Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Asynchronous Distributed Gaussian Process Regression for Online Learning and Dynamical Systems: Complementary Document

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read An asynchronous distributed Gaussian-process predictor that fuses stale local predictions still carries a time-varying error bound, and a closed-loop controller using it keeps tracking error within a computable limit.

desk verdict Companion proof document for an AAAI paper: the core prediction bound is plausible, but the control theorem has a spectral assumption gap and the periodic-kernel Lipschitz constant is wrong. read the letter →

arxiv 2412.11950 v1 pith:XQTK6EYB submitted 2024-12-16 cs.LG cs.SYeess.SY

Zewen Yang , Xiaobing Dai , Sandra Hirche This is my paper · ORCID
classification cs.LGcs.SYeess.SY MSC 62M2068T0560G15
keywords asynchronousdistributedGaussianprocessesonlinelearningpredictionerrorbounddelayedpredictionstrackingcontrolkernelLipschitzconstantsprocessregression
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 companion document supplies the proofs behind an asynchronous distributed Gaussian-process regression method for online learning and dynamical systems. It aims to establish that even when local GP models send back predictions computed from outdated data, their weighted fusion still satisfies a time-varying prediction-error bound that never exceeds $\beta\sigma_f$. It then shows that a closed-loop controller built on this predictor has tracking error converging to a computable limit proportional to that bound. If these results hold, real-time control need not wait for fresh GP updates: stale but certified predictions can be fused safely, with the price of staleness explicitly quantified. The document also derives Lipschitz constants for several common kernels, which are used to bound the error contributed by delays.

What carries the argument

The load-bearing object is the aggregation rule and the associated time-varying bound $\omega(t)$ defined in the companion paper. The weights $\omega_i^k(t)$ are chosen so that the weighted sum of delayed-error bounds can be compressed by Cauchy-Schwarz into $\omega(t)$, and the information-set rule sets $\rho_i^k(t)=0$ whenever $\eta_i^k(t) \ge \beta\sigma_f$, which is what forces the global bound below $\beta\sigma_f$. For the control half, the machinery is the matrix-exponential estimate $\|\exp(At)\| \le \|Q\|\|Q^{-1}\|e^{\bar{\Lambda}t}$ under the Hurwitz assumption on $A$; that estimate is what carries the prediction certificate into the control guarantee.

What would settle it

Recompute the Lipschitz constant of the periodic kernel by differentiating $\sigma_f^2\exp(-2\sin^2(\pi\|x-x'\|/p)/\sigma_l^2)$ with respect to $\|x-x'\|$ while keeping the chain-rule cosine factor; if the maximum is not the two-case value in Corollary 4, the delayed-prediction bound for that kernel is unsupported and the global certificate does not apply to it.

Watch

Extended reading notes

Core claim

The central claim is Theorem 7 of the companion paper: for the asynchronous distributed GP predictor $\hat{f}(x(t)) = \sum_{i}\sum_{k}\omega_i^k(t)\mu_i(x(t_k^i)) + \omega_m(t)m(x(t))$, the prediction error satisfies $|f(x(t)) - \hat{f}(x(t))| \le \omega(t)$, where $\omega(t)$ is a time-varying quantity built from the delayed-error bounds $\eta_i^k(t)$ and the prior bound $\beta\sigma_f$. Corollary 8 sharpens this to $\omega(t) \le \beta\sigma_f$ by exploiting the information-set rule that admits only delayed predictions with $\eta_i^k(t) < \beta\sigma_f$. Theorem 9 feeds this bound into the error dynamics $\dot{e} = Ae + b(f - \hat{f})$ and obtains the ultimate tracking bound $\lim_{t\to\infty}\|e(t)\| \le \|Q\|\|Q^{-1}\|\bar{\omega}/|\bar{\Lambda}|$. The contribution is a formal certificate that fusing stale local GP predictions in real time remains safe, with the cost of asynchrony measured by $\omega(t)$.

Load-bearing premise

The whole argument depends on the companion paper's per-local-model error bounds and on the algorithm's rule admitting only delayed predictions with a sufficiently small error bound; for the periodic kernel that foundation is missing because the stated Lipschitz constant comes from a derivative that omits the chain-rule cosine factor.

