Learning Curve Extrapolation Learner

Abstract base class for learners that extrapolate the surrogate-model quality trajectory of an active-learning run. Subclasses fit a model on a TaskLCE and produce point and (optionally) distributional predictions of the surrogate's performance at unseen batch_nr values.

The supported predict types are:

• "response": the point forecast (predictive median on the natural scale).

• "se": additionally the total predictive standard error se and the epistemic standard error of the mean se_epistemic, both on the task's lce_link scale (see PredictionLCE).

• "quantiles": a matrix of predictive quantiles at the probabilities given by the quantile_probs parameter.

• "samples": a matrix of predictive draws (joint sample paths).

• "target_reached": a matrix of reach probabilities at the targets given by the reach_target parameter; needs the task's optimization direction (from the task measure, or from a directed codomain for best-so-far tasks).

Subclasses that model the curve as Gaussian on the link scale assemble their predictions with lce_distr_predict(); subclasses that produce explicit predictive draws use lce_samples_predict().

Creates a new LCE learner.

Arguments

id

(character(1))
Learner id.

param_set

(paradox::ParamSet).

predict_types

(character())
One or more of "response", "se", "quantiles", "samples", "target_reached".

feature_types

(character())
Supported feature types. Defaults to "integer" (the type of batch_nr).

properties

(character()).

packages

(character()).

label

(character(1)).

man

(character(1)).

Parameters

Depending on the supported predict types, the following parameters are added automatically:

• quantile_probs :: numeric()
  Probabilities for the "quantiles" predict type. Defaults to c(0.05, 0.25, 0.5, 0.75, 0.95).

• reach_target :: numeric()
  Target value(s) for the "target_reached" predict type. No default.