REVIEW 2 major objections 3 minor 22 references
Safety in safe Bayesian optimization and its ramifications for control
T0 review · 2 major / 3 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read This paper argues that SafeOpt-type safe Bayesian optimization algorithms cannot currently be guaranteed safe in practice, and proposes LoSBO, whose safety rests on a known Lipschitz bound and bounded noise.
desk verdict A credible critique of SafeOpt-style safety and a sensible Lipschitz-based alternative, but this extended abstract is a summary of the journal paper and the Lipschitz assumption is shaky for the motivating control problems. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central mechanism is the Lipschitz cone safety certificate. From a safe observation $(x_i, y_i)$ with noise bounded by $\varepsilon$ and a target function $f_*$ known to be $L$-Lipschitz, every point $x$ satisfying $y_i - \varepsilon - L\|x - x_i\| \ge h$ is guaranteed safe, because $f_*(x) \ge f_*(x_i) - L\|x - x_i\| \ge y_i - \varepsilon - L\|x - x_i\|$. LoSBO uses the union of such certified sets as its safety mechanism, leaving the Gaussian process mean and variance to guide exploration, so the scaling factors $\beta_t$ are ordinary hyperparameters rather than safety-critical quantities. LoS-GP-UCB keeps the same certificate but optimizes the acquisition function with local search and random restarts, removing the grid that limits SafeOpt-type algorithms to low dimensions.
What would settle it
Run Real-$\beta$-SafeOpt with a correctly evaluated rigorous bound on a benchmark where the target function's true RKHS norm is computable and the only prior knowledge is of the kind an engineer would normally have; a reliable bracketing of the true norm from that prior knowledge would undercut the paper's claim that such bounds are presently unobtainable. Separately, for LoSBO, take any $L$-Lipschitz function with noise bounded by $\varepsilon$ and check every queried point: with valid constants no $f_*(x_t) < h$ should occur, so one such violation would refute the algorithm's safety guarantee.
Extended reading notes
Core claim
The paper's central claim is that SafeOpt-type safe Bayesian optimization cannot currently be made safe in practice. The first obstacle is that the high-probability confidence bounds on which SafeOpt rests are hard to evaluate, and all implementations known to the authors replace them with heuristics such as $\beta_t = 2$; numerical experiments show that these heuristic bounds are violated even in relatively benign settings. The second obstacle is that the theoretically valid bounds require a known upper bound on the reproducing kernel Hilbert space norm of the target function, a function-space smoothness measure that the paper argues cannot be reliably bounded from established engineering prior knowledge; an invalid bound produces safety violations. The proposed remedy, LoSBO, decouples exploration from safety: a Gaussian process model drives optimization, while safety is enforced by a deterministic lower bound built from a known Lipschitz constant and noise bound. The paper concludes that as long as safety relies on a quantitative RKHS norm bound, SafeOpt-type safety cannot be ensured in practice, whereas Lipschitz-only safety rests on assumptions with clear interpretations in control applications.
Load-bearing premise
LoSBO's safety guarantee collapses if the practitioner does not actually have a valid Lipschitz constant and noise bound for the real plant; the paper itself acknowledges that applicability hinges on these assumptions.
Editorial extensions
If this is right
- SafeOpt-type algorithms should not be treated as safe for online controller tuning unless a valid upper bound on the RKHS norm of the target function is available; heuristic choices like $\beta_t = 2$ forfeit the theoretical guarantee.
- Safety and exploration can be decoupled: the Gaussian process model can drive optimization while a Lipschitz-based lower bound enforces safety, turning the GP scaling factors into tunable hyperparameters.
- For continuous parameter spaces in moderate dimensions, LoS-GP-UCB offers a gridding-free route to safe optimization, which matters for controller tuning in practice.
- If the critique is correct, published results reporting safe behavior of SafeOpt-type algorithms with heuristic bounds should be re-examined, since their safety was not guaranteed by the algorithm's theory.
- A practical precondition for safe online tuning is that a valid Lipschitz constant and noise bound can be identified from physical knowledge or high-fidelity simulation.
Reading between the lines
- The paper leaves implicit that the same Lipschitz-cone certificate could replace the confidence-bound safety mechanism in other safe-exploration schemes, such as stagewise or global-optimization variants, potentially giving them practical safety without any RKHS norm bound.
- An obvious next question, not addressed here, is how to maintain a valid Lipschitz bound and noise bound over time; any scheme that adapts $L$ or $\varepsilon$ from data would have to keep the bounds conservative enough to preserve the guarantee.
- The critique suggests a concrete audit for earlier safe-BO applications: if the implementation used a constant $\beta_t$ instead of a rigorously evaluated bound, reported safe behavior should not be credited to the algorithm's theoretical safety guarantee.
- In high-dimensional control problems a single global Lipschitz constant may be conservative, so the practical reach of LoSBO is likely to depend on local or dimension-aware Lipschitz bounds; that is a testable extension, not a result of the paper.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper argues that SafeOpt-type safe Bayesian optimization algorithms have two significant practical obstacles: (i) typical implementations replace theoretically justified uncertainty bounds by heuristics that are not guaranteed to be valid, and (ii) the theoretically valid uncertainty bounds require an upper bound on the RKHS norm of the target function, which the authors claim cannot be reliably obtained from standard engineering prior knowledge. The authors propose Lipschitz-only Safe Bayesian Optimization (LoSBO), which uses only a known Lipschitz bound and bounded noise for safety, and a grid-free variant LoS-GP-UCB for higher-dimensional problems. The manuscript is explicitly an extended abstract that disseminates results from the authors' journal paper [7], and it states that all plots and results are taken verbatim from [7].
Significance. If the central claims are correct, the paper identifies a real and important gap in the practical applicability of SafeOpt-type algorithms and offers a conceptually simple alternative with more interpretable assumptions. The critique of heuristic uncertainty bounds and the RKHS-norm requirement is timely and could influence practice in learning-based control. However, the manuscript's contribution is largely a summary of prior work, and the practical value of LoSBO depends crucially on whether the required Lipschitz and noise bounds can be supplied in the very applications used to motivate the method.
major comments (2)
- [Section 3 (LoSBO)] The safety certificate of LoSBO, ℓ(x) = max_i(y_i − ε − L‖x − x_i‖), requires the target function to be L-Lipschitz continuous on the entire search space D. The paper motivates the need for safe BO with control applications such as LQR tuning, where the performance measure is typically unbounded near the stability boundary. A function that is unbounded on D is not Lipschitz continuous on D, and no finite L exists. The paper acknowledges that 'the applicability of the proposed algorithm clearly hinges on these assumptions,' but it does not address the mismatch between the motivating applications and the existence of a finite, usable global L. This is load-bearing because the central claim is that LoSBO overcomes the practical safety issues of SafeOpt-type algorithms; for the examples in Section 1, a finite L may not exist and LoSBO is not applicable as stated. The authors should clarify the domain over which L is assumed, discuss whether the safe sublevel set can be chosen to avoid non-Lipschitz regions, and provide guidance on how a practitioner can verify L.
- [Title page / preface] The manuscript explicitly states that it is an extended abstract that disseminates results from the journal paper [7] and that all plots and results are taken verbatim from [7]. As a submission to a serious journal, the paper does not contain the derivations, pseudocode, or experimental protocols required to independently verify the central claims. In particular, the claim in Section 2 that 'as long as the safety of the algorithm relies on a quantitative upper bound on the RKHS norm of the target function, it is at present not possible to ensure that SafeOpt-type algorithms are safe in practice' is not substantiated in this manuscript; the supporting numerical experiments are only referenced, not presented. This lack of self-containedness prevents a full assessment of correctness and of the relationship between the stated Lipschitz assumptions and the algorithm's behavior. The authors should include the necessary technical content or restructure the submission to make clear that it is a research summary rather than a full journal paper.
minor comments (3)
- [Section 3] The sentence 'they have to be judge on a case-by-case base by practitioners' contains a typo; it should read 'they have to be judged on a case-by-case basis by practitioners.'
- [Figure 2 caption] The caption uses 'LosBO' while the text uses 'LoSBO'; the spelling should be made consistent.
- [Section 3 / Figure 2] The text refers to 'the constant E' in the caption of Figure 2, but E is never defined in the body. Please define the notation or remove it.
Circularity Check
No circularity: the safety certificate is derived from stated external Lipschitz and noise bounds, not from the conclusion it supports.
full rationale
The paper's central derivation chain is self-contained in the required sense. The critique of SafeOpt-type algorithms is conditional: safety guarantees are shown to depend on a quantitative RKHS norm bound, and the paper argues that such a bound is not reliably available from engineering knowledge. That argument is an external-assumption concern, not a self-referential loop. The proposed LoSBO safety mechanism is built explicitly from a known Lipschitz constant and a bounded-noise assumption, yielding a conservative lower bound from observed samples; this is a standard set-membership inference and is not equivalent to the conclusion that LoSBO is practical. The paper transparently states that all plots and results are taken verbatim from the authors' earlier journal paper [7], and it cites its own prior bound [8] as a mathematical tool; these are separate published results with stated assumptions and are used as evidence, not as a definitional substitute for the present conclusion. The acknowledged dependence on the validity of the Lipschitz and noise bounds is a limitation of the method's applicability, not a circular step. No equation is observed to reduce to its own output, and no fitted parameter is renamed as a prediction. Therefore the circularity score is 0.
