Pith. sign in

REVIEW 4 major objections 4 minor 33 references

No-Regret Gaussian Process Optimization of Time-Varying Functions

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A windowed sparse GP-bandit algorithm tracks time-varying optima with almost no extra queries.

desk verdict Solid windowing idea and a worthwhile Fano lower bound, but Assumptions 1 and 2 cannot both hold, so the main theorem is about an empty class. read the letter →

arxiv 2512.00517 v3 pith:46UDHJFM submitted 2025-11-29 stat.ML cs.LGmath.OC

Eliabelle Mauduit , Eloïse Berthier , Andrea Simonetto This is my paper · ORCID
classification stat.MLcs.LGmath.OC MSC 68T0562G99
keywords Gaussianprocessbanditstime-varyingoptimizationnon-stationaryregretboundsuncertaintyinjectionsparsevariationalinferencedeterminantalpointprocessesdynamic
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 claims that no-regret optimization of a time-varying black-box function is achievable even when the function drifts persistently, provided the learner may occasionally ask for a refreshed evaluation of previously visited points. The authors propose W-SparQ-GP-UCB, an algorithm that refreshes only a sparse, determinantal-point-process-selected subset of past points at the start of each growing window; they prove that the average number of these additional queries per round tends to zero while the cumulative dynamic regret stays sublinear. They also prove that in the pure bandit setting, no algorithm can achieve sublinear regret once the drift is fast (α ≥ 1), and that any algorithm must make Ω(T^{α/(α+1)}) additional queries in that regime. Together these results give a nearly tight characterisation of how much side information is needed to track a drifting optimum.

What carries the argument

The uncertainty-injection model (Assumption 2), which replaces arbitrary temporal drift by independent zero-mean sub-Gaussian noise with variance σ²(t2−t1)^α; the windowed scheduling of sparse refreshes that keeps effective observation variance below O(t^ᾱ) with ᾱ<1/3; and the determinantal-point-process selection of which past points to refresh, together with a trace-residual monotonicity lemma for low-rank kernel approximations that transfers posterior closeness between the sparse and virtual Gaussian-process models.

What would settle it

Construct a sequence f_t where the drift increments v_t(x) = f_{t+1}(x)−f_t(x) are strongly negatively autocorrelated (e.g., f_t alternates between two fixed functions), so the cumulative variance of f_t(x)−f_s(x) is bounded rather than growing as (t−s)^α. If W-SparQ-GP-UCB still achieves sublinear regret with o(T) queries, the lower-bound premise is not necessary; if it suffers linear regret, the independence/variance-growth assumption is confirmed as load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that temporal drift can be modelled as injected uncertainty: an observation of the function at time t1, viewed from time t2, behaves like a noisy measurement with variance σ²(1+(t2−t1)^α), where α measures how fast the function evolves. Under this model, the paper proves that W-SparQ-GP-UCB has, with probability 1−δ, cumulative dynamic regret O(T^{2ᾱ+1} d log^{d+1} T (log(1/δ) + d T^ᾱ log^{d+1} T)) for any chosen ᾱ<1/3, which is sublinear for all α, while requiring only N_T = O(T^{−ᾱ/α} log^d T) additional queries on average — a quantity that vanishes. The companion lower bound shows that with fewer than Ω(T^{α/(α+1)}) additional queries, any algorithm suffers linear reg

Load-bearing premise

The guarantees rest on Assumption 2: that drift is zero-mean, independent across time pairs, sub-Gaussian, with variance growing exactly like σ²(t2−t1)^α and with α known (or bounded); if real drift is adversarial, correlated, or of unknown rate, the regret and query bounds can collapse—as the paper's own real-data experiment appears to show.

Editorial extensions

If this is right

  • If the algorithm and guarantees are correct, a learner can track a drifting optimum with sublinear regret while asking the expert only o(T) times in total, a cost that vanishes relative to the T rounds of the game.
  • The α=1/3 and α=1 thresholds separate regimes: standard GP-UCB remains no-regret for α<1/3, becomes superlinear for 1/3≤α<1, and no pure-bandit algorithm can beat linear regret for α≥1.
  • The query lower bound N_T ≥ Ω(T^{α/(α+1)}) means that in fast-drift environments any no-regret method must refresh a growing number of past points; the proposed algorithm's query count matches this rate up to polylog factors.
  • The same sparse-refresh mechanism reduces the computational cost of GP regression compared to full GP-UCB, since regression is performed on small, representative subsets.