Editorial extensions

If this is right

  • A control designer can use asynchronous distributed GP predictions with stale data and still quote a formal worst-case prediction error at each time step, instead of treating staleness heuristically.
  • The ultimate tracking error scales linearly with the worst-case prediction bound and with the condition number $\|Q\|\|Q^{-1}\|$ of the error dynamics, so better GP accuracy directly improves the tracking guarantee.
  • The information-set rule that discards delayed predictions whose bound reaches $\beta\sigma_f$ yields a uniform cap on the prediction error, making the certificate independent of how large the delay grows.
  • The kernel Lipschitz constants give an explicit way to compute $\eta_i^k(t)$ for linear, squared-exponential, ARD-SE, and rational-quadratic kernels, so the bound can be evaluated in practice.
  • If the periodic-kernel Lipschitz constant were established, the same delayed-prediction bound would extend the certificate to periodic kernels as well.

Reading between the lines

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

  • The same weighting argument would likely survive replacing Gaussian-process experts with any base learner that supplies a per-expert error bound, since the proof consumes only the bounds $\eta_i^k(t)$ and $\beta\sigma_f$ rather than the GP likelihood itself.
  • A direct stress test of the certificate is to run the algorithm with artificially growing delay patterns while enforcing the information-set rule; the empirical prediction error should never cross $\omega(t)$.
  • If a simpler but looser certificate is preferred, the monotonicity in Corollary 8 suggests replacing the time-varying $\omega(t)$ by the constant $\beta\sigma_f$ in the tracking bound, at the price of a larger ultimate bound.
  • The periodic-kernel Lipschitz constant in Corollary 4 is not established by the given derivation, because that derivation drops the chain-rule cosine factor when differentiating the periodic kernel with respect to the distance; until this is corrected, the certificate should not be claimed for periodic kernels.
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 / 5 minor

Summary. This complementary document supplies proofs and additional simulations for a companion paper on asynchronous distributed Gaussian process regression. It proves a prediction-error bound |f(x(t)) - fhat(x(t))| <= omega(t) with omega(t) <= beta*sigma_f (Theorem 7, Corollary 8), derives a closed-loop tracking-error bound for a control law with error dynamics e_dot = Ae + b(f - fhat) (Theorem 9, eq. (24)), gives Lipschitz constants for several kernels (Table I, Corollaries 1-4), and reports delay-time and Monte-Carlo control simulations comparing AsyncDGP with BCM, rBCM, POE, gPOE, and MOE. The paper is explicitly a companion to [1] and relies on that paper for definitions of aggregation weights, the information-set rule, and supporting lemmas.

Significance. If the results are correct, the paper provides a formal time-varying error certificate for fusing stale local GP predictions in real-time control, which is a valuable contribution: the prediction bound is derived analytically rather than fitted, and the control bound gives a quantitative ultimate tracking error. The proof of Theorem 7 is internally coherent under the weight definitions inherited from [1], and the supplementary simulations illustrate the claimed advantage of aggregating delayed information. However, the control theorem's proof contains a load-bearing unproved spectral assumption, and the periodic-kernel Lipschitz derivation contains a concrete mathematical error. These issues do not necessarily invalidate the central prediction-bound result, but they prevent the paper from being accepted in its current form.

major comments (3)
  1. [Section III, eqs. (19)-(24)] The proof of Theorem 9 assumes that A admits a decomposition A = Q*Lambda*Q^{-1} with Lambda = diag(Lambda_1, ..., Lambda_n) and Lambda_i in R_{<0}, and then uses ||exp(At)|| <= ||Q|| ||Q^{-1}|| exp(Lambda_bar t). No justification is given that the matrix A defined in eq. (22) of the companion paper is diagonalizable with real negative eigenvalues. A general Hurwitz matrix can have complex-conjugate eigenvalues or be defective; in those cases the stated bound (20) and the ultimate bound (24) do not follow as written. This is load-bearing for the central control claim. The authors should either prove the spectral property for their specific A, add it as an explicit assumption, or redo the proof using the maximal real part of eigenvalues and a Jordan-form bound to handle defective cases.
  2. [Section IV-F, Corollary 4, eq. (48)] The derivative of the periodic kernel with respect to d_Per is computed incorrectly: differentiating exp(-2 sin^2(pi d/p)/sigma_l^2) produces an additional factor cos(pi d/p), which is omitted in eq. (48). Also, the line 'due to the fact s(x,x") >= 1' is false for s(x,x") = sin(pi ||x-x"||/p); the sine satisfies |s| <= 1. Consequently the Lipschitz constants for the periodic kernel in Table I are not established. The authors need to recompute the derivative, redo the maximization over the appropriate range, and correct the resulting constants. This does not directly affect the control experiments, which use the ARD-SE kernel, but it is a concrete defect in a stated result.
  3. [Section II-B, Theorem 7 proof] The proof of Theorem 7 depends essentially on quantities and facts that are not reproduced in this document: the definitions of the aggregation weights omega_i^k(t) and omega_m(t), the definition of omega(t) in eq. (12) of the main paper, the statement of Lemma 2 (posterior concentration), the statement of Lemma 3 (delayed prediction bound), and the information-set rule in Algorithm 1 that guarantees eta_i^k(t) < beta*sigma_f. As a complementary document this is understandable, but for a standalone verification the central prediction bound is not self-contained. I recommend restating the inherited lemmas and the weight definitions (or providing a precise, versioned pointer to the companion paper) so the proof can be checked without guessing.
minor comments (5)
  1. [Section IV-B] The subsection titled 'Corollaries of the Other Lipschitz constant of Kernel' contains no corollaries and appears to be a placeholder; either fill it with content or remove it.
  2. [Eq. (38)] The rational quadratic kernel is written as sigma_f^2 (1 + ||x-x'||^2/(2 alpha sigma_l^2)) without the exponent -alpha. The later derivative in eq. (41) is consistent with the usual definition with exponent -alpha, so eq. (38) appears to be a typo and should be corrected.
  3. [Section IV-A, eqs. (25)-(27)] The notation d_x(x, x') is used in the SE kernel proof without a clear definition; it should be defined explicitly as the chosen distance d_SE = ||x - x'||, or the notation should be unified with the rest of the section.
  4. [Section V-C] The control simulation paragraph states that the reference follows a uniform distribution between [0.4, 0.6] and [3, 5] for each simulation; it would be clearer to specify whether these are the ranges for the two components of the reference signal or two different simulation settings.
  5. [References] Reference [1] is cited as 'The 39th Annual AAAI Conference on Artificial Intelligence, 2024' with an OpenReview link; if the paper has since been published, the final venue and DOI should be provided.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the prediction and tracking bounds are analytic consequences of stated premises, not equivalents of their inputs.

full rationale

The document is a proof appendix to the authors' companion paper, and its derivations are not circular. Theorem 7 is obtained by an explicit Cauchy-Schwarz chain starting from the triangle inequality, Lemma 3's delayed-error bound, and the definition of omega(t) in the main paper; omega(t) is a formula, not a fitted or measured quantity, and the theorem establishes an inequality rather than renaming an input. Corollary 8 is a monotonicity argument using the Algorithm 1 condition eta_i^k(t) < beta sigma_f. Theorem 9 applies a standard exponential-decay estimate to the error dynamics (17); its questionable step, the unsupported assumption that A is diagonalizable with real negative eigenvalues, is a correctness gap, not a circular reduction, because the bound (24) is not assumed in the premises. The kernel Lipschitz computations are direct calculus; the periodic-kernel derivative in eq. (48) omits a cos factor and the text's 's(x,x') >= 1' is erroneous, but these are local mathematical defects. The paper does rely on Lemma 2, Algorithm 1, and eq. (12) from the companion paper; this is modular self-reference rather than circularity, since Lemma 2 is a standard GP concentration bound with stated assumptions external to the target result and no fitted parameter is relabeled as a prediction. Thus no step reduces by construction to its own input.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The theoretical results introduce no fitted constants: the only hand-set design value in the companion is the information set threshold I. The bounds are derived from standard GP prior assumptions (bounded RKHS norm, Lipschitz kernel, bounded prior error) inherited from the main paper. No new physical or algorithmic entities are postulated.

