Add Trust Regions

[CompRhys]

This is a WIP Draft PR to add trust regions.

This introduces a breaking change renaming LSRBO to LSRBOConfig.

TODO:

• [ ] Tests
• [ ] Ideally we could fill up the TR in one shot rather than needing to iterate

[CompRhys]

@jduerholt This is now in a "working" state but it has many rough edges, would appreciate a review and thoughts of how some things may be cleaned up

[jduerholt]

@CompRhys: I let some comments, when we have decided on how to handle these things, I would then have a look at the tests in the next iteration.

[jduerholt]

@CompRhys: short info: I try to get over it, today or tomorrow. Best, Johannes

[CompRhys]

Not sure why it's trying to test the base class and then giving validation errors

strategy_spec = SoboStrategy: {'domain': {'type': 'Domain', 'inputs': {'type': 'Inputs', 'features': [{'type': 'ContinuousInput', 'key... 'failure_counter': 0, 'experiment_batch_sizes': [[]], 'use_independent_tr': False, 'length': 0.8, 'X_center_idx': -1}}

    def test_strategy_should_be_serializable(strategy_spec: Spec):
        spec = strategy_spec.typed_spec()
>       obj = strategy_spec.cls(**spec)
E       TypeError: Can't instantiate abstract class TrustRegionConfig without an implementation for abstract methods 'init_trust_region', 'update_trust_region', 'validate_domain', 'validate_surrogates'

tests/bofire/data_models/serialization/test_serialization.py:60: TypeError

--- this error was caused by adding the base class to the AnyLocalSearchConfig typing Union.

[jduerholt]

Hi @CompRhys, I saw that you made some commits but also that you transferred it into a draft PR. Just ping me here, when I should review again! Best, Johannes

[CompRhys]

I want to actually have a chance to check that it works rather than just doesn't fail before moving to have it merged, thought it best to make as draft whilst not actively working on it

[jduerholt]

Ok, thanks for the info! I saw some nice stuff from you regarding SEBO in botorch/Ax. Nice! At some point, I also have to make it available here ;)

[jduerholt]

It appears to me that there are also quite a few changes needed in botorch itself to make SEBO really usable. Can you give me a "ping" when you have the feeling that it is now really usable? So that I can try to port it also to BoFire. And ofc it would be nice to also get this TR PR here at some point ready, as I really like it!

[CompRhys]

The majority of those changes do not impact how SEBO might be implemented here because the homotopy stuff is in botorch and then it's just using qnehvi so would require a similar amount of wrapper code here to implement.

I would say that it would be a great thing to add given that sparsity in high-dimensional formulation/mixtures problems is common in material science.

A missing part is a a sum-to-one constraint an equality constraints aren't yet in Ax. Not sure if that's an API design or bandwidth issue - potentially also just that equality constraints are harder to optimize so maybe ties to your cyipopt stuff relevant. And then you're into the rabbit whole of it being a pain to package cyipopt as you can't install via pip.

[jduerholt]

I agree, so far we do not had so many high-dim optimizations, but I think it would be a great benefit. And I think it should be relatively easy to implement in BoFire. I have an idea ;) We will have a hackathon next Thursday (on-site, in Germany), and I will try to come up with a solution then. So I hope, that we have Sebo in BoFire at the end of next week together with all the other constraints ;)

[CompRhys]

Looking forward to it, this PR https://github.com/pytorch/botorch/pull/2588 to botorch should have exposed everything needed to allow all the constraints allowed by botorch.

[jduerholt]

Yeah, I saw this, optimization over mixed domains is still missing there to have it fully compabible with BoFire. But I created already a PR for it: https://github.com/pytorch/botorch/pull/2639