Difficulty combining SingleTaskMultiFidelityGP and qMultiFidelityMaxValueEntropy acquisition function

[jkhanfri]

Hello, I am trying to create a multi-fidelity framework with 3 discrete fidelities that is optimized with the qMultiFidelityMaxValueEntropy function as described in the documentation. I am having some issues with the dimensions of the inputs for each of the two pieces.

As described in the documentation I set my inputs to the SingleTaskMultiFidelityGP function to be a "n x (d + s)" size tensor. Since I have a 3D input space the size of my input tensor is [nx4]. Then for the qMultiFidelityMaxValueEntropy definition I set my candidate_set arg to be a [nx3] size tensor (the documentation asks for a n x d Tensor) randomly sampling the 3D input design space. I have specified my data_fidelities arg to be the correct index and followed much of the setup in this example

However, when I run the BO loop I get the error " Sizes of tensors must match except in dimension 0. Expected size 3 but got size 4 for tensor number 1 in the list."

When I change the candidate_set size to [nx4] it seems to function correctly.
Does this mean that I have to include a fidelity index in the candidate_set? What fidelity do I use for this index, the desired highest fidelity, or a random sampling of the 3 fidelities, like the random sampling of the 3D input design space?

Thank you very much!

[saitcakmak]

Hi @jkhanfri. The model inputs for multi-fidelity models will need to be n x (d + s). In the case of the multi-fidelity acquisition functions, this is often guaranteed using. the project input. Are you passing in project (that adds the missing dimension to the tensor) when constructing the acquisition function?

Alternatively, you could pass in candidate_set of shapep n x (d + s) where the fidelity dimension is fixed to the target fidelity.

[jkhanfri]

Hi Sait,
Yes I am using the project input and am still recieving the error. I initialized a separate function to use the existing project_to_target_fidelity function as below:

target_fidelities = {3: 2.0}
def project(X):
    return project_to_target_fidelity(X=X, target_fidelities=target_fidelities)

It gave me the error I posed above even with the project argument.

[saitcakmak]

Can you try passing d=4 to project_to_target_fidelity. It seems like it might be silently ignoring the projection since it is supposed to add a new dimension in this case. Maybe we should make d a required input there to fix this issue

[jkhanfri]

Strangely enough, it works if I pass my candidate_set directly into the project() function with the d=4 it works, but if I set project=project directly in the qMultiFidelityMaxValueEntropy function.

def project(X):
    return project_to_target_fidelity(X=X, target_fidelities=target_fidelities,d=4)

def get_mfmes(model):
    candidate_set = project(torch.rand(1000, bounds.size(1)-1))
    return qMultiFidelityMaxValueEntropy(
        model=model,
        candidate_set = candidate_set,
        # project=project, # if I uncomment this it does not work
        cost_aware_utility=cost_aware_utility
    )

[saitcakmak]

I looked into it a bit more and looks like candidate_set does need the fidelity dimensions included (since you might be running some candidates that are not at the target fidelity, it cannot be just filled using project). I'll put up a PR with documentation updates about this and also fix the issue with project_to_target_fidelity

[jkhanfri]

So from an understanding standpoint, should I be running candidate that are not the target fidelity (randomly search the fidelity space) in my candidate points? I am confused whether for my application they should all be the target fidelity.

In terms of the documentation I also noticed that the example in the qMultiFidelityMaxValueEntropy documentation is only for a SingleTaskGP model.

    Example:
        >>> model = SingleTaskGP(train_X, train_Y)
        >>> candidate_set = torch.rand(1000, bounds.size(1))
        >>> candidate_set = bounds[0] + (bounds[1] - bounds[0]) * candidate_set
        >>> MF_MES = qMultiFidelityMaxValueEntropy(model, candidate_set)
        >>> mf_mes = MF_MES(test_X)
    """

[saitcakmak]

When the candidate_set eventually gets consumed within the acquisiition function, it is passed through project to ensure the entropy is measured at the target fidelity. The acqf constructor wasn't doing this, so it was mixing tensors of different sizes. This is being fixed in #3205