Pith. sign in

REVIEW 3 major objections 4 minor 34 references

FUSION couples offline forecasting and contracting with online potential-game scheduling to raise social welfare in air-ground edge networks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-03 16:11 UTC pith:DUJIARN5

load-bearing objection FUSION's online-stage convergence proof doesn't survive contact with its own payment terms; the forecasting and auction pieces are solid but the end-to-end evaluation is circular. the 3 major comments →

arxiv 2512.14323 v4 pith:DUJIARN5 submitted 2025-12-16 cs.NI

FUSION: Forecast-Embedded Agent Scheduling with Service Incentive Optimization over Distributed Air-Ground Edge Networks

classification cs.NI
keywords air-ground edge computingUAV-assisted computingliquid neural network forecastingauction mechanism designpotential gamestask schedulinghuman-machine coexistencesocial welfare
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes FUSION, a two-stage scheduler for air-ground edge networks in which vehicles ferry UAVs to overloaded servers. In the offline stage, a compact liquid neural network forecasts spatio-temporal demand, an ant-colony routine plans vehicle routes, and a critical-price double auction signs capacity contracts between servers and agent pairs. In the online stage, human and machine users decide locally or offload in a congestion game that the paper argues is an ordinal potential game, so best-response updates converge to a pure-strategy Nash equilibrium. The paper claims this combination raises social welfare and resource utilization while keeping latency, energy, and interaction overhead comparable to baselines, and it preserves individual rationality, near-truthfulness, and budget balance.

Core claim

FUSION's central claim is that proactive, forecast-driven deployment of vehicle–UAV agent pairs can be cleanly separated from reactive, congestion-aware task assignment without losing efficiency. The offline stage uses demand forecasts to pick high-value server–agent contracts and routes; the online stage frames each user's local-versus-offload choice as a finite game with an ordinal potential function Φ(π) that aggregates user-side benefits, congestion costs, and server operating costs. Any unilateral improvement that raises a user's utility is asserted to increase Φ, so asynchronous best-response dynamics reach a pure-strategy Nash equilibrium in finitely many steps, and that equilibrium i

What carries the argument

The load-bearing object is the ordinal potential function Φ(π) of Eq. (21), which sums non-congestion user valuations, subtracts the integrated congestion cost at each server, and subtracts server operating cost. Its separable form—user-specific load factors θ and server congestion functions τ(y) with an integral–pointwise sandwich inequality—is what aligns individual user incentives with system-wide welfare. The second pillar is the two-stage decomposition: Pro-LNN (a compact liquid neural network), eACO-VRP (ant-colony routing), and Off-AIC2 (a critical-price double auction) act offline, while PG-BRD (best-response dynamics over Φ) acts online.

Load-bearing premise

The claimed convergence of the online scheduler to a stable equilibrium rests on payments being neutral transfers that do not affect the potential, even though payments appear in users' utilities; if a change in price outweighs a congestion benefit, a utility-improving move might not raise the system potential, and the proof would not apply.

What would settle it

Take a two-server, two-user instance, run PG-BRD with different service prices, and record Φ(π) after each accepted best response; if any accepted update strictly raises the user's utility yet leaves Φ unchanged or lowers it, the ordinal-potential theorem fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the ordinal-potential claim holds, distributed best-response scheduling converges without a central optimizer, making the scheme deployable at the edge.
  • Forecast-driven contracting lets servers reserve UAV capacity before demand spikes, which the simulations indicate improves social welfare at scale.
  • The compact forecaster (roughly 11k parameters) degrades less under missing/noisy inputs than larger recurrent or attention baselines, suggesting edge-friendly deployment.
  • The auction's critical-value pricing yields approximately truthful reports in the numerical experiments, supporting the use of such mechanisms under spatial coupling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • I read the paper as arguing that forecasting quality—not just online adaptation—is the main lever for welfare; a stronger test would predict user-type composition, not only aggregate workload.
  • The ordinal-potential proof deliberately excludes payments from Φ; a natural extension is to check whether a potential that includes payment-induced externalities restores the convergence guarantee under price discrimination.
  • The framework could be stress-tested with real traffic traces and higher prediction error to see whether the offline routing gains persist when the forecaster is less accurate.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes FUSION, a two-stage framework for distributed air-ground edge networks. In the offline stage, a liquid-neural-network forecaster (Pro-LNN) predicts per-ES demand; an ant-colony route planner (eACO-VRP) constructs vehicle-UAV routes; and a double auction (Off-AIC2) determines ES-AP contracts. In the online stage, HU/MU task offloading is modeled as an ordinal potential game, and a best-response algorithm (PG-BRD) is claimed to converge to a pure-strategy Nash equilibrium. The paper claims higher social welfare, better resource utilization, comparable latency/energy, and individual rationality, budget balance, and near-truthfulness.

