Convergence-Aware Multi-Fidelity Bayesian Optimization

26 Sept 2024 (modified: 05 Feb 2025)Submitted to ICLR 2025EveryoneRevisionsBibTeXCC BY 4.0
Keywords: Bayesian Optimization, Multi-Fidelity Bayesian Optimization, Gaussian process, dynamic systems
TL;DR: This work brings SOTA multi-fidelity Bayesian Optimization to the next level by proposing a new framework to design fidelity models that respect convergence behavior wrt fidelity.
Abstract:

Multi-fidelity Bayesian Optimization (MFBO) has emerged as a powerful approach for optimizing expensive black-box functions by leveraging evaluations at different fidelity levels. However, existing MFBO methods often overlook the convergence behavior of the objective function as fidelity increases, leading to inefficient exploration and suboptimal performance. We propose CAMO, a novel Convergence-Aware Multi-fidelity Optimization framework based on Fidelity Differential Equations (FiDEs). CAMO explicitly captures the convergence behavior of the objective function, enabling more efficient optimization. We introduce two tractable forms of CAMO: an integral Automatic Relevance Determination (ARD) kernel and a data-driven Deep Kernel. Theoretical analysis demonstrates that CAMO with the integral ARD kernel achieves a tighter regret bound compared to state-of-the-art methods. Our empirical evaluation on synthetic benchmarks and real-world engineering design problems shows that CAMO consistently outperforms existing MFBO algorithms in optimization efficiency and solution quality, with up to 4x improvement in optimal solution. This work establishes a foundation for tractable convergence-aware MFBO and opens up new avenues for research in this area.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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: 8025
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview