Double inversion for transformed minimization targets

[AdrianSosic]

As described in #459, there is an unintended second inversion of minimization targets when bounds are involved, as they add a corresponding additional transformation to the target, resulting in:

• one inversion happening when going from experimental to computation target values and
• another inversion (i.e. the desired one) happening when evaluating the acquisition values, introduced in MIN mode via acquisition function #340

Minimal Example

import numpy as np
import pandas as pd

from baybe.parameters.numerical import NumericalContinuousParameter
from baybe.recommenders import BotorchRecommender
from baybe.targets.numerical import NumericalTarget

searchspace = NumericalContinuousParameter("p", [0, 1]).to_searchspace()
objective = NumericalTarget("t", "MIN", (0, 1)).to_objective()
recommender = BotorchRecommender()
measurements = pd.DataFrame({"p": np.linspace(0, 1, 100), "t": np.linspace(0, 1, 100)})
rec = recommender.recommend(1, searchspace, objective, measurements)
print(rec)

Sketch of Fix

The problem can be fixed by setting the descending argument on the following line to False

baybe/baybe/targets/numerical.py

Line 46 in fd78099

return partial(linear_transform, descending=True)

However, this would result in an incorrect behavior for targets entering a desirability objective, for which the inversion is needed. Unfortunately, the information whether an single target or a desirability objective is used is unavailable at this point. So we need to find an alternative hotfix until a "proper" solution is implemented (which is around the corner for Pareto)

[AdrianSosic]

I guess the hotfix solution could be to simply (temporarily) invert a third time in SingleTargetObjective when triggering the transformation on a MIN target. Pretty dumb, but it ain't dumb if it works 😄 can try

[AdrianSosic]

@Scienfitz, @Alwalid-Abushanab, a hotfix is up under hotfix/minimization. Can one of you confirm if it works?

[AVHopp]

I briefly executed your exact code as well as minor variations (changing bounds a bit) of it. The recommendation is consistently equal to

     p 
0   0.0

which is what we expect, right?

[AdrianSosic]

Yes. But I was hoping more on some alternative example / doing some other sanity check. I tested my example above + the original example posted in #459

[AVHopp]

Then let me do some more testing.