FixedFeatureAcquisitionFunction wraps a BoTorch acquisition function so optimization can vary some inputs while holding selected features at specified values. This is useful when a multifidelity model contains both search variables and a fidelity variable, but the policy should operate at one chosen fidelity. section 9.7 section 9.4
Configure the wrapper by supplying:
- An acquisition function, such as PosteriorMean or qMaxValueEntropy. section 9.4
- The full input dimension d, the feature indices in columns, and their fixed values in values. section 9.4
- A standard optimizer such as optimize_acqf to maximize the wrapped score over the remaining dimensions. section 9.4
Optimize the high-fidelity posterior mean
For a two-dimensional model whose second dimension represents fidelity, wrap PosteriorMean(model) with columns=[1] and values=[1]. optimize_acqf then searches for the posterior-mean maximizer while keeping fidelity at its high-fidelity value. The resulting point can be augmented with a fidelity value of 1 for final evaluation. section 9.4
