Skip to contents

Convenience Constructor for Pool-Based Active Learning Optimizers

Source: R/optimizer_pool_al.R
optimizer_pool_al.Rd

Creates an OptimizerAL with appropriate components for a specific active learning method.

Usage

optimizer_pool_al(
  method = c("gsx", "gsy", "igs", "qbc", "random", "ideal"),
  learner = NULL,
  delta = 1,
  n_init = NULL,
  init_method = NULL,
  k_qbc = 5L,
  batch_size = 1L,
  n_candidates = NULL,
  distance = NULL
)

Arguments

method

(character(1))
One of "gsx", "gsy", "igs", "qbc", "random", "ideal".

learner

(mlr3::LearnerRegr | NULL)
Regression learner. Required for "gsy", "igs", "qbc", "ideal".

delta

(numeric(1))
Exploration weight for IDEAL (default 1).

n_init

(integer(1) | NULL)
Number of initial samples. NULL uses OptimizerAL's default initialization policy.

init_method

(character(1) | NULL)
Optional initialization override: "gsx", "random", or "kmeans". NULL keeps the method-specific default ("kmeans" for IDEAL, "gsx" for GSx/GSy/iGS, "random" for random/QBC).

k_qbc

(integer(1))
Number of QBC committee members (default 5).

batch_size

(integer(1))
Points per iteration (default 1).

n_candidates

(NULL | integer(1))
Optional number of candidate points to subsample uniformly before scoring. NULL keeps exhaustive pool scoring and does not enable continuous-space use.

distance

(character(1) | ALDistance | NULL)
Distance used by every distance-based component (the GSx / iGS / IDEAL acquisition functions and the "gsx" / "kmeans" initializations). NULL (default) keeps the papers' method-specific scalings (standardization; per-dimension affine for IDEAL), which support numeric search spaces only. For mixed-type pools pass "gower" (a mlr_al_distances key) or an ALDistance object. A "kmeans" initialization combined with a non-geometry distance (such as Gower) uses the medoid-based SpaceSamplerKMedoids instead, its mixed-type analogue. Note that on mixed-type pools the surrogate learner must support character features (e.g. lrn("regr.ranger")), since archives store paradox::ParamSet factor parameters as character columns.

Value

A configured OptimizerAL.