Pith. sign in

REVIEW 3 major objections 6 minor 75 references

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

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Ride-hailing fleets can serve passengers and fine-tune models in one policy

desk verdict New joint order-serving/fine-tuning MARL scenario with a coherent formulation, but the unspecified data-utility function f_k makes the empirical claims unverifiable. read the letter →

arxiv 2502.04399 v1 pith:5S3AA32X submitted 2025-02-06 cs.LG cs.AIcs.SYeess.SY

Bokeng Zheng , Bo Rao , Tianxiang Zhu , Chee Wei Tan , Jingpu Duan , Zhi Zhou , Xu Chen , Xiaoxi Zhang This is my paper · ORCID
classification cs.LGcs.AIcs.SYeess.SY
keywords vehiclecrowdsensingfoundationmodelfine-tuningride-hailingdispatchingmulti-agentreinforcementlearninggraphneuralnetworksageofinformationLoRAsmartcity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a fleet of ride-hailing vehicles can jointly serve passenger orders and collect fresh urban data for fine-tuning foundation models, and that a decentralized multi-agent reinforcement learning framework with graph-neural-network state embeddings is the right way to decide where each vehicle should go next. It defines a single quality-of-service objective, $\text{QoS} = \alpha\,\text{ADI} + \beta\,\text{ADU}$, that weights driver income from completed orders against a data-utility term based on the volume and age of collected data. It argues that the two tasks can be balanced online even though their spatial distributions and time-sensitivity differ. If correct, the result would let ride-hailing platforms turn their fleets into moving data-collection and model-update infrastructure without giving up order revenue.

What carries the argument

The carrying object is the topology graph plus R-GCN embedding: at each time slot, vehicle nodes, order nodes, PoI nodes, grid nodes, and a shortcut node form a heterogeneous graph, and relational graph convolutional layers propagate features so each agent's state has a fixed dimension while still reflecting the whole city. The objective $\text{QoS} = \alpha\,\text{ADI} + \beta\,\text{ADU}$ is realized through matching rewards: order acceptance gives $\alpha$ times the order price, data collection gives $\beta f_k(d, \lambda)$, a task-specific utility of data volume and age of information, and dispatching gives zero immediate reward. Training uses MAPPO with generalized advantage estimation, and the RankTuner heuristic decides the LoRA rank for each fine-tuning round by hill-climbing on recent ADU.

What would settle it

Run the same MARL training in a simulator where $f_k$ is replaced by an empirically measured accuracy surface from actual LoRA fine-tuning on time-stamped urban data, and compare QoS against the same baselines; if the margin shrinks or reverses, the claimed advantage is an artifact of the simulated utility function. A cheaper check is to hold out half the fine-tuning accuracy measurements, train the agent on the other half, and see whether the learned policy predicts the held-out ADU.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that the proposed GNN-enhanced MARL algorithm achieves the best QoS, best accumulated driver income, best accumulated data utility, and best average fine-tuning accuracy compared with Random, Greedy-in-OS, Greedy-in-FT, MAB, and IQL in a simulator driven by the New York City Taxi dataset and three real fine-tuning tasks: image classification with ViT on CIFAR-100, image segmentation with SAM on satellite imagery, and object detection with YOLOv7. The learning problem is modeled as a Markov game in which each vehicle is an agent choosing among dispatching to a neighboring grid, accepting an order, or collecting data from a point of interest. R-GCN turns a dynamic topology graph of vehicles, orders, PoIs, and grids into a fixed-dimension state embedding that gives each local agent a global view without inflating the policy input. The paper also claims that its RankTuner module, which adjusts the LoRA rank based on recent ADU changes, converges to the best rank faster than greedy and Thompson sampling alternatives and improves QoS.

Load-bearing premise

The data-collection reward in equation (14) is built on an unstated function $f_k(d, \lambda)$ that maps data volume and freshness to fine-tuning utility; if that function does not match how real fine-tuning accuracy responds to volume and age, the learned policy optimizes the wrong objective and the claimed advantages may not transfer.

Editorial extensions

If this is right

  • If the central claim holds, a ride-hailing platform can dispatch idle vehicles to data-rich grids without a separate optimization stage; the same learned policy covers passenger income and data utility.
  • Because the GNN embedding fixes the input dimension, the method should scale to larger fleets and longer time horizons without enlarging each agent's policy network.
  • The RankTuner result implies that dynamically choosing LoRA rank during operation can improve aggregate utility compared with any fixed rank, since lower ranks allow more fine-tuning rounds per vehicle.
  • The framework offers a way to operationalize data freshness: instead of hand-crafting freshness-decay models, the agent learns when old data is worth collecting from the reward signal.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves $f_k(d, \lambda)$ unspecified, so a natural extension is to learn it online from the fine-tuning accuracy actually observed after each data upload; that would close the loop between the reward model and the real accuracy surface the paper uses only to motivate the experiments.
  • Because the QoS weights $\alpha$ and $\beta$ are stakeholder-set, the trained policies lie on a Pareto front of order-serving versus data utility; a platform could retune these weights and retrain or condition the policy to shift fleet behavior, a direction the paper does not explore.
  • The simulator uses one-step travel to any neighboring grid; on real street networks with traffic, travel times vary, so the claimed advantages should be re-tested with a time-expanded graph before deployment.
  • The three PoI distributions tested suggest the method is robust to mismatch between order and PoI geography; one could test an adversarial distribution where PoIs lie in sparse or hard-to-reach grids to probe the limits of that robustness.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper studies an online decision-making problem in which ride-hailing vehicles jointly serve passenger orders and collect urban data for fine-tuning foundation models. The authors formulate a weighted QoS objective (Eq. 5) over accumulated driver income (ADI) and accumulated data utility (ADU), model the problem as a Markov game, and propose a decentralized MARL framework based on MAPPO with R-GCN state embedding over a topology graph, plus a RankTuner heuristic for LoRA rank selection. The system is evaluated in a custom simulator using the New York City taxi dataset and three fine-tuning tasks (ViT/CIFAR-100, SAM/satellite imagery, YOLOv7/vehicle detection), comparing against Random, Greedy-in-OS, Greedy-in-FT, MAB, and IQL baselines. The paper claims the proposed method achieves the best QoS, ADI, ADU, and fine-tuning accuracy in these experiments.

Significance. If the reported results are substantiated, the work addresses a timely and under-explored problem: integrating vehicle crowdsensing with foundation-model fine-tuning under spatio-temporal heterogeneity. The graph-based state embedding is a reasonable approach to a high-dimensional, time-varying state space, and the idea of dynamically adjusting LoRA rank during fleet-level optimization is interesting. The central strength is the problem formulation, which captures two different utility decay structures (order expiration and data staleness). However, the evaluation is built on a reward function f_k(d, lambda) that is never specified, and the experimental comparison lacks statistical grounding. Since the claimed advantages hinge on the relationship between simulated data utility and true fine-tuning accuracy, the empirical contribution is currently not verifiable. With the missing components supplied (explicit f_k, parameter values, seeds, ablations, and a reproducibility artifact), the framework could be a valuable contribution to vehicular crowdsensing and edge fine-tuning research.

major comments (3)
  1. [Section IV-B, Eq. (14)] The data-collection reward is defined as r_{m,t} = beta * f_k(d^{p_{m,t,k}}_t, lambda^{p_{m,t,k}}_t), but the function f_k(d, lambda) is never specified. The text only says that it "encapsulates the fine-tuning accuracy" and Figure 3 gives qualitative curves without equations or fitted parameters. Because f_k enters every ADU computation (Eq. 4), the QoS objective (Eq. 5), and the learning signal of the MARL agents, the entire evaluation depends on an unverifiable quantity. The paper must provide the functional form of f_k, all fitted coefficients, the measurement protocol that connects data volume and AoI to fine-tuning accuracy, and preferably the simulator code; otherwise the reported ADU and accuracy gains cannot be reproduced or independently checked.
  2. [Section III-C, Eq. (5) and Section IV-B, Eqs. (13)-(14)] The QoS metric is not independent of the RL reward: the per-step rewards are exactly alpha * order_price and beta * f_k(d, lambda), i.e., the components of QoS. Consequently, the QoS values in Tables III and V are cumulative training returns rather than an external performance measure. Moreover, the values of alpha and beta are never stated, and no normalization is described. In Table V, for "Ours One Task," QoS = 66.24 with ADI = 2,834.13 and ADU = 8,474.18, which is inconsistent with the simple linear form QoS = alpha*ADI + beta*ADU unless extremely small weights are used. The authors should state alpha and beta (and any scaling), or re-report QoS in a way that is numerically self-consistent, and ideally evaluate the learned policy by measuring actual fine-tuning accuracy on a held-out dataset rather than through the same f_k used in the reward.
  3. [Section V-B, Figs. 8 and Tables IV-V] All reported results appear to come from a single run. There are no error bars, no number of random seeds, and no statistical tests (e.g., t-test, Mann-Whitney, or confidence intervals). Given that MAPPO and IQL are stochastic and MARL training is notoriously seed-sensitive, the claim that the proposed method outperforms all baselines across QoS, ADI, ADU, and accuracy is not statistically supported. The authors should run multiple seeds and report means and standard deviations across seeds for all plots and tables. In addition, the baseline set omits the closest MARL methods for ride-hailing dispatching and vehicular crowdsensing, such as the ones cited as [12], [50], and [51]; including them would strengthen the comparison.
minor comments (6)
  1. [Section I, paragraph 1] "PETF techniques" in the scenario description should be "PEFT techniques" (parameter-efficient fine-tuning).
  2. [Section III-A, end of first paragraph] The text says "such as DiDi or Urber"; "Urber" is a typo for "Uber."
  3. [Section V-A, evaluation metrics] The paper lists "the average AoI" as a performance metric but never reports it in any figure or table. Either present the result or remove the metric from the list.
  4. [Section V-C, Table II] The text states that rank 6 has fine-tuning time "nearly double" that of rank 3, but Table II reports normalized fine-tune times of 3.71 (rank 3) and 5.31 (rank 6), which is only about 43% longer. Please correct the wording to match the numbers.
  5. [Section IV-D, Algorithm 1] The RankTuner description is ambiguous: it is not clear whether ADU_curr in Algorithm 1 refers to the total fleet ADU, the agent-local ADU, or an ADU computed during a single fine-tuning iteration. Clarify the scope of ADU used in the rank update decision.
  6. [Section IV-C, Definition 4] The "shortcut node" that connects all grid nodes is an ad hoc architectural choice with no explanation or ablation. A small ablation showing QoS with and without the shortcut node would justify its inclusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reward/QoS alignment is by design, and the empirical claims are benchmarked against external data and baselines.

