Research map / Multi-agent systems and games / Human-AI and goal-conditioned
i-Rebalance: Personalized Vehicle Repositioning for Supply Demand Balance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes i-Rebalance, a personalized ride-hailing vehicle repositioning method using dual DRL agents (Grid Agent orders idle vehicles, Vehicle Agent recommends destinations), with driver acceptance modeled from an on-field user study of 99 drivers.
Problem
Ride-hailing supply-demand imbalance where existing repositioning methods assume drivers always comply with recommendations, ignoring individual cruising preferences and the freedom to reject them.
Approach
Sequential repositioning with two DRL agents: a Grid Agent learns the order in which idle vehicles in a grid are recommended, and a Vehicle Agent gives personalized recommendations in that order, shrinking the action space versus joint action; driver acceptance is modeled by a logistic regression fit on a user study and driver preferences by an LSTM.
Key results
- i-Rebalance reaches 109.97% normalized total driver income and 84.43% acceptance rate, versus 109.00% and 66.48% for the best baseline (By Collective Preference).
- Joint-Action variant drops to 103.15% TDI and 52.53% acceptance, showing the sequential order-learning design helps.
- Abstract reports acceptance rate improvement of 38.07% and total driver income improvement of 9.97%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use personalized sequential repositioning with learned vehicle order instead of By Collective Preference | i-Rebalance · Chengdu taxi simulator | Norm. TDI: 109.00% -> 109.97% | Helped |
| use personalized sequential repositioning with learned vehicle order instead of By Collective Preference | i-Rebalance · Chengdu taxi simulator | Acceptance Rate: 66.48% -> 84.43% | Helped |
| learn the recommendation order instead of using a fixed preference-correlation order (ablation) | i-Rebalance · Chengdu taxi simulator | Norm. TDI: 109.89% -> 109.97% | Helped |
| learn the recommendation order instead of using a random order (ablation) | i-Rebalance · Chengdu taxi simulator | Norm. TDI: 109.77% -> 109.97% | Helped |
| use sequential dual agents instead of a single joint-action agent (ablation) | i-Rebalance · Chengdu taxi simulator | Acceptance Rate: 52.53% -> 84.43% | Helped |
| add driver preference observation and preference satisfaction reward (ablation) | i-Rebalance Vehicle Agent · Chengdu taxi simulator | Acceptance Rate: 67.95% -> 84.43% | Helped |
| add driver preference observation to the Vehicle Agent state (ablation) | i-Rebalance Vehicle Agent · Chengdu taxi simulator | Acceptance Rate: 68.11% -> 84.43% | Helped |
| add preference satisfaction reward (ablation) | i-Rebalance Vehicle Agent · Chengdu taxi simulator | Acceptance Rate: 81.65% -> 84.43% | Helped |
| add preference observation and preference satisfaction reward (ablation) | i-Rebalance Vehicle Agent · Chengdu taxi simulator | Norm. TDI: 109.37% -> 109.97% | No clear effect |
Surprising findings
- Joint-action learning performed worst, even below simple heuristic baselines such as Random and Min Cost Flow in TDI.
Limitations
- Driver acceptance is estimated from a small hypothetical survey (99 drivers, 10 questions per trial) and a logistic regression, and the simulator assumes random visits to top-4 preferred grids upon rejection.
- Evaluation is entirely in a simulator built from one city's taxi data (Chengdu) with a sampled subset of 500 taxis.
Experimental setup
Datasets: Chengdu taxi trajectory dataset on-field driver user study (99 drivers) OpenStreetMap road network
Benchmarks: Chengdu taxi trajectory simulator
Metrics: Total Driver Income Reposition Income / TDI Request Response Rate Acceptance Rate Number of Repositions
Baselines: No Reposition Random Demand Greedy Reward Greedy Min Cost Flow Contextual DDQN Proportional Repositioning By Collective Preference
Builds on: advantage actor-critic Adam min cost flow repositioning contextual DDQN
Closest papers by meaning
- One Step is Enough: Multi-Agent Reinforcement Learning Based on One-Step Policy Optimization for Order Dispatch on Ride-Sharing Platforms (NeurIPS 2026 · similarity 0.87)
- GARLIC: GPT-Augmented Reinforcement Learning with Intelligent Control for Vehicle Dispatching (AAAI 2025 · similarity 0.84)
- NondBREM: Nondeterministic Offline Reinforcement Learning for Large-Scale Order Dispatching (AAAI 2024 · similarity 0.84)
- DriveGPT: Scaling Autoregressive Behavior Models for Driving (ICML 2025 · similarity 0.82)
- Advancing Multi-agent Traffic Simulation via R1-Style Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.82)
- IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.