This PR adds a TURBO implementation based on the new batch-trust-region classes and replaces the original TURBO implementation.
The entry point filter_datasets in acquisition rules from the earlier local-models PR is called after acquisition and observation of new query-points, but before model training. Trust region algorithms do not update their regions in acquire anymore, instead the update is now in this new method. This enables the behaviour of the new TURBOBox class to match the previous TURBO implementation.
Existing notebook, unit tests and integration tests have been updated to use the new classes. The trust_region notebook shows how local models can be used with TURBOBox.
Fully backwards compatible: no
We have agreed that we don't need backwards compatibility.
Related issue(s)/PRs: None
Summary
This PR adds a TURBO implementation based on the new batch-trust-region classes and replaces the original
TURBO
implementation.The entry point
filter_datasets
in acquisition rules from the earlier local-models PR is called after acquisition and observation of new query-points, but before model training. Trust region algorithms do not update their regions inacquire
anymore, instead the update is now in this new method. This enables the behaviour of the newTURBOBox
class to match the previousTURBO
implementation.Existing notebook, unit tests and integration tests have been updated to use the new classes. The
trust_region
notebook shows how local models can be used withTURBOBox
.Fully backwards compatible: no
We have agreed that we don't need backwards compatibility.
Previously, a
TURBO
rule was created like this:This will need to be replaced with the following:
PR checklist