calc_wIMSE: Weighted Integrated Mean-Square Error Given a New Inducing...

Description Usage Arguments Details Value Author(s) References Examples

View source: R/wimse.R

Description

Calculates the Weighted Integrated Mean-Square Error (wIMSE) given a prediction location, local neighborhood, design of inducing points, and new proposed inducing point location.

Usage

1
2
3
4
calc_wIMSE(xm1, Xm = NULL, Xn, theta = NULL, g = 1e-4,
           w_mean, w_var = NULL, integral_bounds = NULL,
           epsK = sqrt(.Machine$double.eps),
           epsQ = 1e-5, mult = NULL)

Arguments

xm1

a vector containg the location of a proposed inducing point

Xm

a design matrix of existing inducing points; ncol(Xm) = length(xm1)

Xn

a matrix of the local neighborhood; ncol(Xn) = length(xm1)

theta

the lengthscale parameter (positive number) in a Gaussian correlation function; a (default) NULL value sets the lengthscale at the square of the 10th percentile of pairwise distances between neighborhood points (similar to darg in laGP package)

g

the nugget parameter (positive number) in the covariance

w_mean

a vector of the mean (center) of the Gaussian weight; length(w_mean) should equal ncol(Xn)

w_var

a positive number or vector of positive numbers (length equal to ncol(Xn)) denoting the variance(s) in the Gaussian weight. If NULL (default), theta is used.

integral_bounds

a 2 by d matrix containing the domain bounds for the data; first row contains minimum values for each dimension, second row contains maximum values; if integral_bounds=NULL, defaults to range of the local neighborhood Xn

epsK

a small positive number added to the diagonal of the correlation matrix, of inducing points, K, for numerically stability for inversion

epsQ

a small positive number added to the diagonal of the Q matrix (see Cole (2021)) for numerically stability for inversion

mult

a vector of length nrow(X) that contains the number of replicates for each design location in X

Details

The function calculates the integrated mean-square error with a Gaussian weight with mean w_mean (i.e. predictive location) and variance w_var. By using a Gaussian weight along with a Gaussian kernel for the GP, the wIMSE is calculated in closed-form.

Value

the weighted integrated mean-sqaure error

Author(s)

D. Austin Cole austin.cole8@vt.edu

References

D.A. Cole, R.B. Christianson, and R.B. Gramacy (2021). Locally Induced Gaussian Processes for Large-Scale Simulation Experiments Statistics and Computing, 31(3), 1-21; preprint on arXiv:2008.12857; https://arxiv.org/abs/2008.12857

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
## Build a "local neighborhood" and existing inducing point design
X_center <- c(.5, .5)
Xn <- matrix(runif(100), ncol=2)
Xm <- matrix(runif(10), ncol=2)

integral_bounds <- rbind(c(0,0), c(1,1))
xm1_new <- c(.4, .2)

## Calculate the weighted integrated mean-square error
calc_wIMSE(xm1=xm1_new,  Xm=Xm, Xn=Xn, w_mean=X_center,
           integral_bounds=integral_bounds)

## Define weight's variance
calc_wIMSE(xm1=xm1_new, Xm=Xm, Xn=Xn, w_mean=X_center,
           w_var=c(.1, .2), integral_bounds=integral_bounds)

## Without an exisiting inducing point design
calc_wIMSE(xm1=xm1_new, Xm=NULL, Xn=Xn, w_mean=X_center,
           integral_bounds=integral_bounds)

liGP documentation built on July 17, 2021, 9:08 a.m.

Related to calc_wIMSE in liGP...