nugget

Specify a nugget to handle ill-conditioning

Specification

• Alias: None

• Arguments: REAL

• Default: None

Description

By default, the Surfpack GP handles ill-conditioning and does not use a nugget. If the user wishes to specify a nugget, there are two approaches.

• The user can specify the value of a nugget with model-surrogate-global-gaussian_process-surfpack-nugget.

• Have Surfpack find the optimal value of the nugget. This is specified by model-surrogate-global-gaussian_process-surfpack-find_nugget. There are two options for find_nugget.

  • find_nugget = 1: assume that the reciprocal condition number of the correlation matrix R, rcondR, is zero and calculate the nugget needed to make the worst case of R not ill-conditioned.

  • find_nugget = 2: calculate rcondR, which requires a Cholesky factorization. If rcondR indicates that R is not ill-conditioned, then kriging uses the Cholesky factorization. Otherwise, if rcondR says R is ill conditioned, then kriging will calculate the nugget needed to make the worst case of R not ill conditioned. find_nugget = 1 and 2 are similar, the second option just takes more computation (the initial Cholesky factorization) for larger problems.