Reading between the lines

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

  • The independence and known-α requirements suggest an immediate practical extension: estimate α online from the observed stream (e.g., from the growth of empirical prediction errors) and adapt window sizes; the theory would then need a regret bound that accounts for estimation error.
  • The windowed-refresh recipe is general: any online optimizer that periodically re-samples a sparse, informative subset of past actions converts a drifting problem into a near-stationary one; this could carry over to bandits beyond Gaussian processes, such as linear, contextual, or multi-armed bandits with drifting means.
  • The lower-bound technique—sliding windows plus a pigeonhole argument on cumulative information—could extend to other noise-growth laws (e.g., variance proportional to t^α log t) and to settings with delayed or batched side queries.
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

4 major / 4 minor

Summary. The paper studies time-varying Gaussian-process bandit optimization in a frequentist RKHS setting. Temporal drift is modeled by Assumption 2: for t1<t2, f_t1(x1)=f_t2(x1)+v_{t1,t2}, where v is zero-mean sub-Gaussian with variance σ^2(t2−t1)^α and is independent across time pairs. The paper derives a bandit regret upper bound for GP-UCB (Prop. 2.2), a lower bound in the bandit setting (Prop. 2.5), and proposes W-SparQ-GP-UCB, which uses expert side queries on a DPP-selected sparse set at window starts. The main result (Thm. 3.8) claims that, under Assumptions 1–2 and a window parameter \tildeα<1/3, W-SparQ-GP-UCB achieves R_T = O(T^{2\tildeα+1} d log^{d+1} T (log(1/δ)+dT^{\tildeα} log^{d+1} T)) with N_T/T = O(T^{-\tildeα/α} log^d T) = o(1). A lower bound (Thm. 5.2) states that N_T=O(T^{α/(α+1)}) side queries imply linear regret for α>1. The paper also contains numerical experiments on synthetic and real data.

Significance. The side-query model is a potentially interesting relaxation of the strict bandit setting, and the paper contains some useful auxiliary material: Lemma 3.7 on monotonicity of the Nyström residual trace is proved in detail, and the window-counting argument leading to N_T/T = o(1) is plausible. If the main theorem were correct, it would be a meaningful contribution to non-stationary GP optimization. However, the central theoretical claims rest on an Assumption pair that is internally inconsistent, and the key KL-transfer step in Theorem 3.8 is not proved in the required heteroscedastic form. As a result, the significance of the paper as a frequentist no-regret result is not established.