free parameters (1)
  • Information set threshold I = 4 or 10 in the simulations
    User-chosen number of retained past predictions per agent; it controls how many terms enter the bound and is set by hand, not estimated from data.
assumptions (6)
  • domain assumption The unknown function f lies in the RKHS of the GP prior with norm bound ||f||_kappa <= Gamma (used in Lemma 3, eq. (2)).
    This is a standard regularity assumption for GP regression error bounds, inherited from the main paper.
  • domain assumption The kernel kappa is Lipschitz continuous with constant L_kappa with respect to the chosen distance d (used in eq. (3) of Lemma 3).
    The delayed prediction error bound depends on this Lipschitz property; the companion derives constants for specific kernels, one of which is incorrect.
  • domain assumption The prior mean error is bounded as |f(x)-m(x)| <= beta*sigma_f (used in Theorem 7, eq. (6)).
    This is the standard GP prior confidence bound assumed by the aggregation framework.
  • domain assumption The matrix A is Hurwitz with all real negative eigenvalues Lambda_i (used in Theorem 9, eq. (19)).
    The tracking error bound uses exp(Lambda t) with the maximal real eigenvalue; complex eigenvalues are not handled in the proof.
  • domain assumption Algorithm 1's information set management guarantees eta_i^k(t) < beta*sigma_f (used in Corollary 8).
    Corollary 8's upper bound omega(t) <= beta*sigma_f relies on this algorithmic condition.
  • standard math Standard matrix inequalities, Cauchy-Schwarz, and the triangle inequality are used throughout the proofs.
    These are unproved background tools, used without issue except where noted in the periodic kernel derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Asynchronous Distributed Gaussian Process Regression for Online Learning and Dynamical Systems: Complementary Document." pith.science (2026). https://pith.science/paper/XQTK6EYB

@misc{pith2026241211950,
  author       = {Pith},
  title        = {Pith review of: Asynchronous Distributed Gaussian Process Regression for Online Learning and Dynamical Systems: Complementary Document},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XQTK6EYB}},
  note         = {Machine review of arXiv:2412.11950}
}
read the original abstract

This is a complementary document for the paper titled "Asynchronous Distributed Gaussian Process Regression for Online Learning and Dynamical Systems".

Figures

Figures reproduced from arXiv: 2412.11950 by the authors.

Figure 1
Figure 1. The delay time of the results of GP models. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 3
Figure 3. The delay time of the results on SARCOS. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. The prediction errors and tracking errors over 1000 iterations. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figures from the paper (1 more)
Figure 6
Figure 6. Figure 6: The violin plots for mean and maximum values of the Monte-Carlo simulation. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 39 canonical work pages

  1. [1]

    Asynchronous distributed gaussian process regression,

    X. D. Zewen Yang and S. Hirche, “Asynchronous distributed gaussian process regression,” in The 39th Annual AAAI Conference on Artificial Intelligence, 2024. [Online]. Available: https://openreview.net/forum? id=OklB7DL8Z0

    work page 2024

  2. [2]

    Can Learning Deteriorate Control? Analyzing Computational Delays in Gaussian Process-Based Event-Triggered Online Learning,

    X. Dai, A. Lederer, Z. Yang, and S. Hirche, “Can Learning Deteriorate Control? Analyzing Computational Delays in Gaussian Process-Based Event-Triggered Online Learning,” in Learning for Dynamics and Con- trol Conference. PMLR, 2023, pp. 445–457

    work page 2023

  3. [3]

    Sparse gaussian processes using pseudo-inputs,

    E. Snelson and Z. Ghahramani, “Sparse gaussian processes using pseudo-inputs,” Advances in neural information processing systems , vol. 18, 2005

    2005

  4. [4]

    Variational learning of inducing variables in sparse gaussian processes,

    M. Titsias, “Variational learning of inducing variables in sparse gaussian processes,” in Artificial intelligence and statistics . PMLR, 2009, pp. 567–574

    2009

  5. [5]

    A sparse covariance function for exact gaussian process inference in large datasets,

    A. Melkumyan and F. T. Ramos, “A sparse covariance function for exact gaussian process inference in large datasets,” in Twenty-first international joint conference on artificial intelligence , 2009

    work page 2009

  6. [6]

    A framework for bayesian optimization in embedded subspaces,

    A. Nayebi, A. Munteanu, and M. Poloczek, “A framework for bayesian optimization in embedded subspaces,” in International Conference on Machine Learning. PMLR, 2019, pp. 4752–4761

    work page 2019

  7. [7]

    When Gaussian Process Meets Big Data: A Review of Scalable GPs,

    H. Liu, Y .-S. Ong, X. Shen, and J. Cai, “When Gaussian Process Meets Big Data: A Review of Scalable GPs,” vol. 31, no. 11, pp. 4405–4423, 2020

    work page 2020

  8. [8]

    Gaussian processes for high- dimensional, large data sets: A review,

    M. Jiang, G. Pedrielli, and S. H. Ng, “Gaussian processes for high- dimensional, large data sets: A review,” in 2022 Winter Simulation Conference (WSC). IEEE, 2022, pp. 49–60. 9 AsyncDGP rBCM gPOE BCM POE MOE 3 × 10−1 4 × 10−1 6 × 10−1 ‖e‖ Tracking errors Mean Median AsyncDGP rBCM gPOE BCM POE MOE 2 × 100 3 × 100 4 × 100 6 × 100 |f (x) − ˆf (x)| Predictio...

    work page 2022

Show all 44 references
  1. [9]

    Mixtures of Gaussian Processes,

    V . Tresp, “Mixtures of Gaussian Processes,” in Advances in Neural Information Processing Systems , T. Leen, T. Dietterich, and V . Tresp, Eds., vol. 13. MIT Press, 2000. [On- line]. Available: https://proceedings.neurips.cc/paper files/paper/2000/ file/9fdb62f932adf55af2c0e09...

    2000

  2. [10]

    Variational mixture of gaussian process experts,

    C. Yuan and C. Neubauer, “Variational mixture of gaussian process experts,” Advances in neural information processing systems , vol. 21, 2008

    2008

  3. [11]

    Mixture of Experts: A Literature Survey,

    S. Masoudnia and R. Ebrahimpour, “Mixture of Experts: A Literature Survey,” vol. 42, no. 2, pp. 275–293, 2014. [Online]. Available: http://link.springer.com/10.1007/s10462-012-9338-y

    2014 doi

  4. [12]

    Training Products of Experts by Minimizing Contrastive Divergence,

    G. E. Hinton, “Training Products of Experts by Minimizing Contrastive Divergence,” Neural Computation, vol. 14, no. 8, pp. 1771–1800, 2002

    2002

  5. [13]

    Hierarchical Mixture-of-Experts Model for Large-Scale Gaussian Process Regression,

    J. W. Ng and M. P. Deisenroth, “Hierarchical Mixture-of-Experts Model for Large-Scale Gaussian Process Regression,” 2014

    2014

  6. [14]

    Generalized Product of Experts for Automatic and Principled Fusion of Gaussian Process Predictions,

    Y . Cao and D. J. Fleet, “Generalized Product of Experts for Automatic and Principled Fusion of Gaussian Process Predictions,” 2015. [Online]. Available: http://arxiv.org/abs/1410.7827

    2015 arXiv

  7. [15]

    Healing Products of Gaussian Process Experts,

    S. Cohen, R. Mbuvha, T. Marwala, and M. Deisenroth, “Healing Products of Gaussian Process Experts,” in Proceedings of the 37th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, H. D. III and A. Singh, Eds., vol

  8. [16]

    Correlated product of experts for sparse gaussian process regression,

    M. Sch ¨urch, D. Azzimonti, A. Benavoli, and M. Zaffalon, “Correlated product of experts for sparse gaussian process regression,” Machine Learning, pp. 1–22, 2023

    2023

  9. [17]

    A Bayesian committee machine,

    V . Tresp, “A Bayesian committee machine,” Neural computation , vol. 12, no. 11, pp. 2719–2741, 2000

    2000

  10. [18]

    Distributed Gaussian Processes,

    M. Deisenroth and J. W. Ng, “Distributed Gaussian Processes,” in Proceedings of the 32nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, F. Bach and D. Blei, Eds., vol. 37. Lille, France: PMLR, 07–09 Jul 2015, pp. 1481–1490. [Online...

    2015

  11. [19]

    Generalized robust bayesian committee machine for large-scale gaussian process regression,

    H. Liu, J. Cai, Y . Wang, and Y . S. Ong, “Generalized robust bayesian committee machine for large-scale gaussian process regression,” in International Conference on Machine Learning . PMLR, 2018, pp. 3131–3140

    2018

  12. [20]

    Nested Kriging predictions for datasets with a large number of observations,

    D. Rulli `ere, N. Durrande, F. Bachoc, and C. Chevalier, “Nested Kriging predictions for datasets with a large number of observations,” Statistics and Computing, vol. 28, pp. 849–867, 2018

    2018

  13. [21]

    Decentralized nested gaussian pro- cesses for multi-robot systems,

    G. P. Kontoudis and D. J. Stilwell, “Decentralized nested gaussian pro- cesses for multi-robot systems,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021, pp. 8881–8887

    2021

  14. [22]

    Cooperative Learning with Gaussian Processes for Euler- Lagrange Systems Tracking Control Under Switching Topologies,

    Z. Yang, S. Dong, A. Lederer, X. Dai, S. Chen, S. Sosnowski, G. Hattab, and S. Hirche, “Cooperative Learning with Gaussian Processes for Euler- Lagrange Systems Tracking Control Under Switching Topologies,” in 2024 American Control Conference (ACC) , 2024, pp. 560–567

    2024

  15. [23]

    Whom to Trust? Elective Learning for Distributed Gaussian Process Regression,

    Z. Yang, X. Dai, A. Dubey, S. Hirche, and G. Hattab, “Whom to Trust? Elective Learning for Distributed Gaussian Process Regression,” in Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems , ser. AAMAS ’24, 2024, p. 2020–2028

    2024

  16. [24]

    Distributed Learning Consensus Control for Unknown Nonlinear Multi- Agent Systems based on Gaussian Processes,

    Z. Yang, S. Sosnowski, Q. Liu, J. Jiao, A. Lederer, and S. Hirche, “Distributed Learning Consensus Control for Unknown Nonlinear Multi- Agent Systems based on Gaussian Processes,” in 2021 60th IEEE Conference on Decision and Control (CDC) . IEEE, 2021, pp. 4406– 4411

    2021

  17. [25]

    Cooperative Control of Uncertain Multiagent Systems via Distributed Gaussian Processes,

    A. Lederer, Z. Yang, J. Jiao, and S. Hirche, “Cooperative Control of Uncertain Multiagent Systems via Distributed Gaussian Processes,” IEEE Transactions on Automatic Control, vol. 68, no. 5, pp. 3091–3098, 2023

    2023

  18. [26]

    Collective online learning of gaussian processes in massive multi-agent systems,

    T. N. Hoang, Q. M. Hoang, K. H. Low, and J. How, “Collective online learning of gaussian processes in massive multi-agent systems,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 01, 2019, pp. 7850–7857

    2019

  19. [27]

    Learning- based Control for PMSM Using Distributed Gaussian Processes with Optimal Aggregation Strategy,

    Z. Yin, X. Dai, Z. Yang, Y . Shen, G. Hattab, and H. Zhao, “Learning- based Control for PMSM Using Distributed Gaussian Processes with Optimal Aggregation Strategy,” in IECON 2023- 49th Annual Confer- ence of the IEEE Industrial Electronics Society , 2023, pp. 1–7

    2023

  20. [28]

    Distributed Online Sparse Gaussian Process Regression for Multi-Agent Coverage Control,

    S. He, M. Tang, J. Fu, and J. Liang, “Distributed Online Sparse Gaussian Process Regression for Multi-Agent Coverage Control,” in 2023 42nd Chinese Control Conference (CCC) . IEEE, 2023, pp. 5464–5469. [Online]. Available: https://ieeexplore.ieee.org/document/10239920/

    2023

  21. [29]

    Coop- erative Online Learning for Multiagent System Control via Gaussian Processes With Event-Triggered Mechanism,

    X. Dai, Z. Yang, S. Zhang, D.-H. Zhai, Y . Xia, and S. Hirche, “Coop- erative Online Learning for Multiagent System Control via Gaussian Processes With Event-Triggered Mechanism,” IEEE Transactions on Neural Networks and Learning Systems , pp. 1–15, 2024

    2024

  22. [30]

    Decentralized event-triggered online learning for safe consensus control of multi-agent systems with Gaussian process regression,

    X. Dai, Z. Yang, M. Xu, S. Zhang, F. Liu, G. Hattab, and S. Hirche, “Decentralized event-triggered online learning for safe consensus control of multi-agent systems with Gaussian process regression,” European Journal of Control , vol. 80, p. 101058, 2024. [Online]. Available: ...

