Best-So-Far Optimization Trace as an LCE Task — task_lce_best_so

Builds a TaskLCE whose target is the best objective value observed up to (and including) each archive batch – the progress curve of an optimization run. This is the optimization-mode counterpart of the surrogate-performance tasks built by CallbackSurrogatePerformance and replay_surrogate_performance(): the same LCE learners, measures, and resamplings apply, forecasting future best-so-far values instead of future model quality.

The task has one row per archive evaluation, carrying the archive's feature and target columns in the archive_x / archive_y roles. The target column best_so_far is constant within a batch (batches are evaluated as a whole, so mid-batch improvements only become visible at the batch's end). The optimization direction is taken from the archive codomain's single minimize/maximize-tagged target and travels with the task via its stored codomain; direction-dependent operations (e.g. the "target_reached" predict type, lce_batches_to_target()) work without a task measure.

Usage

task_lce_best_so_far(
  archive,
  link = "identity",
  id = "best_so_far",
  label = NA_character_
)

Arguments

archive: (bbotk::Archive)
Archive of a completed single-target optimization run. Its codomain target must be tagged "minimize" or "maximize" (a "learn"-tagged target has no best value).
link: (character(1))
Name of the predictive lce_link for the resulting task. "identity" by default; note that objective values are in general not sign-constrained, so non-identity links only make sense for suitably bounded objectives.
id: (character(1))
Task id. Defaults to "best_so_far".
label: (character(1))
Optional task label.

Value

TaskLCE.