Deep GP Regression Learner

Two-layer deep Gaussian process regression. Calls deepgp::fit_two_layer() from package deepgp.

Predictions return the posterior predictive mean and the square root of the predictive variance as standard error. The learner always predicts with lite = TRUE and return_all = TRUE, then recomputes the predictive variance from the per-iteration outputs via the law of total variance. This is more robust than the package's built-in variance summary for the exposed response/se outputs.

• If D is left unset, deepgp::fit_two_layer() defaults it to the number of input columns.

• The upstream settings list is represented by explicit learner hyperparameters tagged "settings", and reconstructed internally before calling deepgp::fit_two_layer().

• train_cores maps to the cores argument of deepgp::fit_two_layer() when vecchia = TRUE; predict_cores maps to the cores argument of stats::predict().

• Only mean_map is exposed at predict time. Other upstream predict options that only generate discarded auxiliary outputs are fixed internally.

Creates a new instance of this learner.

Initial Parameter Values

• verb: Set to FALSE (upstream default is TRUE) to suppress progress output.

• train_cores, predict_cores: Set to 1 to disable threading by default. They carry the "threads" tag so mlr3::set_threads() can update them uniformly.