CoreSet

CoreSet is a diversity-based acquisition function that selects candidates by maximising coverage of the input space using greedy k-centres. Rather than relying on uncertainty estimates, it calls surrogate.featurise() to obtain embeddings, then iteratively picks the candidate most distant from the current training set and previously selected candidates. This makes it compatible with any model and independent of uncertainty calibration.

Candidates are scored by selection rank (n_select − step), so the first chosen candidate receives the highest score; unselected candidates receive 0. The model’s featurise() method must return a 2-D array of shape (n_inputs, d).

class alf_tools.optimizer.acquisition_functions.core_set.CoreSet

Bases: AcquisitionFunction

Core-set acquisition function using greedy k-centres.

Greedily selects candidates that maximise the minimum distance to the training set and previously selected candidates (greedy k-centres).

Candidates are scored by their selection rank (n_select - step), so the first selected candidate receives the highest score and the last receives 1. Unselected candidates receive a score of 0. This is a maximising acquisition function that uses features as a 2-D array of shape (n_inputs, d).