[Feature Request] Make gen_batch_initial_conditions more flexible

[jduerholt]

🚀 Feature Request

Motivation

When optimizing an acqf it could be possible that the default starting point sampler is not sufficient (for example when dealing with non-linear constraints or NChooseK constraints). In these case one can provide a initializer method via the ic_generator argument or samples directly via the batch_initial_conditions keyword. When looking at the default generator gen_batch_initial_conditions one sees that there is lot of logic besides simple sampling included, namely boltzmann weighting of the samples, catching of zero acqf values etc.

I propose to make this logic more broadly available to have a easy way to use it in situations where the default intialization does not work and custom samplers has to be used.

Pitch

For this purpose, I propose to add a new optional argument to gen_batch_initial_conditions which accepts a callable of type Callable[[int, int], Tensor] that just returns samples that are then further processed (boltzmann weighted based on the acqf values etc.) The two arguments of the function would be n and q and the returned Tensor would be of shape n x q x d.

What do you think?

Are you willing to open a pull request? Yes.

[Balandat]

Sounds like a great improvement to me! Happy to review a PR for this.