Research map / Reinforcement learning / Visual and representation RL
GARLIC: GPT-Augmented Reinforcement Learning with Intelligent Control for Vehicle Dispatching
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 GARLIC, a GPT-2-augmented multi-agent RL vehicle dispatching framework that uses multiview GCN graphs for hierarchical traffic states, a GRU-based dynamic driver-behavior reward, and a custom GeoLoss, evaluated on Manhattan and Hangzhou taxi data in SUMO.
Problem
Vehicle dispatching for online ride-hailing must handle limited local observability, multi-granularity traffic dynamics, and diverse driver behaviors while reducing empty-loaded travel.
Approach
Combines multiview honeycomb-grid GCN traffic representations (micro/meso/macro with limited V2V hops), a GRU contrastively trained driver-behavior probability folded into a fare-weighted dynamic reward, and a GPT-2 decision-transformer-style policy trained with a circular-angle-aware GeoLoss.
Key results
- On Hangzhou, GARLIC achieves error 0.1582 km and empty-loaded rate 40.71%, the best of all methods listed in Table 1.
- On Manhattan, GARLIC reports error 0.3044 km (best), but empty-loaded rate 32.38% is slightly above SS-DT's 32.25%.
- Single-view graphs yield significantly higher error than multiview graph combinations in the ablation study.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Multiview graph (micro+meso+macro, diameters 2+5+10 km) vs single-view graphs (ablation) | GARLIC · Manhattan/Hangzhou ablation | scheduling error | Helped |
| Driving behavior weight alpha (higher weight on driver behavior vs. fare) (ablation) | GARLIC · Hangzhou ablation | trajectory error | Helped |
| Dynamic reward (driving behavior + fare) vs. imitation learning baseline using traffic states only (ablation) | GARLIC · IL comparison | trajectory error | Helped |
| Custom GeoLoss vs. standard losses | GARLIC · Manhattan and Hangzhou overall | Error (km) | Helped |
| Dynamic behavior-aware reward and GPT policy vs. best baseline on Hangzhou | GARLIC · Hangzhou | Error (km): 0.1582 vs 0.1843 (SS-DT) | Helped |
| Full GARLIC vs. SS-DT on empty-loaded rate | GARLIC · Manhattan | Empty-loaded rate (%): 32.38 vs 32.25 (SS-DT) | Hurt |
| Offline RL methods (TD3+BC, CQL, RLPD) vs. larger Hangzhou dataset | conventional offline RL baselines · Hangzhou | Error (km) | Hurt |
Surprising findings
- On the Manhattan dataset, GARLIC's empty-loaded rate is slightly higher than SS-DT's, which the authors attribute to weighting driver familiarity over shortest routes.
Limitations
- Hangzhou dataset is private and will only be shared through academic collaboration.
- Evaluation is entirely in the SUMO simulator rather than real deployment.
- Empty-loaded rate on Manhattan is slightly worse than SS-DT.
Experimental setup
Datasets: NYC 2018 Yellow Taxi Trip Data (Manhattan) Hangzhou taxi trajectory data (private)
Benchmarks: Manhattan (NYC taxi) simulation Hangzhou taxi simulation
Metrics: Error (km) Empty-loaded rate (%)
Baselines: MT FTPEDEL CQL TD3+BC Decision Transformer RLPD Latent offline RL SS-DT DGS A-RTRS
Builds on: Decision Transformer TD3+BC Conservative Q-Learning SS-DT
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 0.89)
- NondBREM: Nondeterministic Offline Reinforcement Learning for Large-Scale Order Dispatching (AAAI 2024 · similarity 0.87)
- Hierarchical Optimization via LLM-Guided Objective Evolution for Mobility-on-Demand Systems (NeurIPS 2025 · similarity 0.85)
- i-Rebalance: Personalized Vehicle Repositioning for Supply Demand Balance (AAAI 2024 · similarity 0.84)
- Prompt to Transfer: Sim-to-Real Transfer for Traffic Signal Control with Prompt Learning (AAAI 2024 · similarity 0.84)
- URB - Urban Routing Benchmark for RL-equipped Connected Autonomous Vehicles (NeurIPS 2025 · similarity 0.84)
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.