Online Location Planning for AI-Defined Vehicles: Optimizing Joint Tasks of Order Serving and Spatio-Temporal Heterogeneous Model Fine-Tuning

B. Zheng, Bo Rao, Tianxiang Zhu, Chee Wei Tan, Jingpu Duan, Zhi Zhou, Xu Chen, Xiaoxi Zhang

TL;DR

This paper tackles the challenge of optimizing ride-hailing vehicles' dual roles: serving passengers and collecting data for fine-tuning AI foundation models in smart cities. They propose a novel multi-agent reinforcement learning framework that balances these competing tasks, considering factors like data freshness and spatial distribution, to maximize overall urban intelligence and service utility.

What Problem Does This Paper Address?

The rise of AI foundation models (FMs) and smart cities presents a new opportunity: using the vast network of ride-hailing vehicles for "vehicle crowdsensing" (VCS). These vehicles can collect diverse urban data to improve AI models. However, integrating this data collection task with the primary job of serving ride orders is complex. The core challenge lies in the inconsistent spatio-temporal characteristics of these two tasks. Ride orders have strict deadlines and their availability is unpredictable, while data for model fine-tuning degrades over time (staleness) and its utility depends on both volume and freshness. Furthermore, the geographical distribution of ride orders and data collection points (PoIs) often don't align, making it difficult for vehicles to efficiently balance both. Existing approaches often focus on one task or the other, or use simpler dispatching strategies that don't account for the nuances of FM fine-tuning.

Methods

The authors propose a GNN-enhanced Multi-Agent Reinforcement Learning (MARL) framework. Here's a breakdown of the key components:

  • MARL for Joint Optimization: Each vehicle is treated as an independent agent aiming to maximize a shared Quality of Service (QoS) metric. This metric balances the accumulated driver income (ADI) from ride orders and the accumulated data utility (ADU) from fine-tuning.
  • Graph Neural Networks (GNNs) for State Representation: To handle the complex, dynamic, and high-dimensional state space (involving vehicles, orders, PoIs, and grids), they use Relational Graph Convolutional Networks (R-GCN). This allows agents to capture global environmental information and inter-dependencies without overwhelming individual agents with raw data.
  • Low-Rank Adaptation (LoRA) for Fine-Tuning: They leverage LoRA, a parameter-efficient fine-tuning technique, to adapt foundation models. This is crucial for resource-constrained vehicles.
  • RankTuner Module: A novel component that dynamically adjusts the LoRA rank during fine-tuning. This is important because a higher rank can improve accuracy but increases fine-tuning time. The RankTuner balances this trade-off to maximize overall utility over time.
  • Customized Reward Functions: Distinct reward functions are designed for dispatching, order acceptance, and data collection, incorporating factors like order price and data utility (which considers both volume and freshness).

Datasets

  • New York City Taxi Dataset: Used to simulate ride order generation, providing realistic spatio-temporal patterns of passenger requests.
  • Foundation Model Fine-Tuning Tasks: The framework is evaluated on three representative tasks:
  • Image Classification: Using a ViT pre-trained model fine-tuned on CIFAR-100.
  • Image Segmentation: Using a SAM pre-trained model fine-tuned on satellite imagery.
  • Object Detection: Using a YOLOv7 model fine-tuned on a vehicle detection dataset. (Note: For YOLOv7, rank selection was not applicable as it used full fine-tuning).

Key Insights and Ideas

  • The "Double Duty" Vehicle: The core idea is that ride-hailing vehicles can be more than just transport; they can be active contributors to urban intelligence by collecting data for AI model improvement.
  • Spatio-Temporal Heterogeneity is Key: The paper highlights that the mismatch in how orders and data PoIs appear and decay over time and space is the central challenge. Simply optimizing for one task at a time is insufficient.
  • GNNs as a Scalability Solution: Using GNNs to create a unified, yet manageable, state representation is crucial for MARL to scale in complex urban environments.
  • Dynamic Rank Selection for Efficiency: The RankTuner module is a clever way to optimize the fine-tuning process itself, acknowledging that the "best" rank isn't static and depends on the overall objective (maximizing utility over time).
  • Data Freshness Matters (A Lot): The experiments show that data freshness significantly impacts fine-tuning accuracy, and the proposed method explicitly accounts for this, unlike simpler approaches.

