Learning Curve Extrapolation Task

Task class for predicting the surrogate-model quality trajectory of an active learning run. Each row corresponds to one archive evaluation and carries the batch in which it was evaluated, the untransformed archive feature values, the archive target values, and the surrogate's performance for that batch.

The task has a single feature, batch_nr, which is the only column passed to a LearnerLCE at predict time. The archive feature and target columns are carried in the dedicated column roles archive_x and archive_y so that more sophisticated extrapolators may inspect them during training without the columns being required for prediction on new data.

Optionally, the task also carries the paradox::ParamSet search space and the bbotk::Codomain of the originating optimization run. These cannot be recovered from the archive columns alone (which only retain storage types), but they are required by extrapolators that replay the active-learning loop, since they define parameter bounds, transformations, dependencies, and the codomain optimization directions.

Two further pieces of run information can be carried for replay extrapolators: the regression mlr3::Measure that produced the performance target column (so a simulation can score on the same scale) and, for finite-pool active learning, the candidate pool (so a simulation can propose from the same candidate set). The originating surrogate is not stored here, as it is a property of the optimizer rather than of the data.

TaskLCE objects are typically constructed by tracking an OptimizerAL optimization run with a CallbackSurrogatePerformance callback.

Creates a new LCE task.

True surrogate performance for the given rows (defaults to all active rows).

Returns the archive feature data (columns with role archive_x).

Returns the archive target data (columns with role archive_y).

Arguments

id

(character(1))
Task id.

backend

(mlr3::DataBackend | data.frame())
Data backend. Coerced via mlr3::as_data_backend().

target

(character(1))
Name of the numeric column holding the surrogate performance.

batch_nr

(character(1))
Name of the integer column holding the batch number. This column becomes the task's single feature.

archive_x

(character())
Names of the archive feature columns (untransformed x values). At least one column is required at construction.

archive_y

(character())
Names of the archive target columns (objective y values). At least one column is required at construction.

search_space

(paradox::ParamSet | NULL)
Optional search space of the originating optimization run. Its parameter ids must match archive_x. Cloned and stored for replay extrapolators.

codomain

(bbotk::Codomain | NULL)
Optional codomain of the originating optimization run. Its target ids must match archive_y. Cloned and stored for replay extrapolators.

measure

(mlr3::Measure | NULL)
Optional regression measure that produced the performance target column. Cloned and stored so replay extrapolators can score on the same scale.

pool

(data.table::data.table | NULL)
Finite candidate pool of the originating run, for pool-based active learning. Its columns must be exactly archive_x (the search-space parameter ids). NULL for continuous search spaces.

link

(character(1))
Name of the predictive lce_link on whose scale the learners model the performance curve and the distributional measures interpret se. A property of the target metric; "identity" by default. See lce_link (and lce_link_from_range to derive one from a measure's range).

label

(character(1))
Optional label.

extra_args

(list())
Extra constructor arguments preserved for cloning.

rows

(integer())

Value

numeric().

data.table::data.table().

data.table::data.table().

Fields

batch_nr

(character(1))
Name of the batch_nr column.

archive_x

(character())
Names of the archive feature columns.

archive_y

(character())
Names of the archive target columns.

batch_nrs

(integer())
Batch number for every active row, in the order of task$row_ids.

search_space

(paradox::ParamSet | NULL)
Search space of the originating optimization run.

codomain

(bbotk::Codomain | NULL)
Codomain of the originating optimization run.

measure

(mlr3::Measure | NULL)
Regression measure that produced the performance target column.

pool

(data.table::data.table | NULL)
Finite candidate pool of the originating run.

link

(character(1))
Name of the predictive link scale (see lce_link).