Safe Time-Varying Optimization based on Gaussian Processes with Spatio-Temporal Kernel

Ensuring safety is a key aspect in sequential decision making problems, such as robotics or process control. The complexity of the underlying systems often makes finding the optimal decision challenging, especially when the safety-critical system is time-varying. Overcoming the problem of optimizing an unknown time-varying reward subject to unknown time-varying safety constraints, we propose TVSafeOpt, a new algorithm built on Bayesian optimization with a spatio-temporal kernel. The algorithm is capable of safely tracking a time-varying safe region without the need for explicit change detection. Optimality guarantees are also provided for the algorithm when the optimization problem becomes stationary. We show that TVSafeOpt compares favorably against SafeOpt on synthetic data, both regarding safety and optimality. Evaluation on a realistic case study with gas compressors confirms that TVSafeOpt ensures safety when solving time-varying optimization problems with unknown reward and safety functions.

Paper

References (48)

Scroll for more · 36 remaining

Similar papers

Peer review

Reviewer yo4J6/10 · confidence 3/52024-07-08

Summary

The authors propose a time-varying extension of SAFEOPT to overcome the problems of time-varying rewards under time-varying safety constraints. Under stationarity conditions, optimality guarantees are provided and the numerical simluation shows a (favorable) comparison to the SAFEOPT.

Strengths

1. The paper is very well written and easy to follow. 2. Based on related work, the problems of time-varying rewards under time-varying safety constraints are an open problem in literature, and his paper addresses that. 3. The paper provides formal safety guarantees for their TVSAFEOPT algorithm.

Weaknesses

1. *Some delineation to related work seems rather vague and requires stronger justification.* An example for TVSBO: the time-variable and temporal aspect of the kernel can just as well be interpreted as context using existing results. Perhaps a table would help here to highlight key aspects. 2. *Lack of real-world data experiments and comparison to related work.* To support the downsides of existing approaches, an empirical comparison to existing TVSBO approaches mentioned in the related work section would be needed. 3. The *empirical results could be more convincing* by adding a variety of initial safe sets and revised plots. The current plots/results are hard to parse. 4. It would be beneficial if the *theoretical/technical challenge of extending safety to the time-varying case were more detailed*. This would streamline the presentation and help in assessing the impact of the contribution.

Questions

1. *In the Appendix, a spatio-temporal SE kernel is introduced. How is this construction different from using an SE-ARD kernel with a composite variable $z = [x^T,t]^T$?* If I am not mistaken, for the SE-ARD kernel there would be no different than defining a single kernel with $z$. 2. It is mentioned that the Lipschitz constants are to be known beforehand. However, while commonly assumed, *how do you get a hold of an RKHS norm bound $B$ (related to Assumption 2.1) to compute the UCB?* 3. *Could you provide Figure 1 sooner in the manuscript?* It would be super helpful to see this central illustration already on page 2.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

1. Practicality of the safety guarantee: requiring many Lipschitz constants for both for space and time while also requiring an RKHS norm bound. 2. Theoretical and empirical impact: Lack of comparisons to TVSBO approaches makes the implications of the contribution unclear both theoretically and empirically.

Reviewer g7A95/10 · confidence 3/52024-07-12

Summary

This paper presents a safe Bayesian optimization algorithm TVSAFEOPT with a spatial-temporal kernel and time Lipschitz constants, which improves on SAFEOPT with time-varying reward and safety constraints. The optimality guarantee is proved for the stationary case and the safety guarantee for more general settings. The method is tested on a synthetic problem and gas compressors.

Strengths

1. The use of a spatio-temporal kernel in Bayesian optimization for time-varying safety constraints is novel. 2. A formal proof of safety and optimality guarantee under certain assumptions.

Weaknesses

1. More discussion on how to make a tradeoff between optimality and safety is encouraged. 2. Will this conservatism in safety become too large in high-dimensional problems? 2. The method to choose the proper initial safe set and kernel parameters is unclear.

Questions

1. How to find an initial safe set for complex problems? 2. How to find the kernel parameter for each task? 2. What is the computational complexity compared to other BO baselines?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The societal impact is discussed.

Reviewer o5DK5/10 · confidence 2/52024-07-13

Summary

The paper introduces the TVSAFEOPT algorithm, which is based on Gaussian processes with spatio-temporal kernels, designed specifically for optimizing time-varying rewards under time-varying safety constraints. The algorithm provides formal safety guarantees in a general time-varying setting, ensuring safety even when exploring non-stationary safe regions. It robustly subtracts safety margins to prevent unsafe decisions, adapting in real-time to changing environments. Furthermore, they provide optimality guarantees for locally stationary optimization problems, ensuring near-optimal solutions when the optimization problem becomes stationary.

Strengths

They provide formal safety guarantees in dynamic environments, ensuring safe decision-making even in non-stationary settings. Additionally, the algorithm offers optimality guarantees for stationary optimization problems, enhancing its reliability and performance Extensive numerical simulations were provided to validate the proposed approach.

Weaknesses

They extend the Safeopt algorithm from literature. However, it is clear on what are the additional contributions and difference between these two different approaches.

Questions

-

Rating

5

Confidence

2

Soundness

2

Presentation

2

Contribution

2

Limitations

-

Authorsrebuttal2024-08-07

Dear Program chairs, It looks like our uploaded pdf and general answer is not visible to the reviewers. Could this be the case, and is it possible to make it visible? I also saw this in other rebuttals which I am reviewing where there is a reference to a pdf but no pdf uploaded. Thank you for your help.

Reviewer yo4J2024-08-11

Reviewer Reply for Submission20789 by Reviewer yo4J

I thank the authors for the rebuttal and the answers to my questions and concerns, and also appreciate them including a table for related work. Since I was already on the positive side in my initial review, so I prefer to keep my overall score with an increases in parts of evaluation.

Reviewer g7A92024-08-11

Thank you for the rebuttal. Based on the responses I am comfortable with my current score for this paper.

Reviewer o5DK2024-08-13

Response to authors

Thanks for the clarifications. I would like to maintain my score

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC