thresholdIdentification {fanovaGraph}

Function to identify a suitable threshold for an estimateGraph object

Description

From an estimateGraph object and a corresponding data set, candidate treshold values are compared on the prediction performance of the corresponding additive Kriging model. The candidate thresholds are chosen by the biggest jumps in plotDeltaJumps together with 0 (the full model) and 1 (the complete additive model). For each of them the Kriging model with corresponding kernel is estimated and the leave-one-out crossvalidiations on the original data sets are compared on scatterplots and RMSE-values.

Usage

thresholdIdentification(g, x, y, n.cand = 3, covtype = "matern5_2", KM = NULL)

Arguments

g

object of class graphlist as obtained from estimateGraph

x

design matrix of input variables corresponding to g, number of columns should be number of variables

y

vector of output variables of the same length as the columns of x

n.cand

integer, the n.cand biggest jumps are chosen as candidate threshold values. The default value is 3

covtype

optional character string specifying the covariance structure to be used. The default is "matern5_2"

KM

optional object of class km, the full kriging model corresponding to g. With default value NULL, this kriging model is computed by the function itself

Value

a list including

delta

vector of threshold candidates

models

list of full model and models with applied thresholds

y.cv

list of vectors containing crossvalidation predictions for each model

RMSE

vector of residual mean squared errors for each model

Author(s)

J. Fruth, M. Jastrow

See Also

plotDeltaJumps, plotGraphChange

Examples

############ simple 3-dimensional example with one interaction

### data (usually existing)
x <- matrix(seq(0,1,,20), 20, 3)
x <- apply(x,2,sample)
y <- 2*(x[,1]-0.5) * (x[,2]-0.5) + 0.1*sin(10*x[,3])

### FANVOA graph (usually estimated from a meta model over the data)
g <- list(d=3, 
  tii = matrix(c(0.0140, 0.0008, 0.0002)),
  V = 0.0222, 
  tii.scaled = matrix(c(0.6976, 0.0432, 0.0113))
)
class(g) <- "graphlist"

### plot complete graph
plot(g, plot.i1=FALSE)

### Compare candidate thresholds on prediction performance
set.seed(1)
comparison <- thresholdIdentification(g, x, y, n.cand = 1)