Skip to content

Beef up GaussianProcessSurrogate#745

Draft
AdrianSosic wants to merge 68 commits intomainfrom
dev/gp
Draft

Beef up GaussianProcessSurrogate#745
AdrianSosic wants to merge 68 commits intomainfrom
dev/gp

Conversation

@AdrianSosic
Copy link
Collaborator

@AdrianSosic AdrianSosic commented Feb 10, 2026

Rough agenda:

#743 (Cleanup)

  • Clean up GP code

#746 (GP Components)

  • Extract current GP setting into preset
  • Add EDBO / SMOOTHED_EDBO preset
  • Generalize kernel machinery into component machinery using generics
  • Add configurability for mean / likelihood
  • Turn current mean / noise factories into serializable classes using new generic mechanism
  • Add low-level support for gpytorch GP components

#747 (Index Kernel)

  • Add BayBE IndexKernel class
  • Add BayBE PositiveIndexKernel class

#748 (Active Dimensions)

  • Enable active dimension control for kernels
  • Absorb hardcoded IndexKernel logic into kernel factory
  • Implement deprecation mechanism for breaking change

#757 (BoTorch Preset)

  • Add BOTORCH preset

Open TODO

  • Enable serialization for gpytorch GP components
  • Set BOTORCH preset as default
  • Add CHEN preset
  • Add preset Setting attribute
  • Make optimizer configurable

AdrianSosic and others added 4 commits February 10, 2026 17:18
* Cleans up the `GaussianProcessSurrogate` class
* Adds transfer learning tests asserting that the mechanism works
regardless of which tasks are represented in the training data
@AdrianSosic AdrianSosic self-assigned this Feb 10, 2026
@AdrianSosic AdrianSosic added enhancement Expand / change existing functionality new feature New functionality on hold PR progress is awaiting for something else to continue refactor and removed on hold PR progress is awaiting for something else to continue labels Feb 10, 2026
@Scienfitz Scienfitz mentioned this pull request Mar 4, 2026
AdrianSosic and others added 11 commits March 6, 2026 15:39
DevPR, parent is #745 

* Generalizes the existing GP kernel into multi-purpose component logic
using generics
* Adds support for GPyTorch GP components
* Enables configuration of GP mean and likelihood
* Adds `EDBO` and `SMOOTHED_EDBO` GP preset
* Completes preset logic by including mean and likelihood components
DevPR, parent is #745 

Adds the `IndexKernel` and `PositiveIndexKernel` classes

The new factory design added in #746 offers full customization of the
GP, following a composition approach. To reach the full potential that
this offers, having BayBE `IndexKernel` classes to let the user to
construct custom architectures is the natural next step, in particular
because there other reasonable choices than just having the standard ICM
product, such as:
* Selecting the type of index kernel (e.g. regular vs positive)
* Implementing multi-fidelity approaches (see #715)
* Creating kernels that can handle multiple task parameters

In addition:
* It finally provides an interface for configuring the kernel itself,
e.g. its rank (which used to be a TODO for ages)
* And it offers the new parameter-named BayBE for conveniently
specifying active dimensions for the above tasks.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

dev enhancement Expand / change existing functionality new feature New functionality refactor

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants