[Bug] qMultiFidelityLowerBoundMaxValueEntropy fails when X_pending is not None

[AlexanderMouton]

🐛 Bug

When the qMultiFidelityLowerBoundMaxValueEntropy acquisition function is constructed with a number of points that are pending evaluation (X_pending) and with the SingleTaskMultiFidelityGP model, the optimize_acqf method fails. This does not happen when the qMultiFidelityMaxValueEntropy acquisition function is used.

To reproduce

** Code snippet to reproduce **

import torch
from botorch.acquisition.cost_aware import InverseCostWeightedUtility

from botorch.acquisition.max_value_entropy_search import (
    qMultiFidelityLowerBoundMaxValueEntropy,
    qMultiFidelityMaxValueEntropy
)
from botorch.acquisition.utils import project_to_target_fidelity

from botorch.fit import fit_gpytorch_model
from botorch.models.cost import AffineFidelityCostModel
from botorch.models.gp_regression_fidelity import SingleTaskMultiFidelityGP
from botorch.models.transforms.outcome import Standardize
from botorch.optim.optimize import optimize_acqf
from gpytorch.mlls import ExactMarginalLogLikelihood

# create training data
train_x = torch.rand(20, 4)
train_y = (torch.sin(train_x[:, 0]) + torch.cos(train_x[:, 1])).unsqueeze(-1)

# create model
model = SingleTaskMultiFidelityGP(
    train_X=train_x,
    train_Y=train_y,
    nu=2.5,
    linear_truncated=True,
    outcome_transform=Standardize(m=1),
    data_fidelities=[2, 3],
)
# fit the model
mll = ExactMarginalLogLikelihood(model.likelihood, model)
fit_gpytorch_model(mll)

# define function for projecting to the highest fidelity
def project(X):
    return project_to_target_fidelity(X=X, target_fidelities={2: 1.0, 3: 1.0})

# create cost aware utility
cost_utility = InverseCostWeightedUtility(
    cost_model=AffineFidelityCostModel(
        fidelity_weights={2: 1.0, 3: 1.0},
    )
)

for mve_af in [qMultiFidelityMaxValueEntropy, qMultiFidelityLowerBoundMaxValueEntropy]:
    for pending_points in [None, torch.rand(1, 4)]:
        af = mve_af(
            model=model,
            candidate_set=torch.rand(20, 4),
            cost_aware_utility=cost_utility,
            project=project,
            X_pending=pending_points,
        )

        bounds = torch.tensor([[0.0, 0.0, 0.0, 0.0], [1.0, 1.0, 1.0, 1.0]])

        candidate, acq_value = optimize_acqf(
            acq_function=af,
            bounds=bounds,
            q=1,
            num_restarts=20,
            raw_samples=100,
        )

        print(candidate)
        print(acq_value)

** Stack trace/error message **

Traceback (most recent call last):
  File "/home/alexander/workspace/ingenious-framework/IngeniousFrame/ParameterTuning/MFBO/test2.py", line 57, in <module>
    candidate, acq_value = optimize_acqf(
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/optim/optimize.py", line 563, in optimize_acqf
    return _optimize_acqf(opt_acqf_inputs)
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/optim/optimize.py", line 584, in _optimize_acqf
    return _optimize_acqf_batch(opt_inputs=opt_inputs)
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/optim/optimize.py", line 274, in _optimize_acqf_batch
    batch_initial_conditions = opt_inputs.get_ic_generator()(
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/optim/initializers.py", line 417, in gen_batch_initial_conditions
    Y_rnd_curr = acq_function(
  File "/home/alexander/.local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/home/alexander/.local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/utils/transforms.py", line 259, in decorated
    output = method(acqf, X, *args, **kwargs)
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/acquisition/max_value_entropy_search.py", line 808, in forward
    ig = self._compute_information_gain(
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/acquisition/max_value_entropy_search.py", line 860, in _compute_information_gain
    return qLowerBoundMaxValueEntropy._compute_information_gain(
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/acquisition/max_value_entropy_search.py", line 564, in _compute_information_gain
    posterior_m = self.model.posterior(
  File "/home/alexander/.local/lib/python3.10/site-packages/botorch/models/gpytorch.py", line 383, in posterior
    mvn = self(X)
  File "/home/alexander/.local/lib/python3.10/site-packages/gpytorch/models/exact_gp.py", line 310, in __call__
    batch_shape = torch.broadcast_shapes(batch_shape, input.shape[:-2])
  File "/home/alexander/.local/lib/python3.10/site-packages/torch/functional.py", line 124, in broadcast_shapes
    raise RuntimeError("Shape mismatch: objects cannot be broadcast to a single shape")
RuntimeError: Shape mismatch: objects cannot be broadcast to a single shape

Expected Behavior

Considering that optimze_acqf works and correctly returns a candidate and its acquisition value for the qMultiFidelityMaxValueEntropy acquisition function, I expected the same would be true for the qMultiFidelityLowerBoundMaxValueEntropy acquisition function

System information

Please complete the following information:

• BoTorch Version 0.9.5
• GPyTorch Version 1.11
• PyTorch Version 2.1.2+cu121
• Ubuntu 22.04.3 LTS

[Balandat]

Thanks for raising this issue, this indeed looks like a bug. Not entirely sure what's going on, but it seems it may be related to the fact that some parts of qMultiFidelityLowerBoundMaxValueEntropy make an assumption of q=1 (due to sequential handling), which is violated if there is a pending point.

[esantorella]

Thanks for flagging this. As a quick "fix," we'd be happy to accept a PR that raises a NotImplementedError when pending points are passed to qMultiFidelityLowerBoundMaxValueEntropy. Enabling q > 1 would of course be ideal, but we may not get to it quickly.

[AlexanderMouton]

Hi @esantorella

Please see PR 2193, thanks!

[esantorella]

Thanks for raising this issue, this indeed looks like a bug. Not entirely sure what's going on, but it seems it may be related to the fact that some parts of qMultiFidelityLowerBoundMaxValueEntropy make an assumption of q=1 (due to sequential handling), which is violated if there is a pending point.

The docstring seems to say that both q > 1 and X_pending are supported. qMFLBMVE takes functionality from both qLowerBoundMaxValueEntropy and qMultiFidelityMaxValueEntropy, both of which explicitly support q > 1. So it's not obvious where any assumptions or errors might be sneaking in.

[Balandat]

Note: #2193 removes the X_pending arg from qMultiFidelityLowerBoundMaxValueEntropy, technically resolving the issue / not making this a bug anymore. But we'd still want to enable qMultiFidelityLowerBoundMaxValueEntropy to handle X_pending, so let's keep this open.

[saitcakmak]

Fixed in https://github.com/pytorch/botorch/issues/2183