Kernel dimension control

[AdrianSosic]

DevPR, parent is #745

• Enables to specify a kernel's active dimensions in a user-friendly way, i.e. by passing parameter_names
• Adds the functionality for parameter subselection
• Absorbs the ICM construction for multi-task models from the GP code into a new ICMKernelFactory class

[jpenn2023]

GaussianProcessPreset no longer exists but is used in this file.

[AdrianSosic]

Not sure what you mean 🤔 GaussianProcessPreset still exists and is available via this path. Can you elaborate?

[AdrianSosic]

should we better call this selectors.py instead of selector.py?

[Copilot]

Pull request overview

This PR introduces kernel dimension control by deriving GPyTorch kernel active_dims/ard_num_dims from a SearchSpace, adds parameter (sub)selection utilities, and updates GP surrogate behavior around multi-task kernels (including a deprecation guard for custom kernels).

Changes:

• Replace Kernel.to_gpytorch(**kwargs) dimension arguments with Kernel.to_gpytorch(searchspace) and add BasicKernel.parameter_names to control which parameters a kernel operates on.
• Add parameter selector primitives and update default GP kernel preset to use an explicit ICM-style (base-kernel × task-kernel) construction for multi-task search spaces.
• Update tests and changelog to reflect the new kernel translation API and the new multi-task custom-kernel deprecation behavior.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
baybe/kernels/base.py	Adds searchspace-driven dimension resolution and whitelisting for BayBE-only attrs (e.g., parameter_names) during GPyTorch translation.
baybe/parameters/selector.py	Introduces a parameter selector protocol + implementations to support convenient parameter subselection.
baybe/surrogates/gaussian_process/components/kernel.py	Adds KernelFactory base (with parameter selector support) and introduces ICMKernelFactory for multi-task kernel composition.
baybe/surrogates/gaussian_process/presets/baybe.py	Makes the default kernel preset multi-task aware by dispatching to an ICM kernel factory and defining a default task kernel factory.
baybe/surrogates/gaussian_process/presets/edbo.py / edbo_smoothed.py	Wires kernel factories to pass selected parameter_names into base kernels; updates likelihood factory typing.
baybe/surrogates/gaussian_process/core.py	Updates surrogate to call kernel.to_gpytorch(searchspace=...) and adds a DeprecationError gate for custom kernels in multi-task contexts.
baybe/surrogates/gaussian_process/components/{generic,mean,likelihood}.py	Renames the component factory protocol type and aligns runtime typing.
baybe/settings.py	Validates the new BAYBE_DISABLE_CUSTOM_KERNEL_WARNING env var as a boolean-like string.
tests/test_kernels.py	Updates kernel translation tests to build a SearchSpace and verify derived active_dims/ard_num_dims.
tests/hypothesis_strategies/kernels.py	Extends kernel strategies to optionally generate kernels with parameter_names.
tests/test_deprecations.py	Adds a test covering the new multi-task custom-kernel deprecation behavior.
CHANGELOG.md	Documents the new kernel API, parameter subselection, and the multi-task custom-kernel deprecation behavior.

---

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

[Copilot]

missing is computed from unmatched (the loop-local from the last base class iteration) instead of the aggregated unmatched_attrs. This makes the sanity check incomplete and can miss BayBE attrs that don’t map to any GPyTorch constructor arg. Use unmatched_attrs when computing missing (or otherwise ensure the check considers all unmatched attributes across base classes).

Suggested change

	missing = set(unmatched) - set(kernel_attrs) - self._whitelisted_attributes
	missing = set(unmatched_attrs) - set(kernel_attrs) - self._whitelisted_attributes

[Copilot]

The comment about filtering kwargs to avoid passing ard_num_dims=None / batch_shape=None is now out of sync with the implementation: kw only contains active_dims, while ard_num_dims is always passed explicitly later. Please update/remove the comment or restore the original filtering pattern so the code and rationale stay consistent.

Suggested change

	# passed or not. For instance, `ard_num_dims = kwargs.get("ard_num_dims", 1)`
	# fails if we explicitly pass `ard_num_dims=None`.
	# passed or not. Explicitly passing `None` for a kwarg can change behavior
	# compared to omitting it, so we filter out any `None` values here.

[Copilot]

ParameterSelector defines _is_match with @abstractmethod but the class doesn’t inherit from abc.ABC / use ABCMeta, so the abstractness won’t be enforced and ParameterSelector() can be instantiated accidentally. Inherit from ABC (e.g., class ParameterSelector(ABC, ParameterSelectorProtocol):) to make the abstract contract effective.

[Copilot]

ICMKernelFactory’s base_kernel_factory / task_kernel_factory fields are typed as factories and have no converter, so passing a plain Kernel (as suggested by the deprecation error message’s guidance to “include the task kernel explicitly”) will fail at runtime when the kernel is called. Consider adding the same to_component_factory converter pattern used by GaussianProcessSurrogate.kernel_factory so these fields accept either a kernel object or a factory.