Active Learning Optimizer Factory

Convenience constructor that wires together an OptimizerAL for uncertainty-based active learning with optional multipoint proposal heuristics.

Usage

optimizer_active_learning(
  learner,
  se_method = c("auto", "bootstrap", "quantile"),
  n_bootstrap = 30L,
  batch_size = 1L,
  multipoint_method = c("greedy", "local_penalization", "diversity", "constant_liar"),
  acq_optimizer = opt("random_search"),
  acq_evals = 100L
)

Arguments

learner

(mlr3::LearnerRegr)
Base regression learner used as the surrogate.

se_method

(character(1))
How to obtain standard errors:

• "auto": use native "se" if supported by learner, otherwise "bootstrap".

• "bootstrap": wrap via LearnerRegrBootstrapSE.

• "quantile": wrap via LearnerRegrQuantileSE (requires "quantiles" support).

n_bootstrap

(integer(1))
Number of bootstrap replicates for "bootstrap". Ignored otherwise.

batch_size

(integer(1))
Number of points proposed per active-learning iteration.

multipoint_method

(character(1))
Batch selection strategy:

• "greedy": top-k by acquisition score

• "local_penalization": sequential local-penalization heuristic

• "diversity": sequential score/diversity trade-off

• "constant_liar": sequential pseudo-label batching

acq_optimizer

(bbotk::Optimizer | mlr3mbo::AcqOptimizer)
Optimizer used to choose the candidate-generation strategy for acquisition scoring. The current implementation translates common optimizers to a SpaceSampler and ignores optimizer-specific search logic.

acq_evals

(integer(1))
Number of candidate points scored per proposal round.

Value

Configured OptimizerAL.

Details

This helper builds an active-learning optimizer around:

• an uncertainty acquisition function ("sd")

• a surrogate that can provide standard errors (either native "se", LearnerRegrBootstrapSE, or LearnerRegrQuantileSE)

• proposer-based batch construction via ALProposerScore, ALProposerSequentialScore, or ALProposerPseudoLabel

acq_evals controls the size of the candidate pool scored in each proposal round. For continuous search spaces, candidates are sampled from the search space using a coarse translation of acq_optimizer to a SpaceSampler. For finite pools, the same sampler is applied to the remaining pool.