Adaptive hyper-prior: CHENKernelFactory

[emiliencgm]

DevPR, parent is #745

Add an adaptive hyper-prior for GP proposed in the paper:
Guanming Chen, Maximilian Fleck, Thijs Stuyver. Leveraging Hidden-Space Representations Effectively in Bayesian Optimization for Experiment Design through Dimension-Aware Hyperpriors. ChemRxiv. 09 February 2026.
DOI: https://doi.org/10.26434/chemrxiv.10001986/v2

[CLAassistant]

All committers have signed the CLA.

[AdrianSosic]

Hi @emiliencgm, thanks for the commit ⭐ I'll take care of finalizing the PR once the necessary preparations in my dev branch are completed. Will ping you, @maxfleck and @tstuyver here once it's ready for you to review 🙃

[emiliencgm]

[heart] Emilien CHEN reacted to your message:

…

________________________________ From: AdrianSosic ***@***.***> Sent: Monday, March 2, 2026 8:56:08 AM To: emdgroup/baybe ***@***.***> Cc: Emilien CHEN ***@***.***>; Mention ***@***.***> Subject: Re: [emdgroup/baybe] Adaptive hyper-prior: CHENKernelFactory (PR #752) [https://avatars.githubusercontent.com/u/23265127?s=20&v=4]AdrianSosic left a comment (emdgroup/baybe#752)<#752 (comment)> Hi @emiliencgm<https://github.com/emiliencgm>, thanks for the commit ⭐ I'll take care of finalizing the PR once the necessary preparations in my dev branch are completed. Will ping you, @maxfleck<https://github.com/maxfleck> and @tstuyver<https://github.com/tstuyver> here once it's ready for you to review 🙃 — Reply to this email directly, view it on GitHub<#752 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/APPI5XUGQX5QT43BMTKE2RD4OVECRAVCNFSM6AAAAACWAXNCCGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSOBTGAYTMOJUHE>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[AdrianSosic]

Hi @emiliencgm, thanks for the commit ⭐ I'll take care of finalizing the PR once the necessary preparations in my dev branch are completed. Will ping you, @maxfleck and @tstuyver here once it's ready for you to review 🙃

Will have to wait until I'm back from holidays 😬

[Copilot]

Pull request overview

Adds a new Gaussian process preset implementing the adaptive kernel hyper-prior from Chen et al. (2026), and updates documentation/citations to support the new preset and reference handling.

Changes:

• Introduces CHEN GP preset and CHENKernelFactory implementing the adaptive hyper-prior kernel configuration.
• Adds/updates bibliographic references and relocates the rendered bibliography to a dedicated docs page.
• Updates existing EDBO preset docstrings to cite Shields et al. (2021) and updates changelog/contributors accordingly.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
docs/userguide/transfer_learning.md	Removes per-page bibliography block (bibliography now centralized).
docs/references.bib	Adds BibTeX entries for Chen (2026) and Shields (2021).
docs/misc/references.md	New references page rendering the bibliography.
docs/index.md	Adds “References” to the docs navigation.
baybe/surrogates/gaussian_process/presets/edbo_smoothed.py	Updates citations in docstrings (and adjusts docstring formatting).
baybe/surrogates/gaussian_process/presets/edbo.py	Updates citations and reference text in docstrings.
baybe/surrogates/gaussian_process/presets/core.py	Adds CHEN to the GaussianProcessPreset enum with citation.
baybe/surrogates/gaussian_process/presets/chen.py	New CHEN kernel factory implementing the adaptive hyper-prior.
baybe/surrogates/gaussian_process/presets/init.py	Exposes CHENKernelFactory from the presets package.
CONTRIBUTORS.md	Adds contributors for the Chen hyper-prior work.
CHANGELOG.md	Notes availability of the new CHEN preset.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[AdrianSosic]

@emiliencgm @tstuyver @maxfleck

It took a little longer for me to finalize the PR because it turned out there were quite some architectural changes and decisions to be made first. If you are interested, you can follow up on the latest developments (and how your PR fits into these) here: #745

Note: what will probably still change in your code is the way how TaskParameters are handled, but I'll take care of this in a follow-up PR – the current version of the code is good for merge from my perspective. Right now, TaskParameters are circumvented using a selection mechanism, but I'm planning to add some machinery that allows us to outsource their handling to other BayBE default mechanisms. But this is just FYI.

Again: I'm glad that your work eventually finds its way into our package! I've also added a recent simulation result to the PR description which shows the immediate benefits! ⭐

[AdrianSosic]

@emiliencgm @tstuyver @maxfleck

I've sent a collaboration invite for this repo to you. Once you've accepted, I can add you as reviewers

[AVHopp]

Thanks for the contribution :)

[AdrianSosic]

@maxfleck: you are now a reviewer
@emiliencgm: just noticed that this is actually your PR, so you're automatically involved in the reviewing process =)
@tstuyver: invite pending

[maxfleck]

Nice, thank you.
I also like that you are linking the original EDBO paper. It took me quit a while in the beginning of our project to figure out that EDBO != EDBO+ (hyperprior-wise).