authors prove that sequentially ordering vehicle actions can achieve the same expected return as optimizing over the full joint-action space, but with a fraction of the complexity.

Datasets

  • Field User Study: Survey and discrete choice experiment responses from 99 verified Didi drivers across 11 diverse locations in Chengdu, China (filtering out 6 incomplete responses from 105 participants).
  • Trajectory & Demand Data: A real-world Chengdu taxi trajectory dataset containing over 14,800 vehicles and 10.7 million passenger trips over a month. Experiments simulated 500 active taxis and 243,679 passenger requests mapped onto an OpenStreetMap road network.

Key Insights and Ideas

  • Dispatch Order Matters: Drivers who reject recommendations fall back on their own preferences, immediately altering the local supply-demand landscape. An optimal repositioning sequence (e.g., ordering drivers by how well their personal preferences align with existing gaps) prevents later drivers from encountering already saturated or unfavorable recommendations.
  • State Beats Reward for Personalization: Ablation tests revealed that providing driver preferences directly to the agent's observation state yielded higher acceptance gains ($81.65\%$) than merely including a preference bonus in the reward function ($68.11\%$). The agent needs the context before picking an action, not just feedback afterward.
  • Income Moderates Compliance: The behavioral model confirmed that while drivers strongly resist recommendations counter to their habits, high expected fares significantly offset this resistance.

Key Results

  • Driver Acceptance: i-Rebalance achieved an 84.43% acceptance rate, drastically outperforming non-personalized DRL baselines like Proportional Repositioning (22.96%), Contextual DDQN (46.07%), and collective preference matching (66.48%).
  • Revenue and Efficiency: Total Driver Income (TDI) increased by 9.97% relative to the uncoordinated baseline, while achieving the highest Request Response Rate (78.35%) and the fewest unnecessary relocations (6,625 reposition trips).
  • Behavior Fit: The empirical driver acceptance model achieved an AUC of 0.8266 and an accuracy of 76.37%.

Limitations and Future Work

  • Local Action Scope: Repositioning recommendations are currently constrained to immediate neighboring $3 \times 3$ grid cells (10-minute travel radius), excluding longer-range cross-city rebalancing.
  • Simulator Reliance: Acceptance decisions during policy training rely on the fitted binomial regression model; real-world live A/B deployment might reveal unanticipated dynamic behaviors.
  • Future Directions: The authors suggest extending the sequential coordination framework to other cooperative multi-agent tasks and incorporating dynamic monetary incentives (subsidies) to sway driver acceptance on critical routes.

Summary updated

Abstract

Ride-hailing platforms have been facing the challenge of balancing demand and supply. Existing vehicle reposition techniques often treat drivers as homogeneous agents and relocate them deterministically, assuming compliance with the reposition. In this paper, we consider a more realistic and driver-centric scenario where drivers have unique cruising preferences and can decide whether to take the recommendation or not on their own. We propose i-Rebalance, a personalized vehicle reposition technique with deep reinforcement learning (DRL). i-Rebalance estimates drivers' decisions on accepting reposition recommendations through an on-field user study involving 99 real drivers. To optimize supply-demand balance and enhance preference satisfaction simultaneously, i-Rebalance has a sequential reposition strategy with dual DRL agents: Grid Agent to determine the reposition order of idle vehicles, and Vehicle Agent to provide personalized recommendations to each vehicle in the pre-defined order. This sequential learning strategy facilitates more effective policy training within a smaller action space compared to traditional joint-action methods. Evaluation of real-world trajectory data shows that i-Rebalance improves driver acceptance rate by 38.07% and total driver income by 9.97%.
BibTeX
@inproceedings{chen2024i,
  author = {Haoyang Chen and Peiyan Sun and Qiyuan Song and Wanyuan Wang and Weiwei Wu and Wencan Zhang and Guanyu Gao and Yan Lyu},
  title = {{i-Rebalance: Personalized Vehicle Repositioning for Supply Demand Balance}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=vraTGQBG9R},
}

Loading related content...

Figure