major comments (4)
  1. [Section 2, Assumptions 1 and 2 (Eqs. 2.6–2.7)] Assumptions 1 and 2 are mutually inconsistent for α>0. From Eq. (2.6) and k(x,x)≤M_k^2, every f_t satisfies |f_t(x)|≤B M_k, so the difference f_t1(x)−f_t2(x) is a.s. bounded by 2B M_k. A zero-mean random variable in that interval has variance bounded by a constant independent of t2−t1, whereas Eq. (2.7) requires variance proxy σ^2(t2−t1)^α, which is unbounded for any α>0. Thus no nontrivial sequence satisfies Assumptions 1 and 2, and Theorem 3.8 is a statement about an empty class. The independence condition is also inconsistent: for a<b<c, Eq. (2.7) forces v_{a,c}=v_{a,b}+v_{b,c}, so v_{a,c} cannot be independent of v_{a,b} unless the intervening term is degenerate; Example 2.1's claim of disjoint sums is false for overlapping intervals. This is not a presentation issue but a fatal modeling inconsistency in the frequentist framework.
  2. [Section 3.3, Theorem 3.8 proof, Step (1)] The KL control between the algorithm's posterior and the virtual posterior is not proved. The proof invokes 'Inequality (51) in [24]' without stating it, but the cited Proposition 3.2 in this paper is derived for i.i.d. N(0,σ^2) virtual observations. In Theorem 3.8 the virtual posterior P_v has non-uniform variances bounded by 2σ^2T^{\tildeα}, and the KL bound, the DPP approximation tolerance, and the integration over the DPP distribution must be re-derived for this heteroscedastic covariance structure. Moreover, the algorithm needs the posterior-closeness uniformly at every round t≤T, not merely at the final T; the argument only states a bound for the final posterior, so the confidence-transfer step is not supported.
  3. [Section 2.2.2, Proposition 2.5 proof] The linear lower bound for α≥1 is not established. The proof claims that Model (1) yields at least Ω(T) regret 'as α1→1', but for every fixed α1<1 the stated bound is Ω(T^{(α1+1)/2})=o(T). The limiting argument does not provide a uniform constant for a fixed α1, and the coupling via Blackwell's theorem does not automatically transfer a regret lower bound from a less noisy experiment to a more noisy one without an additional argument linking the decision problem and the regret definition. The proof as written does not justify the Ω(T) conclusion.
  4. [Section 5, Theorem 5.2] The lower-bound construction in Theorem 5.2 does not provide a full temporal sequence (f_1,...,f_T) satisfying Assumptions 1 and 2. The adversary family F_M is defined only for the function at the final horizon T, and the argument then states that the same family can be 'repeated' for every horizon t≥1. However, the observation noise in Eq. (5.1) depends on the global T, and the dynamic regret at intermediate times is defined with respect to f_t, not f_T; a valid lower bound must specify how the functions at earlier times are coupled to the final f_T while satisfying Eq. (2.7). This gap is load-bearing for the claimed impossibility threshold N_T=O(T^{α/(α+1)}).
minor comments (4)
  1. [Section 3.3 heading] The heading says 'vanishing mean number of additional queries' but the body and formal statement use 'average number'; this should be made consistent. There are also typos such as 'intoduce' in Section 3 and 'the the bound' after the proof of Proposition 2.2.
  2. [Assumption 2 / Eq. (3.14)] Assumption 2 allows α=0, but the window construction in Eq. (3.14) and the query-complexity bound divide by α. The paper should either assume α>0 or treat the stationary case α=0 separately.
  3. [Figures 2 and 3] The text in Section 4 says the plots show cumulative regret, but the y-axes are labeled 'Average regret' (Figures 2 and 3). This should be clarified: is the plotted quantity R_T/T or cumulative regret?
  4. [Table 2] Table 2 is difficult to parse: the row/column structure mixes 'Upper bound of W-SparQ-GP-UCB' with 'Lower bounds' and 'Open question' entries in a way that could be made more explicit. A clearer separation of upper and lower bounds, and of regimes, would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: W-SparQ-GP-UCB's bounds follow from explicit model assumptions and standard/prior published ingredients, not from fitting the target result.

full rationale

I find no circular step in the paper's derivation chain. The central quantities are not fitted-then-renamed: α and ˜α are input model/algorithm parameters, and Theorem 3.8's regret rate is obtained by substituting the chosen window growth t_j^{tilde α/α}, the SE-kernel information gain bound γ_T=O(dT^{tilde α} log^{d+1} T), and the confidence parameter β_T into the standard GP-UCB regret inequality; none of these steps assumes the claimed no-regret conclusion. The pure-bandit upper bounds in Proposition 2.2 rest on a stated variance-proxy assumption v_t=O(t^α) and the external Lemma 2.3 / Makarova et al. analysis, while the lower bounds use external heteroscedastic lower bounds [15,27] and Fano's lemma. The most self-referential load-bearing ingredient is Proposition 3.2 and Inequality (51) from the authors' own prior work [24], used to control the KL divergence between the sparse and virtual posteriors; however, this is stated as a published theorem from prior work and is itself anchored in the external sparse-GP result of Burt et al. It is therefore a legitimate cited lemma rather than the conclusion of the paper restated by construction. I also note that Assumptions 1 and 2 may be mutually inconsistent for α>0 with bounded RKHS functions, and that the independence claim in Assumption 2 is algebraically problematic for overlapping intervals; these are substantive correctness/existence concerns, but they are not circularity, so they do not affect the circularity score.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central results are conditional on the UI model (Assumption 2) and on the sparse-DPP KL bound imported from the authors' prior paper [24]. No new physical or ontological entities are postulated; the 'expert' and 'virtual observations' are analysis constructs. Hand-set quantities are α, tilde_α, and the standard confidence parameter β_t.

