Train a Gaussian Process Surrogate — train_gp

Trains a GP surrogate model from parameter-observation pairs. Uses squared exponential (RBF) kernel with automatic relevance determination (ARD) via median heuristic for length scale.

Usage

train_gp_surrogate(
  X_train,
  Y_train,
  length_scale = 0,
  signal_var = 0,
  noise_var = 1e-04
)

Arguments

X_train: Matrix (n x npar). Training parameter sets.
Y_train: Matrix (n x nobs). Corresponding model outputs.
length_scale: Numeric. Kernel length scale. If 0 (default), uses the median heuristic (median pairwise distance).
signal_var: Numeric. Signal variance. If 0, uses variance of Y.
noise_var: Numeric. Observation noise variance.

Value

A list of class step_gp containing trained GP components: K_inv (inverse kernel matrix), alpha (weight vectors), hyperparameters.

Details

The GP learns the mapping: parameters -> observations, enabling cheap prediction of model outputs for new parameter sets.

Examples

set.seed(1L)
X_train <- matrix(rnorm(20 * 4), 20, 4)
Y_train <- matrix(rnorm(20 * 6), 20, 6)
gp <- train_gp_surrogate(X_train, Y_train)
pred <- predict_gp_surrogate(gp, X_train)
dim(pred$mean)
#> [1] 20  6
length(pred$uncertainty)
#> [1] 20