predictKrigingReinterpolation: Predict Kriging Model (Re-interpolating)
In bartzbeielstein/SPOT: Sequential Parameter Optimization Toolbox

Description Usage Arguments Details Value See Also Examples

Kriging predictor with re-interpolation to avoid stalling the optimization process which employs this model as a surrogate. This is supposed to be used with deterministic experiments, which do need a non-interpolating model that avoids predicting non-zero error at sample locations. This can be useful when the model is deterministic (i.e. repeated evaluations of one parameter vector do not yield different values) but does have a "noisy" structure (e.g. due to computational inaccuracies, systematical error).

1	predictKrigingReinterpolation(object, newdata, ...)

`object`	Kriging model (settings and parameters) of class `kriging`.
`newdata`	design matrix to be predicted
`...`	not used

Please note that this re-interpolation implementation will not necessarily yield values of exactly zero at the sample locations used for model building. Slight deviations can occur.

list with predicted mean y, uncertainty s (optional) and expected improvement ei (optional). Whether s and ei are returned is specified by the vector of strings object$target, which then contains "s" and "ei.

buildKriging, predict.kriging

## Test-function:
braninFunction <- function (x) {	
	(x[2]  - 5.1/(4 * pi^2) * (x[1] ^2) + 5/pi * x[1]  - 6)^2 + 
10 * (1 - 1/(8 * pi)) * cos(x[1] ) + 10
}
## Create design points
x <- cbind(runif(20)*15-5,runif(20)*15)
## Compute observations at design points (for Branin function)
y <- as.matrix(apply(x,1,braninFunction))
## Create model
fit <- buildKriging(x,y,control=list(reinterpolate=FALSE))
fit$target <- c("y","s")
## first estimate error with regressive predictor
sreg <- predict(fit,x)$s
## now estimate error with re-interpolating predictor
sreint <- predictKrigingReinterpolation(fit,x)$s
## equivalent:
fit$reinterpolate <- TRUE
sreint2 <- predict(fit,x)$s
print(sreg)
print(sreint)
print(sreint2)
## sreint should be close to zero, significantly smaller than sreg