[FEATURE REQUEST]: Discrete handling in optimize_acqf_mixed_alternating

[StefanPSchmid]

Motivation

Oftentimes, discrete / categorical options are represented through features with more than one dimension. In optimize_acqf_mixed_alternating, discrete dimensions are optimized independently from each other, with no way to tell that some can only go together. This feature request solves this problem.

Describe the solution you'd like to see implemented in BoTorch.

Implementation of something analogous to the fixed_feature_list in optimize_acqf_mixed, where the options belonging together need to be given. Distances for the nearest neighbor can then be computed for the options being together, and the discrete step updates entire options instead of each dimension on its own.
The rest of the optimization logic would remain the same.

Describe any alternatives you've considered to the above solution.

Sticking with optimize_acqf_mixed, but that takes very long for relevant search spaces and I would like to try the described function

Is this related to an existing issue in BoTorch or another repository? If so please include links to those Issues here.

No response

Pull Request

None

Code of Conduct

• I agree to follow BoTorch's Code of Conduct

[jduerholt]

Hi @StefanPSchmid,

The main functionality that you are requesting is already possible with plain botorch. Let's assume you have some feature which is a molecule and you represent this molecule by a molecular descriptor with a size larger than 1, then you can encode the molecule as an ordinal feature (give every molecule a number from 0 to n-1 (where n is your number of molecules) and use the NumericToCategoricalEncoding input transform to build your models. By doing this, the optimizer just sees one number, but this number is under the hood broadcasted by the input transform to a vector and the model always operates one the vector whereas the the optimizer operates only on the number and you can use the mixed optimizers of botorch out of the box.

Have a look here:

botorch/botorch/models/transforms/input.py

Line 1643 in 720e8be

class NumericToCategoricalEncoding(InputTransform):

What this approach is not doing is to compute the distances in optimize_acqf_alternating based on the descriptors, it would handle the molecular feature as a categorical feature and would use the machinery for categorical features. But by doing this, one can apply the function directly to molecular descriptors etc.

Best,

Johannes

PS: In case that you need an example script, just say it, I can quickly set one up ;)

[StefanPSchmid]

Thanks @jduerholt! I wasn't aware of the NumericToCategoricalEncoding. I will have a look and am hopeful it works. I will get in touch in case further questions arise.

Thank you again,
Stefan