Assumptions & free parameters
free parameters (1)
- beta_t scaling factors for LoS-GP-UCB acquisition =
user-chosen, not specified
assumptions (4)
- domain assumption A valid Lipschitz constant L for the unknown target function is known to the user.
- domain assumption Noise is bounded by a known constant (e.g., |eta_t| <= epsilon).
- domain assumption A nonempty initial safe set S0 is known and the target function satisfies the assumed regularity.
- domain assumption The RKHS norm of the target function cannot be reliably upper-bounded from typical engineering prior knowledge.
Cite this review
Pith. "Pith review of Safety in safe Bayesian optimization and its ramifications for control." pith.science (2026). https://pith.science/paper/C5ESI5RW
@misc{pith2026250113697,
author = {Pith},
title = {Pith review of: Safety in safe Bayesian optimization and its ramifications for control},
year = {2026},
howpublished = {\url{https://pith.science/paper/C5ESI5RW}},
note = {Machine review of arXiv:2501.13697}
}
read the original abstract
A recurring and important task in control engineering is parameter tuning under constraints, which conceptually amounts to optimization of a blackbox function accessible only through noisy evaluations. For example, in control practice parameters of a pre-designed controller are often tuned online in feedback with a plant, and only safe parameter values should be tried, avoiding for example instability. Recently, machine learning methods have been deployed for this important problem, in particular, Bayesian optimization (BO). To handle safety constraints, algorithms from safe BO have been utilized, especially SafeOpt-type algorithms, which enjoy considerable popularity in learning-based control, robotics, and adjacent fields. However, we identify two significant obstacles to practical safety. First, SafeOpt-type algorithms rely on quantitative uncertainty bounds, and most implementations replace these by theoretically unsupported heuristics. Second, the theoretically valid uncertainty bounds crucially depend on a quantity - the reproducing kernel Hilbert space norm of the target function - that at present is impossible to reliably bound using established prior engineering knowledge. By careful numerical experiments we show that these issues can indeed cause safety violations. To overcome these problems, we propose Lipschitz-only Safe Bayesian Optimization (LoSBO), a safe BO algorithm that relies only on a known Lipschitz bound for its safety. Furthermore, we propose a variant (LoS-GP-UCB) that avoids gridding of the search space and is therefore applicable even for moderately high-dimensional problems.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[7]
arXiv preprint arXiv:2403.12948 (2024)
Fiedler, C., Menn, J., Kreisk¨ other, L., Trimpe, S.: On safety in safe Bayesian optimization. arXiv preprint arXiv:2403.12948 (2024)
arXiv 2024
-
[1]
Doctoral thesis, University of Alberta (2013)
Abbasi-Yadkori, Y.: Online learning for linearly parametrized control problems. Doctoral thesis, University of Alberta (2013)
work page 2013
-
[2]
Advances in neural information processing systems 33, 21,524–21,538 (2020)
Balandat, M., Karrer, B., Jiang, D., Daulton, S., Letham, B., Wilson, A.G., Bak- shy, E.: BoTorch: A framework for efficient Monte-Carlo Bayesian optimization. Advances in neural information processing systems 33, 21,524–21,538 (2020)
work page 2020
-
[3]
In: IEEE International Conference on Robotics and Automation (ICRA), pp
Baumann, D., Marco, A., Turchetta, M., Trimpe, S.: Gosafe: Globally optimal safe robot learning. In: IEEE International Conference on Robotics and Automation (ICRA), pp. 4452–4458. IEEE (2021)
work page 2021
-
[4]
Machine Learning 112(10), 3713–3747 (2023) 8 Fiedler, Menn, Trimpe
Berkenkamp, F., Krause, A., Schoellig, A.P.: Bayesian optimization with safety constraints: safe and automatic parameter tuning in robotics. Machine Learning 112(10), 3713–3747 (2023) 8 Fiedler, Menn, Trimpe
work page 2023
-
[5]
In: IEEE international conference on robotics and automation (ICRA), pp
Berkenkamp, F., Schoellig, A.P., Krause, A.: Safe controller optimization for quadrotors with gaussian processes. In: IEEE international conference on robotics and automation (ICRA), pp. 491–496. IEEE (2016)
work page 2016
-
[6]
34th Interna- tional Conference on Machine Learning (ICML) 2, 1397–1422 (2017)
Chowdhury, S.R., Gopalan, A.: On kernelized multi-armed bandits. 34th Interna- tional Conference on Machine Learning (ICML) 2, 1397–1422 (2017)
work page 2017
-
[8]
Proceedings of the AAAI conference on artificial intelligence 35 (2021)
Fiedler, C., Scherer, C.W., Trimpe, S.: Practical and rigorous uncertainty bounds for Gaussian process regression. Proceedings of the AAAI conference on artificial intelligence 35 (2021)
work page 2021
Show all 22 references
-
[9]
Cambridge University Press (2023)
Garnett, R.: Bayesian optimization. Cambridge University Press (2023)
2023
-
[10]
In: International Workshop on the Foundations of Trustworthy AI Integrating Learning, Optimization and Reasoning, pp
Kim, Y., Allmendinger, R., L´ opez-Ib´ a˜ nez, M.: Safe learning and optimization tech- niques: Towards a survey of the state of the art. In: International Workshop on the Foundations of Trustworthy AI Integrating Learning, Optimization and Reasoning, pp. 123–139. Springer (2020)
2020
-
[11]
IEEE Robotics and Automation Letters 6(2), 1439–1446 (2021)
Marco, A., Baumann, D., Khadiv, M., Hennig, P., Righetti, L., Trimpe, S.: Robot learning with crash constraints. IEEE Robotics and Automation Letters 6(2), 1439–1446 (2021)
2021
-
[12]
In: 2016 IEEE international con- ference on robotics and automation (ICRA), pp
Marco, A., Hennig, P., Bohg, J., Schaal, S., Trimpe, S.: Automatic LQR tuning based on Gaussian process global optimization. In: 2016 IEEE international con- ference on robotics and automation (ICRA), pp. 270–277. IEEE (2016)
2016
-
[13]
Submitted for publication to 63rd IEEE Conference on Decision and Control
Menn, J., Pelizzari, P., Fleps-Dezasse, M., Trimpe, S.: Lipschitz safe Bayesian optimization for automotive control. Submitted for publication to 63rd IEEE Conference on Decision and Control
-
[14]
Au- tomatica 40(6), 957–975 (2004)
Milanese, M., Novara, C.: Set membership identification of nonlinear systems. Au- tomatica 40(6), 957–975 (2004)
2004
-
[15]
IEEE Transactions on Control Systems Technology 28(3), 730–740 (2019)
Neumann-Brosig, M., Marco, A., Schwarzmann, D., Trimpe, S.: Data-efficient auto- tuning with Bayesian optimization: An industrial control study. IEEE Transactions on Control Systems Technology 28(3), 730–740 (2019)
2019
-
[16]
Adaptive Computation and Machine Learning
Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. Adaptive Computation and Machine Learning. MIT Press, Cambridge, MA, USA (2006)
2006
-
[17]
at-Automatisierungstechnik 72(4), 281–292 (2024)
von Rohr, A., Stenger, D., Scheurenberg, D., Trimpe, S.: Local Bayesian opti- mization for controller tuning with crash constraints. at-Automatisierungstechnik 72(4), 281–292 (2024)
2024
-
[18]
Proceedings of the IEEE 104(1), 148–175 (2016)
Shahriari, B., Swersky, K., Wang, Z., Adams, R.P., De Freitas, N.: Taking the human out of the loop: A review of Bayesian optimization. Proceedings of the IEEE 104(1), 148–175 (2016)
2016
-
[19]
27th International Confer- ence on Machine Learning pp
Srinivas, N., Krause, A., Kakade, S., Seeger, M.: Gaussian process optimization in the bandit setting: No regret and experimental design. 27th International Confer- ence on Machine Learning pp. 1015–1022 (2010)
2010
-
[20]
Springer Science & Busi- ness Media (2008)
Steinwart, I., Christmann, A.: Support vector machines. Springer Science & Busi- ness Media (2008)
2008
-
[21]
In: International conference on machine learning, pp
Sui, Y., Gotovos, A., Burdick, J., Krause, A.: Safe exploration for optimization with Gaussian processes. In: International conference on machine learning, pp. 997–1005. PMLR (2015)
2015
-
[22]
In: International conference on machine learning, pp
Sui, Y., Zhuang, V., Burdick, J., Yue, Y.: Stagewise safe Bayesian optimization with Gaussian processes. In: International conference on machine learning, pp. 4781–4789. PMLR (2018)
2018
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.