    2024

  23. [31]

    Gaussian process decentral- ized data fusion and active sensing for spatiotemporal traffic modeling and prediction in mobility-on-demand systems,

    J. Chen, K. H. Low, Y . Yao, and P. Jaillet, “Gaussian process decentral- ized data fusion and active sensing for spatiotemporal traffic modeling and prediction in mobility-on-demand systems,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 3, pp. 901–921, 2015

    2015

  24. [32]

    Gaussian process decentralized data fusion meets transfer learning in large-scale distributed cooperative perception,

    R. Ouyang and B. K. H. Low, “Gaussian process decentralized data fusion meets transfer learning in large-scale distributed cooperative perception,” Autonomous Robots, vol. 44, pp. 359–376, 2020. 10

    2020

  25. [33]

    Decentralized nested gaussian pro- cesses for multi-robot systems,

    G. P. Kontoudis and D. J. Stilwell, “Decentralized nested gaussian pro- cesses for multi-robot systems,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2021, pp. 8881–8887

    2021

  26. [34]

    Distributed multi-agent gaussian regression via finite-dimensional approximations,

    G. Pillonetto, L. Schenato, and D. Varagnolo, “Distributed multi-agent gaussian regression via finite-dimensional approximations,” IEEE Trans- actions on Pattern Analysis and Machine Intelligence, vol. 41, no. 9, pp. 2098–2111, 2018

    2018

  27. [35]

    Learning of gaussian processes in distributed and communication limited systems,

    M. Tavassolipour, S. A. Motahari, and M. T. M. Shalmani, “Learning of gaussian processes in distributed and communication limited systems,” IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 8, pp. 1928–1941, 2019

    1928

  28. [36]

    Aggregating dependent gaussian experts in local approximation,

    H. Jalali and G. Kasneci, “Aggregating dependent gaussian experts in local approximation,” in 2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021, pp. 9015–9022

    2020

  29. [37]

    Gaussian processes on graphs via spectral kernel learning,

    Y .-C. Zhi, Y . C. Ng, and X. Dong, “Gaussian processes on graphs via spectral kernel learning,” IEEE Transactions on Signal and Information Processing over Networks , vol. 9, pp. 304–314, 2023

    2023

  30. [38]

    Mobile sensor network navigation using gaussian processes with truncated observations,

    Y . Xu, J. Choi, and S. Oh, “Mobile sensor network navigation using gaussian processes with truncated observations,” IEEE Transactions on Robotics, vol. 27, no. 6, pp. 1118–1131, 2011

    2011

  31. [39]

    Multi-Robot Active Sensing and Environmental Model Learning With Distributed Gaussian Process,

    D. Jang, J. Yoo, C. Y . Son, D. Kim, and H. J. Kim, “Multi-Robot Active Sensing and Environmental Model Learning With Distributed Gaussian Process,” IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 5905– 5912, 2020

    2020

  32. [40]

    Adaptive Sampling and Online Learning in Multi-Robot Sensor Coverage with Mixture of Gaussian Processes,

    W. Luo and K. Sycara, “Adaptive Sampling and Online Learning in Multi-Robot Sensor Coverage with Mixture of Gaussian Processes,” in 2018 IEEE International Conference on Robotics and Automation (ICRA), 2018, pp. 6359–6364

    2018

  33. [41]

    Resource-Efficient Coop- erative Online Scalar Field Mapping Via Distributed Sparse Gaussian Process Regression,

    T. Ding, R. Zheng, S. Zhang, and M. Liu, “Resource-Efficient Coop- erative Online Scalar Field Mapping Via Distributed Sparse Gaussian Process Regression,” IEEE Robotics and Automation Letters , pp. 1–8, 2024

    2024

  34. [42]

    Learning-based Symbolic Abstractions for Nonlinear Control Systems,

    K. Hashimoto, A. Saoud, M. Kishida, T. Ushio, and D. V . Dimarogonas, “Learning-based Symbolic Abstractions for Nonlinear Control Systems,” Automatica, vol. 146, p. 110646, 2022

    2022

  35. [43]

    D. S. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas . Princeton university press, 2009

    2009

  36. [119]

    2068–2077

    PMLR, 13–18 Jul 2020, pp. 2068–2077. [Online]. Available: https://proceedings.mlr.press/v119/cohen20b.html

    2020

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.