full rationale

The paper's QoS metric (Eq. 5) is defined as a weighted sum of ADI and ADU, and the per-step rewards in Eqs. (12)-(14) are exactly the per-step contributions to those sums. This is not a circular derivation; it is the standard consistency requirement between an RL objective and its reward function. The paper does not claim to derive QoS from rewards, nor to predict a quantity that was used as an input. The central empirical claim is that the GNN-enhanced MARL policy outperforms Random, Greedy-in-OS, Greedy-in-FT, MAB, and IQL on QoS, ADI, ADU, and fine-tuning accuracy in a simulator built from the NYC Taxi dataset and real UFM fine-tuning tasks. That comparison is self-contained and does not reduce to a fit or to a self-citation. The data-utility function f_k in Eq. (14) is never specified, which is a reproducibility and validity risk, but under the quoted-reduction standard it is not circularity because no equation in the paper shows a predicted outcome being identical to a fitted input. The self-citations [6] and [53] are contextual and not load-bearing for the main claims. Therefore no circular step meeting the paper's own equations was found, and the score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim rests on a small number of modeling choices. The QoS objective is a hand-weighted linear sum (alpha, beta unspecified), the data utility function f_k is assumed but never given, and the simulator uses simplified travel and order-expiration rules. None of these are derived from first principles; they are domain assumptions or ad hoc settings. The external anchors, real fine-tuning accuracy measurements and NYC taxi order data, provide some grounding.

free parameters (4)
  • QoS weights alpha, beta = not reported
    Eq. (5) defines QoS as alpha*ADI + beta*ADU; the authors say stakeholders determine them but never report the values used in experiments, so the objective and all comparisons depend on unspecified hand-chosen weights.
  • Data utility function f_k(d, lambda) = unspecified function
    Used as reward in Eq. (14) and in ADU; Fig. 3 shows measured accuracy curves, but no closed form, fitted parameters, or code is given, so the utility model cannot be independently checked.
  • RankTuner parameters = not reported
    Algorithm 1 requires eta_min, eta_max, eta_0, direction, and step; the paper does not state the values, although Table III results depend on them.
  • PoI data volume and distribution settings = 3-12 packages; three manually constructed distributions
    Section V-A generates PoI data volumes and locations by random sampling from author-defined distributions, not real PoI data; this determines ADU and QoS outcomes.
assumptions (6)
  • ad hoc to paper Data utility for fine-tuning is fully captured by a task-specific function f_k(d, lambda) of data volume and AoI.
    Invoked in Definition 2 and Eq. (14); it ignores computation time, communication cost, model state, and other real-world factors.
  • domain assumption The environment is a Markov game with stationary but unknown distributions of orders and PoIs.
    Section IV-B models the problem as a Markov game; MARL convergence and the learned policy rely on this assumption.
  • domain assumption Vehicles that accept orders or collect data become unavailable and return to availability according to simplifications in the problem formulation.
    Section III-A and constraints (7)-(11) abstract away trip durations, queueing, and multi-slot service; reward accumulation is per time slot only.
  • ad hoc to paper QoS defined as a linear weighted sum of ADI and ADU is the correct objective for the system.
    Definition 3; a normative choice by the authors and stakeholders, with no derivation or validation against real stakeholder preferences.
  • domain assumption Travel to any neighboring grid takes exactly one time slot.
    Section V-A states this simplification, citing prior work; it ignores traffic, weather, and route variability.
  • domain assumption An order expires after a fixed number of time slots if not accepted.
    Section V-A sets the active order lifetime to 15 slots, which shapes the ADI reward and the learned dispatching behavior.
invented entities (1)
  • Shortcut node in the topology graph
    purpose: Connects all grid nodes to accelerate information propagation in the R-GCN state encoder
    Introduced in Definition 4 (Section IV-C) as a graph construct; no ablation or external evidence that it improves learning, so it is an unverified design element.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Online Location Planning for AI-Defined Vehicles: Optimizing Joint Tasks of Order Serving and Spatio-Temporal Heterogeneous Model Fine-Tuning." pith.science (2026). https://pith.science/paper/5S3AA32X

@misc{pith2026250204399,
  author       = {Pith},
  title        = {Pith review of: Online Location Planning for AI-Defined Vehicles: Optimizing Joint Tasks of Order Serving and Spatio-Temporal Heterogeneous Model Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5S3AA32X}},
  note         = {Machine review of arXiv:2502.04399}
}
read the original 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 fine-tuning 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.

