km1Nugget.init is used to give good initial values to fit kriging models when there is an unknown nugget effect to be estimated.
km1Nugget.init(model)
a matrix whose rows contain initial vectors of parameters.
a vector containing the function values corresponding to par.
a list containing the covariance objects corresponding to par.
,
vectors containing lower and upper bounds for parameters.
an object of class km.
O. Roustant, David Ginsbourger, Ecole des Mines de St-Etienne.
The procedure can be summarized in 4 stages :
| 1) | Compute the variogram and deduce a first estimation of the total variance. If an initial value is provided for nugget, check its compatibility with the estimated variance. If not, use again the variogram to give a first estimation of the nugget effect. |
| 2) | Simulate several values for the nugget effect and the process variance, around the estimations obtained at stage 1). The number of simulations is the one given in model@control$pop.size. |
| 3) | If no initial value is provided for the other covariance parameters, simulate them uniformly inside the domain delimited by model@lower and model@upper. The number of simulations is the same as in stage 2). |
| 4) | Compute the likelihood at each simulated "point" (variance + nugget effect + other covariance parameters), and take the best(s) one(s). This(these) point(s) gives the first initial value(s). The number of values considered can be set by the argument multistart in km. |
km, kmEstimate