-
Notifications
You must be signed in to change notification settings - Fork 409
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Feature Request] MultiTaskGP with multiple task features #2359
Comments
Hi @AdrianSosic. Your suggestion makes sense to me. It is a straightforward extension to support more broad use cases. I'd be curious to see how this works in practice. Implementation-wise, we just need to make sure to keep the model constructor backwards compatible while adding support for this. If you want to put up a PR for this, I'd be happy to review it. |
Hi @saitcakmak, thanks for the feedback. I can try to free some time for this in the next few weeks 👍🏼 Regarding backwards compatibility, perhaps it's worth discussing what the implications would be upfront. If you want full compatibility without introducing a new class, I think the we need to talk about two separate parts:
If that is what you want, I'm happy to set it up this way. Let me know what you think. |
Thanks for the detailed thoughts on this. BoTorch is technically still at 0.x.x development stage, and does not have a defined public API, so we don't need to have 100% backwards compatibility. Though, we still try our best to avoid BC breaking changes.
I think this one is ok. We can convert it to
This one I do not like. I think |
🚀 Feature Request
Currently, the
MultiTaskGP
expects exactly one task dimension, specified via thetask_feature
argument. It would be great to extend the model such that it can handle several task dimensions.Pitch
Describe the solution you'd like
I see no reason why the internal logic could not be straightforwardly generalized to handle multiple task dimensions, since the corresponding feature enters multiplicatively via an
IndexKernel
. That means, additional task dimensions could be treated in the exact same way, as additional factors to the kernel product.Describe alternatives you've considered
Allowing to specify multiple task dimensions would enable learning models over structured tasks spaces – as opposed to the alternative where several such dimensions would be flattened into a single one before entering the
MultiTaskGP
.The text was updated successfully, but these errors were encountered: