Skip to contents

GPfit Regression Learner

Source: R/LearnerRegrGPfit.R
mlr_learners_regr.gpfit.Rd

Gaussian process regression. Calls GPfit::GP_fit() from package GPfit.

Predictions return the posterior mean and the square root of the predictive mean squared error as standard error.

  • scale is initialized to TRUE and controls whether the learner min-max scales non-constant input features to the unit hypercube before calling GPfit::GP_fit(). When scaling is enabled, constant features are dropped. The fitted GP object stores the full per-feature scaling map in feature_offset and feature_scaling. Constant features are recorded there with offset 0 and scaling 0.

  • control_search, control_best, and control_cluster map to the three components of the control vector expected by GPfit::GP_fit(). Unset components fall back to the package defaults 200 * d, 80 * d, and 2 * d, where d is the number of input columns.

  • The correlation function is configured via the hyperparameters corr_type, corr_power (for "exponential"), and corr_nu (for "matern"), which are assembled into the corr list expected by GPfit::GP_fit().

Creates a new instance of this learner.