Key Results

  • Superior QoS: The proposed GNN-enhanced MARL framework consistently outperformed baselines (Random, Greedy-in-OS, Greedy-in-FT, IQL, MAB) in terms of overall Quality of Service (QoS) across different PoI distributions. For instance, in Distribution 1, the proposed method achieved a QoS of 72.78, significantly higher than the next best (Greedy-in-OS at 68.93).
  • Balanced Performance: The method achieved high scores in both Accumulated Driver Income (ADI) and Accumulated Data Utility (ADU), demonstrating its ability to effectively balance the two competing tasks. It achieved the maximum income among baselines.
  • High Accuracy: The approach achieved the highest accuracy in fine-tuning tasks, attributed to its effective management of both data freshness and volume.
  • RankTuner Efficiency: The RankTuner module converged to an optimal rank faster than greedy or Thompson sampling strategies, leading to better overall QoS.
  • Robustness to Distributions: The framework showed strong performance across different distributions of PoIs and orders, indicating its adaptability.

Limitations and Future Work

  • Simplified Vehicle Movement: The paper assumes vehicles can reach their dispatch destination in one time slot. Real-world travel times are more complex and could be incorporated.
  • Order Expiration: While orders are considered invalid after 15 time slots, the exact impact of this expiration on driver behavior and strategy could be further explored.
  • Computational Overhead: While MARL and GNNs are used to manage complexity, the computational demands for training and real-time inference in large-scale urban environments might still be a consideration.
  • Future Work: The authors suggest exploring more sophisticated GNN architectures, incorporating more realistic environmental dynamics (e.g., traffic congestion), and extending the framework to handle more diverse types of urban sensing tasks and foundation models.

Summary updated

Abstract

Advances in artificial intelligence (AI) including foundation models (FMs), are increasingly transforming human society, with smart city driving the evolution of urban living. Meanwhile, vehicle crowdsensing (VCS) has emerged as a key enabler, leveraging vehicles' mobility and sensor-equipped capabilities. In particular, ride-hailing vehicles can effectively facilitate flexible data collection and contribute towards urban intelligence, despite resource limitations. Therefore, this work explores a promising scenario, where edge-assisted vehicles perform joint tasks of order serving and the emerging foundation model finetuning using various urban data. However, integrating the VCS AI task with the conventional order serving task is challenging, due to their inconsistent spatio-temporal characteristics: (i) The distributions of ride orders and data point-of-interests (PoIs) may not coincide in geography, both following a priori unknown patterns; (ii) they have distinct forms of temporal effects, i.e., prolonged waiting makes orders become instantly invalid while data with increased staleness gradually reduces its utility for model fine-tuning. To overcome these obstacles, we propose an online framework based on multi-agent reinforcement learning (MARL) with careful augmentation. A new quality-of-service (QoS) metric is designed to characterize and balance the utility of the two joint tasks, under the effects of varying data volumes and staleness. We also integrate graph neural networks (GNNs) with MARL to enhance state representations, capturing graph-structured, time-varying dependencies among vehicles and across locations. Extensive experiments on our testbed simulator, utilizing various real-world foundation model fine-tuning tasks and the New York City Taxi ride order dataset, demonstrate the advantage of our proposed method.
BibTeX
@article{zheng2025online,
  author = {B. Zheng and Bo Rao and Tianxiang Zhu and Chee Wei Tan and Jingpu Duan and Zhi Zhou and Xu Chen and Xiaoxi Zhang},
  title = {{Online Location Planning for AI-Defined Vehicles: Optimizing Joint Tasks of Order Serving and Spatio-Temporal Heterogeneous Model Fine-Tuning}},
  journal = {IEEE Transactions on Mobile Computing},
  year = {2025},
  doi = {10.1109/tmc.2025.3608179},
}
Figure