Skip to contents

Proposer-Based Active Learning Optimizer

Source: R/OptimizerAL.R
OptimizerAL.Rd

Active-learning optimizer whose outer loop is fixed and whose proposal logic is delegated to an ALProposer.

The optimizer owns canonical run-local surrogates and unwired acquisition function prototypes. Proposers refer to those objects by registry id; the ephemeral ALContext wires and updates them lazily.

Creates a new active-learning optimizer.

Returns a run-local surrogate by id.

Arguments

proposer

(ALProposer)
Proposer used after initialization.

surrogates

(named list() of mlr3mbo::Surrogate)
Canonical surrogate registry. Defaults to an archive-only surrogate under id "archive" for model-free acquisition functions.

acq_functions

(named list() of mlr3mbo::AcqFunction)
Acquisition-function prototype registry. Registered acquisition functions are treated as unwired prototypes; any pre-set surrogate is removed with a warning.

init_sampler

(NULL | SpaceSampler)
Sampler for the initial evaluations. NULL is only valid when n_init = 0.

result_assigner

(NULL | mlr3mbo::ResultAssigner)
Result assigner used for bbotk::OptimInstance objects. Defaults to ResultAssignerNull.

grid_expansion_limit

(integer(1))
Upper limit for fully-discrete grid expansion inherited from OptimizerPoolAbstract.

surrogate_id

(character(1))
Surrogate registry id.

update

(logical(1))
Whether to update the surrogate before returning it. Repeated calls with an unchanged archive do not retrain the surrogate.

Value

mlr3mbo::Surrogate.

Fields

proposer

(ALProposer) Proposer used after initialization.

init_sampler

(NULL | SpaceSampler) Sampler for initial evaluations.

surrogates

(named list() of mlr3mbo::Surrogate) Canonical surrogate registry.

acq_functions

(named list() of mlr3mbo::AcqFunction) Unwired acquisition-function prototype registry.

result_assigner

(NULL | mlr3mbo::ResultAssigner) Result assigner.

param_set

(paradox::ParamSet) Combined parameter set of the optimizer and directly owned components.

Parameters

batch_size

integer(1)
Number of configurations evaluated per active-learning proposal round.

replace_samples

character(1)
Whether finite-pool points can be proposed again after a batch. "never" exhausts the pool without replacement; "between_batches" allows repeat evaluations in later batches, while still preventing repeats within the current proposal batch.

n_init

integer(1)
Number of initial evaluations requested before the proposer is used. If unset, fresh runs use 4 * d initial evaluations, where d is the search-space dimension, while runs with an already populated archive do not request additional initial evaluations.

The optimizer parameter set also exposes the parameter sets of directly owned components: init_sampler, proposer, surrogates, and acquisition-function constants.