free parameters (2)
  • α (temporal variability) = assumed known or upper-bounded; tuned on a training set in Section 4
    Controls the drift variance σ^2(t2−t1)^α in Assumption 2; all regime thresholds, window lengths, and regret rates depend on it.
  • tilde_α (window growth) = user-selected; set to 0.25 in Section 4
    Determines window length growth t^{tilde_α/α} and the query-vs-computation trade-off; must be <1/3 and, for the vanishing-rate proof, <α.
assumptions (4)
  • domain assumption Assumption 1: each f_t lies in the RKHS H_k with ||f_t||_k ≤ B
    Frequentist smoothness used throughout to apply GP-UCB confidence bounds (Eq. 2.6).
  • domain assumption Assumption 2: f_{t1}(x) = f_{t2}(x) + v_{t1,t2} with independent zero-mean sub-Gaussian v and variance σ^2(t2−t1)^α
    The uncertainty-injection model is the entire temporal-drift hypothesis; all regime thresholds and lower bounds follow from this variance growth (Eqs. 2.7–2.8).
  • standard math Proposition 3.2 (from [24]): a Q_T-DPP sparse subset approximates the full GP posterior in KL with Q_T=O(log^d(T/η))
    Imported without proof; load-bearing for the KL-control step in Theorem 3.8 and for SparQ-GP-UCB.
  • standard math Fano's lemma, the Scarlett–Bogunovic–Cevher adversary construction, and SE-kernel information-gain bounds
    Standard GP-bandit tools used for the lower bounds (Section 5) and for γ_T bounds (Proposition 2.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of No-Regret Gaussian Process Optimization of Time-Varying Functions." pith.science (2026). https://pith.science/paper/46UDHJFM

@misc{pith2026251200517,
  author       = {Pith},
  title        = {Pith review of: No-Regret Gaussian Process Optimization of Time-Varying Functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/46UDHJFM}},
  note         = {Machine review of arXiv:2512.00517}
}
read the original abstract

Sequential optimization of black-box functions from noisy evaluations has been widely studied, with Gaussian Process bandit algorithms such as GP-UCB guaranteeing no-regret in stationary settings. However, for time-varying objectives, no-regret is unattainable under pure bandit feedback unless strong and often unrealistic assumptions are imposed. We propose a novel method for optimizing time-varying rewards in the frequentist setting, where the objective has bounded RKHS norm almost surely. Time variations are captured through uncertainty injection, enabling heteroscedastic Gaussian process regression that adapts past observations to the current time step. As no-regret is unattainable in general in the strict bandit setting, we relax the latter allowing additional queries on previously observed points. Building on sparse inference and the effect of uncertainty injection on regret, we propose W-SparQ-GP-UCB, an online algorithm that achieves no-regret with a vanishing number of additional queries per iteration. To assess the theoretical limits of this approach, we establish a lower bound on the number of additional queries required for no-regret, proving the efficiency of our method. Finally, we provide a comprehensive analysis linking the temporal regime of the function to achievable regret rates, together with upper and lower bounds on the number of additional queries needed in each regime.

Figures

Figures reproduced from arXiv: 2512.00517 by the authors.

Figure 1
Figure 1. Visualization of the partition of {1, . . . , t}. The virtual posterior P v t obtained from (Xv t , Y v t ) will be used as a reference: we will show that the posterior P r t built from the algorithm’s regressors (Xr t , Y r t ) is close to P v t in KL divergence, provided the sparse set is chosen with sufficiently large Qt . Regret and query complexity. In this section, we analyze the regret of W-SparQ-GP￾UCB, and … view at source ↗
Figure 2
Figure 2. Average regret of GP-UCB variants in the time-varying setting. As predicted by the theoretical analysis, both SparQ-GP-UCB and W-SparQ-GP￾UCB achieve sublinear cumulative regret and outperform existing bandit algorithms. Al￾though SparQ-GP-UCB converges slightly faster, it requires significantly more additional queries (O(T logd T) against O(T −α/α ˜ logd T) for W-SparQ-GP-UCB). This illustrates the expected trade-o… view at source ↗
Figure 3
Figure 3. Average regret of GP-UCB variants on Berkeley Earth dataset. On this real-world dataset, although our methods outperform the other GP-UCB vari￾ants, they all appear to exhibit linear regret. While this may seem at odds with the theoretical results, several factors can explain this discrepancy. First and foremost, the true underlying function governing monthly temperature anomalies may not lie in the RKHS associated … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Squared prediction error of GP-UCB variants. Setting and notations. Fix a time horizon T ∈ N. We consider a Gaussian process model with a squared-exponential kernel k and observations following a time-varying noise model: (5.1) yt = ft(xt) + ϵt , ϵt ∼ N 0, σ2 (1 + (T −…
Figure 5
Figure 5. Figure 5: Schematic illustration of six adversary functions for [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 4 linked inside Pith

  1. [24]

    Mauduit, E

    E. Mauduit, E. Berthier, and A. Simonetto , Time-varying G aussian process bandit optimization with experts: no-regret in logarithmically-many side queries , in Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2025, pp. 164--182

    2025

  2. [1]

    Agarwal, D

    A. Agarwal, D. P. Foster, D. J. Hsu, S. M. Kakade, and A. Rakhlin , Stochastic convex optimization with bandit feedback , in Advances in Neural Information Processing Systems, vol. 24, 2011, pp. 213--240

    2011

  3. [2]

    https://berkeleyearth.org/high-resolution-data-access-page/, 2025

    Berkeley Earth , High-resolution europe temperature data . https://berkeleyearth.org/high-resolution-data-access-page/, 2025

    2025

  4. [3]

    Blackwell , Equivalent comparisons of experiments , The Annals of Mathematical Statistics, (1953), pp

    D. Blackwell , Equivalent comparisons of experiments , The Annals of Mathematical Statistics, (1953), pp. 265--272

    1953

  5. [4]

    Bogunovic, J

    I. Bogunovic, J. Scarlett, and V. Cevher , Time-varying G aussian process bandit optimization , in International Conference on Artificial Intelligence and Statistics, vol. 51, 2016, pp. 314--323

    2016

  6. [5]

    Brunzema, A

    P. Brunzema, A. Von Rohr, and S. Trimpe , On controller tuning with time-varying B ayesian optimization , in Conference on Decision and Control, 2022, pp. 4046--4052

    2022

  7. [6]

    Bubeck, N

    S. Bubeck, N. Cesa-Bianchi, et al. , Regret analysis of stochastic and nonstochastic multi-armed bandit problems , Foundations and Trends in Machine Learning, 5 (2012), pp. 1--122

    2012

  8. [7]

    D. R. Burt, C. E. Rasmussen, and M. van der Wilk , Convergence of sparse variational inference in G aussian processes regression , Journal of Machine Learning Research, 21 (2020), pp. 1--63

    2020

Show all 33 references
  1. [8]

    Charpentier, R

    A. Charpentier, R. Elie, and C. Remlinger , Reinforcement learning in economics and finance , Computational Economics, 62 (2023), pp. 425--462

    2023

  2. [9]

    W. C. Cheung, D. Simchi-Levi, and R. Zhu , Learning to optimize under non-stationarity , in The 22nd International Conference on Artificial Intelligence and Statistics, 2019, pp. 1079--1087

    2019

  3. [10]

    V. Dani, T. P. Hayes, and S. M. Kakade , Stochastic linear optimization under bandit feedback , in 21st Annual Conference on Learning Theory, no. 101, 2008, pp. 355--366

    2008

  4. [11]

    Y. Deng, X. Zhou, B. Kim, A. Tewari, A. Gupta, and N. B. Shroff , Weighted G aussian process bandits for non-stationary environments , in International Conference on Artificial Intelligence and Statistics, vol. 151, 2022, pp. 6909--6932

    2022

  5. [12]

    Djolonga, A

    J. Djolonga, A. Krause, and V. Cevher , High-dimensional G aussian process bandits , Advances in Neural Information Processing Systems, 26 (2013), pp. 1025--1033

    2013

  6. [13]

    Drineas, M

    P. Drineas, M. W. Mahoney, and N. Cristianini , On the N ystr \"o m method for approximating a G ram matrix for improved kernel-based learning. , Journal of Machine Learning Research, 6 (2005)

    2005

  7. [14]