Significance. If the theoretical guarantees held, FUSION would be a substantial engineering contribution: it integrates forecasting, routing, contracting, and online scheduling under heterogeneous human/machine QoS requirements, and the decomposition is sensible. The paper has a real strength in the standalone forecasting evaluation: Pro-LNN is compared against LSTM and Transformer on accuracy, model footprint, and robustness to noisy/missing inputs, using both synthetic and real-world data, and the results are internally consistent. However, the central convergence proof for the online stage is invalid as written because the proposed ordinal potential omits payment terms that appear in player utilities, and the system-level evaluation is circular because the realized demand is generated from the framework's own forecast. These are load-bearing issues: they undermine the main theoretical and empirical claims, so the contribution is not currently established.

major comments (3)
  1. [Sec. 5.2, Eq. (19) vs. Eq. (21); Appx. D.7/D.9] The ordinal-potential claim is invalid. The unified SD utility in Eq. (19) includes the payment term -p_{j,i}, but the potential Phi(pi) in Eq. (21) contains no payment term. For a unilateral deviation from i to i', the utility change in Eq. (60) contains -(p_{j,i'} - p_{j,i}), while the potential change in Eq. (61) does not. Appx. D.9 argues that payments are internal transfers and 'can be omitted from the potential without changing ordinal comparisons.' That argument applies to aggregate welfare, not to an ordinal potential, which requires sign agreement between each player's actual utility change and the potential change. A concrete one-SD, two-SP counterexample: let tau_1(y)=y, tau_2(y)=2y, S_j=0, kappa_j theta_j,i=1, p_{j,1}=2, p_{j,2}=0, and the SD's load y=1. Then U(SP1)=-3, U(SP2)=-2 (so the SD strictly prefers SP2), but Phi(SP1)=-0.5 and Phi(SP2)=-1 (so the potential strictly de
  2. [Sec. 6.3, footnote 6] The system-level evaluation is circular. Footnote 6 states that after training Pro-LNN on the UCI traces, 'the forecasted demand profiles drive the time-varying SD task-arrival intensities at each ES.' In other words, the realized online demand is generated from Pro-LNN's own forecast output, so the end-to-end SW comparison evaluates FUSION against demand that it itself produced. This removes the very forecast error that the offline stage is designed to hedge against, and it invalidates the claimed advantage of forecast-aware offline planning in Figs. 2(a)-(b). The standalone forecasting benchmarks in Sec. 6.2 are valid and are compared to ground truth, but they do not repair the circularity of the coupled system experiments.
  3. [Appx. B.2, Eq. (47), Prop. 6] The budget-balance proof is incomplete. Eq. (47) asserts ask_k <= r^ES_{i,k} <= p^ES_{i,k} <= bid_{i,k} for every matched pair, but the binary-search pricing in Step 3 computes the ES payment p^b*_i by varying only the ES's effective bid (Alg. 2, lines 18-25) and the AP reward r^s*_k by varying only the AP's effective ask (lines 26-33), with no coupling that forces p^ES_{i,k} >= r^ES_{i,k}. The surplus-feasible band in Eq. (44) only ensures both quantities lie in [ask_k, bid_{i,k}], not that the payment is at least the reward. Therefore U^P >= 0 in Eq. (40) is not established. This is a gap in a property explicitly claimed in the abstract and conclusion.
minor comments (4)
  1. [Sec. 4.3.2, Eq. (8) and Alg. 1 line 14] The notation for the utility increment is inconsistent: Eq. (8) uses Delta U^A_{r,s}, while Algorithm 1 line 14 and Step 4 use Delta U^W_{r,s}.
  2. [Alg. 2, line 12] Line 12 reads 'K*_k <- Alg. 2;' which appears to be a self-reference; it should invoke the eACO-VRP subroutine (Alg. 1).
  3. [Eq. (2)] The AP utility expression has a typesetting issue: '+ omega2 chard' appears without a preceding plus sign, making the formula ambiguous.
  4. [Appx. B.2, Prop. 5] The near-truthfulness proof is presented as a sequence of informal 'approximately monotone' arguments and is not a formal bound. Since the paper claims 'near-truthfulness' as a property, either a concrete approximation factor or a precise formulation of the guarantee is needed, not just empirical curves.

Circularity Check

1 steps flagged

System-level SW evaluation generates demand from Pro-LNN's own forecasts, making the central end-to-end result a self-consistency loop; standalone forecast validation is non-circular.

specific steps
  1. self definitional [Sec. 6.3 'Performance Evaluation on FUSION' and footnote 6, used in Sec. 6.3.1]
    "We then train Pro-LNN on these profiles to forecast the per-ES demand6, and then evaluate FUSION as follows (i.e., in Sec. 6.3.1–Sec. 6.3.3). ... 6. ... Pro-LNN is trained on these time series and used to forecast the demand N Dem,(ν)_i of ES s_i over future trading rounds; the forecasted demand profiles drive the time-varying SD task-arrival intensities at each ES, while the total population sizes of SDs and SPs are kept fixed to control the overall network scale."

    The system-level simulator defines the realized per-ES demand as Pro-LNN's own forecast output: "the forecasted demand profiles drive the time-varying SD task-arrival intensities at each ES." Thus the offline stage's forecast-driven contracting and routing are evaluated against demand that the same predictor generated, i.e., zero forecast error by construction. The reported social-welfare advantage of FUSION is therefore a self-consistency check of the forecast-embedded pipeline rather than an out-of-sample test of proactive provisioning under forecast error. The standalone forecast evaluation in Sec. 6.2.2 uses ground-truth UCI traces and is non-circular, but the central end-to-end claim is not independently tested.

full rationale

The only concrete circular equivalence in the derivation/evaluation chain is the Sec. 6.3 evaluation loop: the forecasted demand profiles generated by Pro-LNN are used as the actual demand realization in the system-level experiments, so the offline forecast-driven mechanism is matched against its own predictions. This makes the central SW result partly forced by construction. The standalone Pro-LNN experiments (Sec. 6.2.2) are measured against ground-truth UCI data and thus provide independent support for the forecasting module, preventing a score of 8-10. No load-bearing self-citation chain was found; author self-citations appear for parameters and heuristics rather than for a uniqueness theorem or ansatz. A separate correctness gap exists in the online-stage proof: ΔU_j in Eq. (60) contains the payment difference −(p_{j,i'}−p_{j,i}) while ΔΦ in Eq. (61) omits payments, and Appx. D.9 asserts payments "can be omitted from the potential without changing ordinal comparisons." This is a proof error/unsupported inference rather than a circular reduction, so it is noted but not scored as circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The system model introduces numerous hand-chosen coefficients (ω, V, γ, ε, ACO parameters, LNN hyperparameters) and assumes a linearizable, payment-free potential structure that is not justified. The AP and auctioneer are system constructs with no external validation. The central theoretical and empirical claims rest on these ad hoc assumptions.

