Uncertainty Sampling

Uncertainty Sampling scores each candidate by the surrogate’s predictive uncertainty alone, ignoring the predicted mean. It is the pure-exploration counterpart to Greedy (pure exploitation), and is equivalent to UCB in the limit of a very large exploration parameter. Prefer it when the goal is to improve the surrogate itself (model-quality metrics such as test RMSE) rather than to find high-scoring candidates.

Uncertainty is read from variances when the surrogate reports it. For an ensemble surrogate that only reports per-member predictions (empirical_dist), the disagreement between members is used instead.

class alf_tools.optimizer.acquisition_functions.uncertainty_sampling.UncertaintySampling

Bases: AcquisitionFunction

Uncertainty sampling acquisition function.

Scores each candidate by the surrogate’s predictive uncertainty alone, ignoring the predicted mean: acquisition = σ. This is the pure-exploration counterpart to Greedy (pure exploitation), and reproduces UCB’s ranking in the limit of a large exploration parameter. It is the natural baseline when the goal is to improve the surrogate (model-quality metrics such as test RMSE) rather than to find high-scoring candidates.

Uncertainty is read from variances when the surrogate reports it. For an ensemble surrogate that only reports per-member predictions (empirical_dist), the disagreement between members — their standard deviation across the ensemble axis — is used instead.

This is a maximising acquisition function.