    Imamura, N

    H. Imamura, N. Charoenphakdee, F. Futami, I. Sato, J. Honda, and M. Sugiyama , Time-varying G aussian process bandit optimization with non-constant evaluation time , arXiv preprint arXiv:2003.04691, (2020)

    2003 arXiv

  8. [15]

    Iwazaki and S

    S. Iwazaki and S. Takeno , Improved regret analysis in G aussian P rocess bandits: optimality for noiseless reward, R K H S norm, and non-stationary variance , arXiv preprint arXiv:2502.06363, (2025)

    2025 arXiv

  9. [16]

    Kirschner and A

    J. Kirschner and A. Krause , Information directed sampling and bandits with heteroscedastic noise , in Conference On Learning Theory, 2018, pp. 358--384

    2018

  10. [17]

    Krause and C

    A. Krause and C. Ong , Contextual G aussian process bandit optimization , Advances in Neural Information Processing Systems, 24 (2011), pp. 2447--2455

    2011

  11. [18]

    Kulesza , Determinantal point processes for machine learning , Foundations and Trends in Machine Learning , 5 (2012), pp

    A. Kulesza , Determinantal point processes for machine learning , Foundations and Trends in Machine Learning , 5 (2012), pp. 123--286

    2012

  12. [19]

    Kulesza and B

    A. Kulesza and B. Taskar , k - DPP s: Fixed-size determinantal point processes , in Proceedings of the 28th International Conference on Machine Learning (ICML-11), 2011, pp. 1193--1200

    2011

  13. [20]

    Le Cam , On the information contained in additional observations , The Annals of Statistics, 2 (1974), pp

    L. Le Cam , On the information contained in additional observations , The Annals of Statistics, 2 (1974), pp. 630--649

    1974

  14. [21]

    Leibfried, V

    F. Leibfried, V. Dutordoir, S. John, and N. Durrande , A tutorial on sparse G aussian processes and variational inference. arxiv 2020 , arXiv preprint arXiv:2012.13962, (2012)

    2020 arXiv

  15. [22]

    Makarova, I

    A. Makarova, I. Usmanova, I. Bogunovic, and A. Krause , Risk-averse heteroscedastic B ayesian optimization , in Advances in Neural Information Processing Systems , vol. 34, 2021, pp. 17235--17245

    2021

  16. [23]

    J. Mary, R. Gaudel, and P. Preux , Bandits and recommender systems , in International Workshop on Machine Learning, Optimization and Big Data, vol. 9432, 2015, pp. 325--336

    2015

  17. [25]

    C. E. Rasmussen and C. K. I. Williams , Gaussian processes for machine learning , The MIT Press, 2006

    2006

  18. [26]

    V. Roy, A. Simonetto, and G. Leus , Spatio-temporal sensor management for environmental field estimation , Signal Processing, 128 (2016), pp. 369--381

    2016

  19. [27]

    Scarlett, I

    J. Scarlett, I. Bogunovic, and V. Cevher , Lower bounds on regret for noisy G aussian process bandit optimization , in Conference on Learning Theory, 2017, pp. 1723--1742

    2017

  20. [28]

    Scarlett and V

    J. Scarlett and V. Cevher , An introductory guide to F ano's inequality with applications in statistical estimation , arXiv preprint arXiv:1901.00555, (2019)

    1901 arXiv

  21. [29]

    Scholkopf, K

    B. Scholkopf, K. Tsuda, and J.-P. Vert , Kernel Methods in Computational Biology , The MIT Press, 2004

    2004

  22. [30]

    Slivkins et al

    A. Slivkins et al. , Introduction to multi-armed bandits , Foundations and Trends in Machine Learning, 12 (2019), pp. 1--286

    2019

  23. [31]

    Srinivas, A

    N. Srinivas, A. Krause, S. M. Kakade, and M. W. Seeger , Information-theoretic regret bounds for G aussian process optimization in the bandit setting , IEEE Transactions on Information Theory, 58 (2012), pp. 3250–--3265

    2012

  24. [32]

    Zhou and N

    X. Zhou and N. Shroff , No-regret algorithms for time-varying B ayesian optimization , in Annual Conference on Information Sciences and Systems, 2021, pp. 1--6

    2021

  25. [33]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTION or pop #1 'skip if FUNCTION new.block.check...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.