Default recommender for campaigns

[keith-hunter]

I've been looking through the documentation and I have not been able to find an answer. What is the default recommender for campaigns? I assume it has to depend on the type of search space, but I could not find any information in the documentation.

Thanks

[Scienfitz]

Hi @keith-hunter

probably we should just explicitly print the default in the doc, I'll make a note

If you just create any campaign without specifying the recommender (ie recommender=None) and print it in the console, you will get this at the end

This is the equivalent of

from baybe.acquisition import qLogEI
from baybe.recommenders import (
    BotorchRecommender,
    RandomRecommender,
    TwoPhaseMetaRecommender,
)
from baybe.surrogates import GaussianProcessSurrogate

recommender = TwoPhaseMetaRecommender(
    initial_recommender=RandomRecommender(),
    recommender=BotorchRecommender(
        acquisition_function=qLogEI(),
        surrogate_model=GaussianProcessSurrogate(),
        ),
    switch_after=1,
)

It means this is a meta-recommender, using random recommendations until you added the first data point. Then it switches to the botorch recommender.

It does not depend on the search space type, as both underlying recommenders support hybrid / discrete / continuous spaces.

[rossfu7]

Thanks! I checked the campaign I was using (which was created without a Recommender) and it said: _botorch_acqf=qLogNoisyExpectedHypervolumeImprovement. Is this the default acquisition function when the objective is a ParetoObjective?

[AdrianSosic]

Hi @rossfu7. Exactly. Pareto optimization has been added in one of the more recent versions and we adjusted the default setting accordingly so that hypervolume ACQF is selected automatically whenever the objective covers multiple outputs 👍🏼