free parameters (6)
  • Cost coefficients ω1, ω2
    Convert energy into monetary units in AP utility (2) and SP cost (16). No values are given in the paper; SW rankings depend on their relative size.
  • HU/MU valuation coefficients V_H_t, V_H_e, V_M_e
    Set sensitivities to latency and energy in utilities (13) and (15). Values are not reported; they determine best responses and the resulting NE.
  • WPS priority weights γ_{j,i}
    Define congestion load y_i and congestion function τ_i in (10)-(11); no values or distributions are provided, yet they drive the potential function.
  • PG-BRD acceptance threshold ε
    Controls ε-NE vs exact NE in Alg. 3. The abstract claims convergence to ε-NE for ε>0 and pure NE for ε=0, but the proof of potential increase fails even at ε=0.
  • eACO-VRP hyperparameters ε1, ε2, ε3, θ, ant count, iterations
    Appear in the transition rule (6) and pheromone update (7)-(8). No values are specified; route selection directly affects offline matching and SW.
  • Pro-LNN hyperparameters (hidden dim, L, H, learning rate, regularization λ)
    Training details in Appx. B.1 omit concrete dimensions and schedules; forecasting accuracy determines offline contracts and downstream performance.
axioms (6)
  • domain assumption Users' utilities have separable congestion structure U_{j,i}=S_j(i)-κ_j θ_{j,i} τ_i(y_i)-p_{j,i}
    Introduced in Eq. (19); requires HU utilities (13) and linearized MU utilities (18) to share the same congestion term, which only holds after an approximate linearization in Appx D.4.
  • ad hoc to paper Monetary payments can be excluded from the ordinal potential without changing the sign of individual improvements
    Appx D.9 states payments are internal transfers and omitted from Φ(π); this is false for individual utilities because payments differ across actions and appear in Eq. (19).
  • ad hoc to paper MU soft-deadline utility can be linearized around typical delay levels without changing the NE
    Appx D.4 replaces G^M_{j,i} with a linear form and asserts ordinal ranking is preserved in the operating regime; no bound is given for when this holds.
  • domain assumption WPS service discipline with τ_i(y)=y/f_i for all SPs
    Adopted from [13] to model congestion in (10)-(11); not validated in this system.
  • ad hoc to paper Predicted demand traces can stand in for realized future demand in system evaluation
    Sec. 6.3 footnote 6: forecasted profiles drive task-arrival intensities, making the end-to-end evaluation circular.
  • domain assumption Distance-dependent path-loss and SNR ranges follow [31]
    Simulation channel model inherited from prior work; affects all latency and energy numbers.
invented entities (2)
  • Vehicular-UAV agent pairs (APs) no independent evidence
    purpose: Ground vehicles transport UAVs to hotspots, then UAVs augment ES capacity; this is the core deployment mechanism.
    No real-world validation or measured cost model; all AP capabilities and energy parameters are assumed.
  • Auctioneer/third-party platform no independent evidence
    purpose: Runs Off-AIC2, collects bids/asks, and sets payments and rewards.
    A trusted third party is assumed to exist and be honest; standard but unverified.

pith-pipeline@v1.3.0-alltime-deepseek · 37400 in / 15673 out tokens · 130313 ms · 2026-08-03T16:11:34.478555+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of FUSION: Forecast-Embedded Agent Scheduling with Service Incentive Optimization over Distributed Air-Ground Edge Networks." pith.science (2026). https://pith.science/paper/DUJIARN5

@misc{pith2026251214323,
  author       = {Pith},
  title        = {Pith review of: FUSION: Forecast-Embedded Agent Scheduling with Service Incentive Optimization over Distributed Air-Ground Edge Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUJIARN5}},
  note         = {Machine review of arXiv:2512.14323}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this paper, we introduce a first-of-its-kind forecasting-driven, incentive-aware service provisioning framework for distributed air-ground integrated networks that explicitly accounts for human-machine coexistence. In our framework, vehicular-UAV agent pairs (APs) are proactively dispatched to overloaded hotspots to augment the computing capacity of edge servers (ESs), which in turn gives rise to a set of challenges that we jointly address: highly uncertain spatio-temporal workloads, spatio-temporal coupling between road traffic and UAV capacity, forecast-driven contracting risks, and heterogeneous quality-of-service (QoS) requirements of human users (HUs) and machine users (MUs). To address these challenges, we propose FUSION, a two-stage optimization framework, consisting of an offline stage and an online stage. In the offline stage, a liquid neural network-powered module performs multi-step spatio-temporal demand forecasting at distributed ESs, whose outputs are exploited by an enhanced ant colony optimization-based routing scheme and an auction-based incentive-aware contracting mechanism, to jointly determine ES-AP contracts and pre-planned service routes. In the online stage, we model congestion-aware task scheduling as a service demander-side exact-potential game, where HUs and MUs select either local execution or accessible ES/UAV resources, and develop a potential-guided best-response dynamics algorithm that converges to an \varepsilon-NE under a positive improvement threshold and to a pure-strategy Nash equilibrium when the threshold is zero. Experiments on both synthetic and real-world datasets show that FUSION consistently achieves higher social welfare and improved resource utilization, while maintaining latency and energy costs comparable to state-of-the-art baselines and preserving individual rationality, budget balance, and near-truthfulness.

Figures

Figures reproduced from arXiv: 2512.14323 by Houyi Qi, Liqun Fu, Minghui Liwang, Sai Zou, Seyyedali Hosseinalipour, Wei Ni, Xianbin Wang, Yiguang Hong.

Figure 1
Figure 1. Figure 1: Schematic of FUSION over an air–ground integrated network. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance comparisons and economic property analyses, where (a)-(b) SW under different problem sizes (80 HUs, 80 MUs, and 5 APs in [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 1 linked inside Pith

  1. [1]

    Edge-computing-driven Internet of Things: A survey,

    L. Kong, J. Tan, J. Huang, G. Chen, S. Wang, X. Jin, P . Zeng, M. Khan, and S. K. Das, "Edge-computing-driven Internet of Things: A survey,"ACM Comput. Surv., vol. 55, no. 8, Art. no. 174, pp. 1-41, Dec. 2023

  2. [2]

    Mobile edge computing and machine learning in the Internet of unmanned aerial vehicles: A survey,

    Z. Ning, H. Hu, X. Wang, L. Guo, S. Guo, G. Wang, and X. Gao, "Mobile edge computing and machine learning in the Internet of unmanned aerial vehicles: A survey,"ACM Comput. Surv., vol. 56, no. 1, Art. no. 13, pp. 1-31, Aug. 2024

  3. [3]

    Bridge the Present and Future: A Cross-Layer Matching Game in Dynamic Cloud-Aided Mobile Edge Networks,

    H. Qi, M. Liwang, X. Wang, L. Li, W. Gong, J. Jin, and Z. Jiao, "Bridge the Present and Future: A Cross-Layer Matching Game in Dynamic Cloud-Aided Mobile Edge Networks,"IEEE Trans. Mobile Comput., vol. 23, no. 12, pp. 12522- 12523, Dec. 2024

  4. [4]

    Multi-objective optimization for multi-UAV-assisted mobile edge computing,

    G. Sun, Y. Wang, Z. Sun, Q. Wu, J. Kang, D. Niyato, and V . C. M. Leung, "Multi-objective optimization for multi-UAV-assisted mobile edge computing," IEEE Trans. Mobile Comput., vol. 23, no. 12, pp. 14803-14820, 2024

  5. [5]

    UAV-assisted mobile edge computing: Optimal design of UAV altitude and task offloading,

    M. Hui, J. Chen, L. Yang, L. Lv, H. Jiang, and N. Al-Dhahir, "UAV-assisted mobile edge computing: Optimal design of UAV altitude and task offloading," IEEE Trans. Wireless Commun., vol. 23, no. 10, pp. 13633-13647, Oct. 2024

  6. [6]

    Energy efficiency optimization of IRS and UAV-assisted wireless powered edge networks,

    X. Wang, J. Li, J. Wu, L. Guo, and Z. Ning, "Energy efficiency optimization of IRS and UAV-assisted wireless powered edge networks,"IEEE J. Sel. Topics Signal Process., vol. 18, no. 7, pp. 1297-1310, 2024

  7. [7]

    Low latency and accuracy- guaranteed DNN inference for UAV-assisted IoT networks,

    J. Xu, H. Yao, R. Zhang, T. Mai, and M. Guizani, "Low latency and accuracy- guaranteed DNN inference for UAV-assisted IoT networks,"IEEE Trans. Cogn. Commun. Netw., early access, 2025, doi: 10.1109/TCCN.2025.3542443

  8. [8]

    Distributed optimization of resource efficiency for federated edge intelligence in AAV-enabled IoT networks,

    Y. Li, J. Rao, L. Wang, Y. Xiao, X. Ge, and G. Shi, "Distributed optimization of resource efficiency for federated edge intelligence in AAV-enabled IoT networks,"IEEE Internet Things J., vol. 12, no. 8, pp. 9449-9464, Apr. 2025

  9. [9]

    A3D: Adaptive, accurate, and autonomous navigation for edge-assisted drones,

    L. Zeng, H. Chen, D. Feng, X. Zhang, and X. Chen, "A3D: Adaptive, accurate, and autonomous navigation for edge-assisted drones,"IEEE/ACM Trans. Netw., vol. 32, no. 1, pp. 713-728, Feb. 2024

  10. [10]

    UAV-aided vehicular short-packet communication and edge computing system under time-varying channel,

    S. Shen, K. Yang, K. Wang, and G. Zhang, "UAV-aided vehicular short-packet communication and edge computing system under time-varying channel," IEEE Trans. Veh. Technol., vol. 72, no. 5, pp. 6625-6638, May 2023

  11. [11]

    CPDZ: A credibility-aware and privacy-preserving data collection scheme with zero-trust in next-generation crowdsensing networks,

    J. Tang, K. Fan, S. Yang, A. Liu, N. N. Xiong, H. H. Song, and V . C. M. Leung, "CPDZ: A credibility-aware and privacy-preserving data collection scheme with zero-trust in next-generation crowdsensing networks,"IEEE J. Sel. Areas Commun., vol. 43, no. 6, pp. 2183-2199, 2025

  12. [12]

    A Potential Game Theoretic Approach to Computation Offloading Strategy Optimization in End-Edge-Cloud Com- puting,

    Y. Ding, K. Li, C. Liu, and K. Li, "A Potential Game Theoretic Approach to Computation Offloading Strategy Optimization in End-Edge-Cloud Com- puting,"IEEE Trans. Parallel Distrib. Syst., vol. 33, no. 6, pp. 1503-1519, Jun. 2022

  13. [13]

    Potential Game-Based Computation Offloading in Edge Computing With Heterogeneous Edge Servers,

    Z. Zhou, L. Pan, and S. Liu, "Potential Game-Based Computation Offloading in Edge Computing With Heterogeneous Edge Servers,"IEEE Trans. Netw. Sci. Eng., vol. 12, no. 1, pp. 290-301, Jan. 2025

  14. [14]

    Auction approaches for resource allocation in wireless systems: A survey,

    C. Zhang, C. Lee, D. Niyato, and P . Wang, "Auction approaches for resource allocation in wireless systems: A survey,"IEEE Commun. Surveys Tuts., vol. 15, no. 3, pp. 1020-1041, 3rd Quart., 2013

  15. [15]

    Truthful auction mechanisms for dependent task offloading in vehicular edge computing,

    H. Ren, K. Liu, G. Yan, C. Liu, Y. Li, C. Li, and W. Wu, "Truthful auction mechanisms for dependent task offloading in vehicular edge computing," IEEE Trans. Mobile Comput., vol. 23, no. 12, pp. 14987-15002, Dec. 2024

  16. [16]

    CADEC: A combinatorial auction for dynamic distributed DNN inference scheduling in edge-cloud networks,

    X. Xu, Y. Hu, G. Cui, L. Qi, W. Dou, and Z. Cai, "CADEC: A combinatorial auction for dynamic distributed DNN inference scheduling in edge-cloud networks,"IEEE Trans. Mobile Comput., vol. 24, no. 10, pp. 10024-10041, Oct. 2025

  17. [17]

    An online auction approach to computing resource allocation in mobile AIGC networks,

    X. Li, K. Mo, Y. Hou, Z. Li, H. Xu, and N. Guan, "An online auction approach to computing resource allocation in mobile AIGC networks,"IEEE Internet Things J., vol. 12, no. 8, pp. 11182-11193, Apr. 2025

  18. [18]

    Auction- based cluster federated learning in mobile edge computing systems,

    R. Lu, W. Zhang, Y. Wang, Q. Li, X. Zhong, H. Yang, and D. Wang, "Auction- based cluster federated learning in mobile edge computing systems,"IEEE Trans. Parallel Distrib. Syst., vol. 34, no. 4, pp. 1145-1158, Apr. 2023

  19. [19]

    Edge-cloud offloading: Knapsack PG in 5G multi-access edge computing,

    C.-Y. Hsieh, Y. Ren, and J.-C. Chen, "Edge-cloud offloading: Knapsack PG in 5G multi-access edge computing,"IEEE Trans. Wireless Commun., vol. 22, no. 11, pp. 7158-7171, Nov. 2023

  20. [20]

    Computation offloading in multi- cell networks with collaborative edge-cloud computing: A game theoretic approach,

    L. Wu, P . Sun, Z. Wang, Y. Li, and Y. Yang, "Computation offloading in multi- cell networks with collaborative edge-cloud computing: A game theoretic approach,"IEEE Trans. Mobile Comput., vol. 23, no. 3, pp. 2093-2106, Mar. 2024

  21. [21]

    AI-native end-to-end network slicing for next-generation mission-critical services,

