Skip to contents

Active Learning Optimizer Factory

Source: R/optimizer_active_learning.R
optimizer_active_learning.Rd

Convenience constructor that wires together an mlr3mbo::OptimizerMbo for active learning with an uncertainty-based acquisition function (SD) and optional multipoint (batch) proposal via BatchProposer batch strategies.

Usage

optimizer_active_learning(
  learner,
  se_method = c("auto", "bootstrap", "quantile"),
  se_method_n_bootstrap = 30L,
  batch_size = 1L,
  multipoint_method = c("greedy", "local_penalization", "diversity", "constant_liar"),
  aqf_optimizer = opt("pool"),
  aqf_evals = 100L
)

Arguments

learner

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

se_method

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

se_method_n_bootstrap

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

batch_size

(integer(1))
Number of points proposed per MBO iteration.

multipoint_method

(character(1))
Batch selection strategy:

  • "greedy": top-k by acquisition score (equivalent to n_candidates = batch_size)

  • "local_penalization": local penalization based diversity

  • "diversity": score/diversity trade-off

  • "constant_liar": uses mlr3mbo::bayesopt_mpcl (constant liar loop)

aqf_optimizer

(bbotk::Optimizer | mlr3mbo::AcqOptimizer)
Acquisition optimization backend. If an bbotk::Optimizer is provided, it is wrapped into a BatchProposer. Default is opt("pool").

aqf_evals

(integer(1))
Evaluation budget for acquisition optimization (TerminatorEvals$n_evals).

Value

Configured mlr3mbo::OptimizerMbo.

Details

This helper focuses on the common active learning setup in this project: