High-Dimensional Safe Exploration via Optimistic Local Latent Safe Optimization

Yunyue Wei, Zeji Yi, Hongda Li, Saraswati Soedarmadji, Yanan Sui

Primary Area: optimization

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Safe Bayesian optimization, High-dimensional optimization, Sequential online optimization

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose the Optimistic Local Latent Safe Optimization method for online safe optimization over high-dimensional spaces, which efficiently optimizes the high-dimensional function while enjoying theoretical probabilistic safety guarantee.

Abstract:

Optimization over high-dimensional input space is inherently difficult, especially when safety needs to be maintained during sampling. Current safe exploration algorithms ensure safety by conservatively expanding the safe region, which leads to inefficiency in large input settings. Existing high-dimensional constrained optimization methods also neglect safety in the search process. In this paper, we propose Optimistic Local Latent Safe Optimization (OLLSO), which is capable of handling high-dimensional problems under probabilistic safety satisfaction. We first use distance-preserved autoencoder to transform the original input space into a low-dimensional continuous latent space. An optimistic local safe strategy is then applied over the latent space to efficiently optimize the utility function.
Theoretically, we prove the probabilistic safety guarantee from the latent space to the original space.
OLLSO outperforms representative high-dimensional constrained optimization algorithms in simulation experiments. We also show its real application in clinical experiments for safe and efficient online optimization of a neuromodulation therapy.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3774