Figures

Figures reproduced from arXiv: 2502.04399 by the authors.

Figure 1
Figure 1. An illustration of the proposed vehicle scheduling [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Illustrates how an order request that is not completed [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The impact of data freshness and data volume on the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: GNN-based MARL framework. It includes the environment, a GNN embedding module for processing raw state [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: An example of constructing a topology graph. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: The distribution probability of orders and PoIs in the [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Different results of different distributions. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Performance comparison with baselines using Distribution 1 of the data [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Rank Variation Over Training Steps. higher than that based on Greedy-in-FT, which is in line with our expectations because if the priority of accepting orders is higher, the frequency of vehicles choosing to accept orders will be higher compared to collecting data. Our…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 58 canonical work pages

  1. [12]

    Multi-agent rein- forcement learning for urban crowd sensing with for-hire vehicles,

    R. Ding, Z. Yang, Y . Wei, H. Jin, and X. Wang, “Multi-agent rein- forcement learning for urban crowd sensing with for-hire vehicles,” in IEEE INFOCOM 2021-IEEE Conference on Computer Communications . IEEE, 2021, pp. 1–10

    work page 2021

  2. [50]

    Efficient large-scale fleet man- agement via multi-agent deep reinforcement learning,

    K. Lin, R. Zhao, Z. Xu, and J. Zhou, “Efficient large-scale fleet man- agement via multi-agent deep reinforcement learning,” in Proceedings of the 24th ACM SIGKDD , 2018

    work page 2018

  3. [51]

    Optimizing long-term efficiency and fairness in ride-hailing via joint order dispatching and driver repositioning,

    J. Sun, H. Jin, Z. Yang, L. Su, and X. Wang, “Optimizing long-term efficiency and fairness in ride-hailing via joint order dispatching and driver repositioning,” in Proceedings of the 28th ACM SIGKDD , 2022

    work page 2022

  4. [1]

    Alleviating corporate environmental pollution threats toward public health and safety: the role of smart city and artificial intelligence,

    W. Liu, Y . Xu, D. Fan, Y . Li, X.-F. Shao, and J. Zheng, “Alleviating corporate environmental pollution threats toward public health and safety: the role of smart city and artificial intelligence,” Safety Science , vol. 143, p. 105433, 2021

    work page 2021

  5. [2]

    Survey on traffic prediction in smart cities,

    A. M. Nagy and V . Simon, “Survey on traffic prediction in smart cities,” Pervasive and Mobile Computing , vol. 50, pp. 148–163, 2018

    work page 2018

  6. [3]

    Smart health: Big data enabled health paradigm within smart cities,

    M. I. Pramanik, R. Y . Lau, H. Demirkan, and M. A. K. Azad, “Smart health: Big data enabled health paradigm within smart cities,” Expert Systems with Applications , vol. 87, pp. 370–383, 2017

    work page 2017

  7. [4]

    Mobile crowdsourcing in smart cities: Technologies, applications, and future challenges,

    X. Kong, X. Liu, B. Jedari, M. Li, L. Wan, and F. Xia, “Mobile crowdsourcing in smart cities: Technologies, applications, and future challenges,” IEEE Internet of Things Journal , vol. 6, no. 5, pp. 8095– 8113, 2019

    work page 2019

  8. [5]

    Crowdsensing big data: sensing, data selection, and understanding,

    S. Zhai, R. Li, and Y . Yang, “Crowdsensing big data: sensing, data selection, and understanding,” in Journal of Physics: Conference Series , vol. 1848, no. 1. IOP Publishing, 2021

    work page 2021

Show all 75 references
  1. [6]

    Raccoon: Online content recommendation and edge-assisted caching for in-vehicle infotainment,

    Y . You, X. Zhang, T. Zhu, B. Rao, Y . Li, J. Duan, L. Xue, E. Zhai, and X. Chen, “Raccoon: Online content recommendation and edge-assisted caching for in-vehicle infotainment,” in International Conference on Big Data Computing and Communications (BigCom) , 2024

    2024

  2. [7]

    Urban foundation models: A survey,

    W. Zhang, J. Han, Z. Xu, H. Ni, H. Liu, and H. Xiong, “Urban foundation models: A survey,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2024, pp. 6633– 6643

    2024

  3. [8]

    On the opportunities and challenges of foundation models for geospatial artificial intelligence,

    G. Mai, W. Huang, J. Sun, S. Song, D. Mishra, N. Liu, S. Gao, T. Liu, G. Cong, Y . Hu et al. , “On the opportunities and challenges of foundation models for geospatial artificial intelligence,” arXiv preprint arXiv:2304.06798, 2023

    2023 arXiv

  4. [9]

    Ai foundation models in remote sensing: A survey,

    S. Lu, J. Guo, J. R. Zimmer-Dauphinee, J. M. Nieusma, X. Wang, P. VanValkenburgh, S. A. Wernke, and Y . Huo, “Ai foundation models in remote sensing: A survey,” arXiv preprint arXiv:2408.03464 , 2024

    2024 arXiv

  5. [10]

    A survey for foundation models in autonomous driving,

    H. Gao, Z. Wang, Y . Li, K. Long, M. Yang, and Y . Shen, “A survey for foundation models in autonomous driving,” arXiv preprint arXiv:2402.01105, 2024

    2024 arXiv

  6. [11]

    Parameter-efficient fine-tuning for large models: A comprehensive survey,

    Z. Han, C. Gao, J. Liu, J. Zhang, and S. Q. Zhang, “Parameter-efficient fine-tuning for large models: A comprehensive survey,” arXiv preprint arXiv:2403.14608, 2024

    2024 arXiv

  7. [13]

    Intelligent marketing in smart cities: Crowd- sourced data for geo-conquesting,

    B.-W. Chen and W. Ji, “Intelligent marketing in smart cities: Crowd- sourced data for geo-conquesting,” It Professional , vol. 18, no. 4, pp. 18–24, 2016

    2016

  8. [14]

    Data collection through mobile vehicles in edge network of smart city,

    Y . Luo, X. Zhu, and J. Long, “Data collection through mobile vehicles in edge network of smart city,” IEEE access , vol. 7, pp. 168 467–168 483, 2019

    2019

  9. [15]

    Towards fine- grained spatio-temporal coverage for vehicular urban sensing systems,

    G. Fan, Y . Zhao, Z. Guo, H. Jin, X. Gan, and X. Wang, “Towards fine- grained spatio-temporal coverage for vehicular urban sensing systems,” in IEEE INFOCOM 2021 . IEEE, 2021

    2021

  10. [16]

    Privacy-preserving sta- ble crowdsensing data trading for unknown market,

    H. Sun, M. Xiao, Y . Xu, G. Gao, and S. Zhang, “Privacy-preserving sta- ble crowdsensing data trading for unknown market,” in IEEE INFOCOM

  11. [17]

    Privacy-preserving online task assignment in spatial crowdsourcing: A graph-based ap- proach,

    H. Wang, E. Wang, Y . Yang, J. Wu, and F. Dressler, “Privacy-preserving online task assignment in spatial crowdsourcing: A graph-based ap- proach,” in IEEE INFOCOM 2022 . IEEE, 2022

    2022

  12. [18]

    A decentralized location privacy-preserving spatial crowdsourcing for internet of vehi- cles,

    J. Zhang, F. Yang, Z. Ma, Z. Wang, X. Liu, and J. Ma, “A decentralized location privacy-preserving spatial crowdsourcing for internet of vehi- cles,” IEEE Transactions on Intelligent Transportation Systems , vol. 22, no. 4, pp. 2299–2313, 2020

    2020

  13. [19]

    A deep learning-based mobile crowdsensing scheme by predicting vehicle mo- bility,

    X. Zhu, Y . Luo, A. Liu, W. Tang, and M. Z. A. Bhuiyan, “A deep learning-based mobile crowdsensing scheme by predicting vehicle mo- bility,” IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 7, pp. 4648–4659, 2020

    2020

  14. [20]

    Exploring both individuality and cooperation for air-ground spatial crowdsourcing by multi-agent deep reinforcement learning,

    Y . Ye, C. H. Liu, Z. Dai, J. Zhao, Y . Yuan, G. Wang, and J. Tang, “Exploring both individuality and cooperation for air-ground spatial crowdsourcing by multi-agent deep reinforcement learning,” in ICDE

  15. [21]

    Ehta: An environment-cost-based heterogeneous task allocation in vehicular crowdsensing,

    Y . Lu, X. Wang, A. Hawbani, P. Liu, L. Zhao, and Z. Liu, “Ehta: An environment-cost-based heterogeneous task allocation in vehicular crowdsensing,” IEEE Transactions on Mobile Computing , 2024

    2024

  16. [22]

    Privacy-preserving traffic monitoring with false report filtering via fog-assisted vehicular crowdsensing,

    M. Li, L. Zhu, and X. Lin, “Privacy-preserving traffic monitoring with false report filtering via fog-assisted vehicular crowdsensing,” IEEE Transactions on Services Computing , vol. 14, no. 6, pp. 1902–1913, 2019

    1902

  17. [23]

    Machine learning-based models for real-time traffic flow prediction in vehicular networks,

    P. Sun, N. Aljeri, and A. Boukerche, “Machine learning-based models for real-time traffic flow prediction in vehicular networks,” IEEE Net- work, vol. 34, no. 3, pp. 178–185, 2020

    2020

  18. [24]

    Real-time traffic conges- tion prediction using big data and machine learning techniques,

    P. Chawla, R. Hasurkar, C. R. Bogadi, N. S. Korlapati, R. Rajendran, S. Ravichandran, S. C. Tolem, and J. Z. Gao, “Real-time traffic conges- tion prediction using big data and machine learning techniques,” World Journal of Engineering , vol. 21, no. 1, pp. 140–155, 2024

    2024

  19. [25]

    Dynamic routing optimization in logistics us- ing machine learning: Towards efficient and sustainable supply chains,

    N. Kumar and S. Sharma, “Dynamic routing optimization in logistics us- ing machine learning: Towards efficient and sustainable supply chains,” Journal of Innovative Technologies , vol. 6, no. 1, pp. 1–7, 2023

    2023

  20. [26]

    An automated machine learning (automl) method of risk prediction for decision-making of autonomous vehicles,

    X. Shi, Y . D. Wong, C. Chai, and M. Z.-F. Li, “An automated machine learning (automl) method of risk prediction for decision-making of autonomous vehicles,” IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 11, pp. 7145–7154, 2020. 15

    2020

  21. [27]

    Giov: Achieving generative ai services in internet of vehicles via collaborative edge intelligence,

    G. Xie, R. Xie, X. Zhang, J. Nie, Q. Tang, W. Y . B. Lim, and D. Niyato, “Giov: Achieving generative ai services in internet of vehicles via collaborative edge intelligence,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC) . IEEE, 2024, pp. 1–6

    2024

  22. [28]

    Gai-iov: Bridging generative ai and vehicular networks for ubiquitous edge intelligence,

    G. Xie, Z. Xiong, X. Zhang, R. Xie, S. Guo, M. Guizani, and H. V . Poor, “Gai-iov: Bridging generative ai and vehicular networks for ubiquitous edge intelligence,” IEEE Transactions on Wireless Communications , 2024

    2024

  23. [29]

    Transfer learning-driven intrusion detection for internet of vehicles (iov),

    Y . Otoum, Y . Wan, and A. Nayak, “Transfer learning-driven intrusion detection for internet of vehicles (iov),” in 2022 International Wireless Communications and Mobile Computing (IWCMC) . IEEE, 2022, pp. 342–347

    2022

  24. [30]

    Geosam: Fine-tuning sam with sparse and dense visual prompting for automated segmentation of mobility infrastructure,

    R. I. Sultan, C. Li, H. Zhu, P. Khanduri, M. Brocanelli, and D. Zhu, “Geosam: Fine-tuning sam with sparse and dense visual prompting for automated segmentation of mobility infrastructure,” arXiv preprint arXiv:2311.11319, 2023

    2023 arXiv

  25. [31]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4015–4026

    2023

  26. [32]

    Ringmo-sam: A foundation model for segment anything in multimodal remote-sensing images,

    Z. Yan, J. Li, X. Li, R. Zhou, W. Zhang, Y . Feng, W. Diao, K. Fu, and X. Sun, “Ringmo-sam: A foundation model for segment anything in multimodal remote-sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–16, 2023

    2023

  27. [33]

    Geoclip: Clip- inspired alignment between locations and images for effective worldwide geo-localization,

    V . Vivanco Cepeda, G. K. Nayak, and M. Shah, “Geoclip: Clip- inspired alignment between locations and images for effective worldwide geo-localization,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  28. [34]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al. , “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

    2021

  29. [35]

    Accessed: Jul.7,2020

    Didichuxing. Accessed: Jul.7,2020. [Online]. Available: https://www. didiglobal.com/

    2020

  30. [36]

    Accessed: Jul.7,2020

    Uber. Accessed: Jul.7,2020. [Online]. Available: https://www.uber.com

    2020

  31. [37]

    A taxi order dispatch model based on combinatorial optimization,

    L. Zhang, T. Hu, Y . Min, G. Wu, J. Zhang, P. Feng, P. Gong, and J. Ye, “A taxi order dispatch model based on combinatorial optimization,” in Proceedings of the 23rd ACM SIGKDD , 2017

    2017

  32. [38]

    Data- driven transportation network company vehicle scheduling with users’ location differential privacy preservation,

    X. Zhang, J. Wang, H. Zhang, L. Li, M. Pan, and Z. Han, “Data- driven transportation network company vehicle scheduling with users’ location differential privacy preservation,” IEEE Transactions on Mobile Computing, vol. 22, no. 2, pp. 813–823, 2021

    2021

  33. [39]

    Beyond shortest paths: Route recommendations for ride-sharing,

    C. F. Yuen, A. P. Singh, S. Goyal, S. Ranu, and A. Bagchi, “Beyond shortest paths: Route recommendations for ride-sharing,” in The World Wide Web Conference, 2019

    2019

  34. [40]

    Privatehunt: Multi-source data-driven dispatching in for-hire vehicle systems,

    X. Xie, F. Zhang, and D. Zhang, “Privatehunt: Multi-source data-driven dispatching in for-hire vehicle systems,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 2, no. 1, pp. 1–26, 2018

    2018

  35. [41]

    Model predictive control of autonomous mobility-on-demand systems,

    R. Zhang, F. Rossi, and M. Pavone, “Model predictive control of autonomous mobility-on-demand systems,” in 2016 IEEE international conference on robotics and automation (ICRA) . IEEE, 2016

    2016

  36. [42]

    Towards supply-demand equilibrium with ridesharing: An elastic order dispatching algorithm in mod system,

    S. Ge, X. Zhou, T. Qiu, G. Wu, and X. Wang, “Towards supply-demand equilibrium with ridesharing: An elastic order dispatching algorithm in mod system,” IEEE Transactions on Mobile Computing , 2023

    2023

  37. [43]

    Deep reinforcement learning: A brief survey,

    K. Arulkumaran, M. P. Deisenroth, M. Brundage, and A. A. Bharath, “Deep reinforcement learning: A brief survey,” IEEE Signal Processing Magazine, vol. 34, no. 6, 2017

    2017

  38. [44]

    Deep reinforcement learning for intelligent transportation systems: A survey,

    A. Haydari and Y . Yılmaz, “Deep reinforcement learning for intelligent transportation systems: A survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 1, 2020

    2020

  39. [45]

    Multi-task-oriented vehicular crowdsensing: A deep learning approach,

    C. H. Liu, Z. Dai, H. Yang, and J. Tang, “Multi-task-oriented vehicular crowdsensing: A deep learning approach,” in IEEE INFOCOM 2020 . IEEE, 2020

    2020

  40. [46]

    Impala: Scalable dis- tributed deep-rl with importance weighted actor-learner architectures,

    L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V . Mnih, T. Ward, Y . Doron, V . Firoiu, T. Harley, I. Dunninget al., “Impala: Scalable dis- tributed deep-rl with importance weighted actor-learner architectures,” in International conference on machine learning . PMLR, 2018, pp....

    2018

  41. [47]

    Multi-agent reinforce- ment learning for urban crowd sensing with for-hire vehicles,

    R. Ding, Z. Yang, Y . Wei, H. Jin, and X. Wang, “Multi-agent reinforce- ment learning for urban crowd sensing with for-hire vehicles,” in IEEE INFOCOM 2021 . IEEE, 2021

    2021

  42. [48]

    Movi: A model-free approach to dynamic fleet management,

    T. Oda and C. Joe-Wong, “Movi: A model-free approach to dynamic fleet management,” in IEEE INFOCOM 2018 . IEEE, 2018

    2018

  43. [49]

    Context-aware taxi dispatching at city-scale using deep reinforcement learning,

    Z. Liu, J. Li, and K. Wu, “Context-aware taxi dispatching at city-scale using deep reinforcement learning,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 3, pp. 1996–2009, 2020

    1996

  44. [52]

    Multi-agent deep reinforcement learning based scheduling approach for mobile charging in internet of electric vehicles,

    L. Liu, Z. Huang, and J. Xu, “Multi-agent deep reinforcement learning based scheduling approach for mobile charging in internet of electric vehicles,” IEEE Transactions on Mobile Computing , 2024

    2024

  45. [53]

    Tapfinger: Task place- ment and fine-grained resource allocation for edge machine learning,

    Y . Li, T. Zeng, X. Zhang, J. Duan, and C. Wu, “Tapfinger: Task place- ment and fine-grained resource allocation for edge machine learning,” in IEEE INFOCOM 2023 , 2023

    2023

  46. [54]

    Hetero- geneous gnn-rl-based task offloading for uav-aided smart agriculture,

    T. Pamuklu, A. Syed, W. S. Kennedy, and M. Erol-Kantarci, “Hetero- geneous gnn-rl-based task offloading for uav-aided smart agriculture,” IEEE Networking Letters , vol. 5, no. 4, pp. 213–217, 2023

    2023

  47. [55]

    Multi-agent graph-attention communication and teaming

    Y . Niu, R. R. Paleja, and M. C. Gombolay, “Multi-agent graph-attention communication and teaming.” in AAMAS, vol. 21, 2021, p. 20th

    2021

  48. [56]

    Gnn-rl: Dynamic reward mechanism for connected vehicle security using graph neural networks and rein- forcement learning,

    H. Rathore and H. Griffith, “Gnn-rl: Dynamic reward mechanism for connected vehicle security using graph neural networks and rein- forcement learning,” in 2023 IEEE International Conference on Smart Computing (SMARTCOMP) . IEEE, 2023, pp. 201–203

    2023

  49. [57]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  50. [58]

    On the role of age of information in the internet of things,

    M. A. Abd-Elmagid, N. Pappas, and H. S. Dhillon, “On the role of age of information in the internet of things,” IEEE Communications Magazine , vol. 57, no. 12, pp. 72–77, 2019

    2019

  51. [59]

    Freshness-aware incentive mechanism for mobile crowdsensing with budget constraint,

    Y . Cheng, X. Wang, P. Zhou, X. Zhang, and W. Wu, “Freshness-aware incentive mechanism for mobile crowdsensing with budget constraint,” IEEE Transactions on Services Computing , 2023

    2023

  52. [60]

    Scaling laws for neural language models,

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, “Scaling laws for neural language models,” arXiv preprint arXiv:2001.08361 , 2020

    2001 arXiv

  53. [61]

    Multi-agent reinforcement learning: A selective overview of theories and algorithms,

    K. Zhang, Z. Yang, and T. Bas ¸ar, “Multi-agent reinforcement learning: A selective overview of theories and algorithms,” Handbook of rein- forcement learning and control , pp. 321–384, 2021

    2021

  54. [62]

    Modeling relational data with graph convolutional networks,

    M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, and M. Welling, “Modeling relational data with graph convolutional networks,” in The Semantic Web: 15th International Conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, Proceedings 15 . Springer, 2018

    2018

  55. [63]

    Federated fine-tuning of large language models under heterogeneous language tasks and client resources,

    J. Bai, D. Chen, B. Qian, L. Yao, and Y . Li, “Federated fine-tuning of large language models under heterogeneous language tasks and client resources,” arXiv e-prints , pp. arXiv–2402, 2024

    2024

  56. [64]

    The surprising effectiveness of ppo in cooperative multi-agent games,

    C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu, “The surprising effectiveness of ppo in cooperative multi-agent games,” Ad- vances in Neural Information Processing Systems , vol. 35, pp. 24 611– 24 624, 2022

    2022

  57. [65]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,” arXiv preprint arXiv:1707.06347 , 2017

    2017 arXiv

  58. [66]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hinton et al. , “Learning multiple layers of features from tiny images,” 2009

    2009

  59. [67]

    Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set,

    S. Ji, S. Wei, and M. Lu, “Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set,” IEEE Transactions on geoscience and remote sensing , vol. 57, no. 1, pp. 574–586, 2018

    2018

  60. [68]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 7464–7475

    2023

  61. [69]

    Vehicle detection dataset,

    O. detection, “Vehicle detection dataset,” https://universe.roboflow. com/object-detection-agrap/vehicle-detection-nmzlp, mar 2024, visited on 2024-12-12. [Online]. Available: https://universe.roboflow.com/ object-detection-agrap/vehicle-detection-nmzlp

    2024

  62. [70]

    New york city taxi datasets,

    “New york city taxi datasets,” https://www1.nyc.gov/site/tlc/about/ tlc-trip-record-data.page

  63. [71]

    Coride: joint order dispatching and fleet management for multi-scale ride-hailing platforms,

    J. Jin, M. Zhou, W. Zhang, M. Li, Z. Guo, Z. Qin, Y . Jiao, X. Tang, C. Wang, J. Wang et al. , “Coride: joint order dispatching and fleet management for multi-scale ride-hailing platforms,” in Proceedings of the 28th ACM CIKM , 2019

    2019

  64. [72]

    Algorithms for multi-armed bandit prob- lems,

    V . Kuleshov and D. Precup, “Algorithms for multi-armed bandit prob- lems,” arXiv preprint arXiv:1402.6028 , 2014

    2014 arXiv

  65. [73]

    An empirical evaluation of thompson sampling,

    O. Chapelle and L. Li, “An empirical evaluation of thompson sampling,” Advances in neural information processing systems , vol. 24, 2011

    2011

  66. [74]

    Thompson sampling and approximate inference,

    M. Phan, Y . Abbasi Yadkori, and J. Domke, “Thompson sampling and approximate inference,” Advances in Neural Information Processing Systems, vol. 32, 2019. 16 Bokeng Zheng is currently pursuing a bachelor’s degree with the School of Computer Science and Engineering, Sun Yat-se...

    2019

  67. [2018]

    He also works with the Department of Communica- tions, Pengcheng Laboratory, Shenzhen, China

    He is currently a Research Assistant Professor with the Institute of Future Networks, Southern Uni- versity of Science and Technology, Shenzhen, China. He also works with the Department of Communica- tions, Pengcheng Laboratory, Shenzhen, China. His research interest includes ...

    2025

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.