SLGP (version 1.0.1)

retrainSLGP: Retrain a fitted SLGP model with new data and/or estimation method

Description

This function retrains an existing SLGP model using either a Bayesian MCMC estimation, a Maximum A Posteriori (MAP) estimation, or a Laplace approximation. The model can be retrained using new data, new inference settings, or updated hyperparameters. It reuses the structure and basis functions from the original model.

Usage

retrainSLGP(
  SLGPmodel,
  newdata = NULL,
  epsilonStart = NULL,
  method,
  interpolateBasisFun = "WNN",
  nIntegral = 101,
  nDiscret = 101,
  hyperparams = NULL,
  sigmaEstimationMethod = "none",
  seed = NULL,
  opts = list(),
  trend = NULL,
  verbose = FALSE
)

Value

An updated object of class SLGP-class with retrained coefficients and updated posterior information.

Arguments

SLGPmodel

An object of class SLGP-class to be retrained.

newdata

Optional data frame containing new observations. If NULL, the original data is reused.

epsilonStart

Optional numeric vector with initial values for the coefficients \(\epsilon\).

method

Character string specifying the estimation method: one of {"MCMC", "MAP", "Laplace"}.

interpolateBasisFun

Character string specifying how basis functions are evaluated:

• "nothing" — evaluate directly at sample locations;

• "NN" — interpolate using nearest neighbor;

• "WNN" — interpolate using weighted nearest neighbors (default).

nIntegral

Integer specifying the number of quadrature points used to approximate integrals over the response domain.

nDiscret

Integer specifying the discretization grid size (used only if interpolation is enabled).

hyperparams

Optional list with updated hyperparameters. Must include:

• sigma2: signal variance;

• lengthscale: vector of lengthscales for the inputs.

sigmaEstimationMethod

Character string indicating how to estimate sigma2: either "none" (default) or "heuristic".

seed

Optional integer to set the random seed for reproducibility.

opts

Optional list of additional options passed to inference routines: stan_chains, stan_iter, ndraws, etc.

trend

Optional a function that returns the trend of the transformed GP (not to be estimated). If not provided, it defaults to 0.

verbose

Logical; if TRUE, print progress and diagnostic messages during computation. Defaults to FALSE.

References

Gautier, A. (2023). Modelling and Predicting Distribution-Valued Fields with Applications to Inversion Under Uncertainty. PhD Thesis, Universität Bern. https://boristheses.unibe.ch/4377/

Examples

# \donttest{
# Load Boston housing dataset
library(MASS)
data("Boston")
range_x <- c(0, 100)
range_response <- c(0, 50)

#Create a SLGP model but don't fit it
modelPrior <- slgp(medv ~ age,        # Use a formula to specify response and covariates
                 data = Boston,     # Use the original Boston housing data
                 method = "none",    # No training
                 basisFunctionsUsed = "RFF",         # Random Fourier Features
                 sigmaEstimationMethod = "heuristic",  # Auto-tune sigma2 (more stable)
                 predictorsLower = range_x[1],         # Lower bound for 'age'
                 predictorsUpper = range_x[2],         # Upper bound for 'age'
                 responseRange = range_response,       # Range for 'medv'
                 opts_BasisFun = list(nFreq = 200,     # Use 200 Fourier features
                                      MatParam = 5/2), # Matern 5/2 kernel
                 seed = 1)                             # Reproducibility
#Retrain using the Boston Housing dataset and a Laplace approximation scheme
modelLaplace <- retrainSLGP(SLGPmodel=modelPrior,
                            newdata = Boston,
                            method="Laplace")
# }