    A. R. Hossain, W. Liu, and N. Ansari, "AI-native end-to-end network slicing for next-generation mission-critical services,"IEEE Trans. Cogn. Commun. Netw., vol. 11, no. 1, pp. 48-58, 2025

  22. [22]

    Effective two-stage double auction for dynamic resource provision over edge networks via discovering the power of overbooking,

    S. Wu, M. Liwang, D. Wang, X. Wang, C. Wu, J. Tang, L. Li, and X. Xia, "Effective two-stage double auction for dynamic resource provision over edge networks via discovering the power of overbooking,"IEEE Trans. Serv. Comput., early access, pp. 1-13, 2025

  23. [23]

    Robust flight navigation out of distribution with liquid neural networks,

    M. Chahine, R. Hasani, P . Kao, A. Ray, R. Shubert, M. Lechner, A. Amini, and D. Rus, "Robust flight navigation out of distribution with liquid neural networks,"Sci. Robot., vol. 8, no. 77, Art. eadc8892, 2023

  24. [24]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, "Layer normalization," arXiv preprint arXiv:1607.06450, 2016

  25. [25]

    Accelerating stable matching between workers and spatial-temporal tasks for dynamic MCS: A stagewise service trading approach,

    H. Qi, M. Liwang, X. Wang, L. Fu, Y. Hong, L. Li, and Z. Cheng, "Accelerating stable matching between workers and spatial-temporal tasks for dynamic MCS: A stagewise service trading approach,"IEEE Trans. Mobile Comput., early access, pp. 1-16, 2025, doi: 10.1109/TMC.2025.3610915

  26. [26]

    UAV path planning for target coverage task in dynamic environment,

    J. Li, Y. Xiong, and J. She, "UAV path planning for target coverage task in dynamic environment,"IEEE Internet Things J., vol. 10, no. 20, pp. 17734-17745, 2023

  27. [27]

    Strategic Information Perturbation for an Online In-Vehicle Coordinated Routing Mechanism for Connected Vehicles Under Mixed-Strategy Congestion Game,

    S. Spana, L. Du and Y. Yin, "Strategic Information Perturbation for an Online In-Vehicle Coordinated Routing Mechanism for Connected Vehicles Under Mixed-Strategy Congestion Game,"IEEE Trans. Intell. Transp. Syst., vol. 23, no. 5, pp. 4541-4555, May 2022. 16

  28. [28]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, "Long short-term memory,"Neural Comput., vol. 9, no. 8, pp. 1735-1780, Nov. 1997, doi: 10.1162/neco.1997.9.8.1735

  29. [29]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, "Attention is all you need," inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017

  30. [30]

    Future resource bank for ISAC: Achieving fast and stable win-win matching for both individuals and coalitions,

    H. Qi, M. Liwang, S. Hosseinalipour, L. Fu, S. Zou, and W. Ni, "Future resource bank for ISAC: Achieving fast and stable win-win matching for both individuals and coalitions,"IEEE J. Sel. Areas Commun., early access, 2025, doi: 10.1109/JSAC.2025.3608114

  31. [31]

    Dynamic offloading and trajectory control for UAV-enabled mobile edge computing system with energy harvesting devices,

    Z. Yang, S. Bi, and Y.-J. A. Zhang, "Dynamic offloading and trajectory control for UAV-enabled mobile edge computing system with energy harvesting devices,"IEEE Trans. Wireless Commun., vol. 21, no. 12, pp. 10515-10528, 2022

  32. [32]

    New services & applications with 5G ultra-reliable low latency communications,

    5G Americas, "New services & applications with 5G ultra-reliable low latency communications," White Paper, Nov. 2018

  33. [33]

    5G; Study on scenarios and requirements for next generation access technologies,

    3GPP , "5G; Study on scenarios and requirements for next generation access technologies," 3GPP TR 38.913, ver. 17.0.0, Release 17, 2022

  34. [34]

    ElectricityLoadDiagrams20112014,

    A. Trindade, "ElectricityLoadDiagrams20112014," UCI Mach. Learn. Reposi- tory, 2015. [Online]. Available: https://doi.org/10.24432/C58C86. 17 APPENDIXA KEYNOTATIONS Key notations in this paper are summarized in Table 5. APPENDIXB DETAILED OFOFFLINESTAGE B.1 Detailed Pro-LNN Architecture and Training Proce- dure In this section, we provide the detailed upd...