Noise Strategies

libKriging supports three observation-noise models, selected via the noise argument of Kriging (and WarpKriging):

\[ y_i = f(\mathbf{x}_i) + \varepsilon_i \]

Strategy	noise =	Noise model	When to use
Noise-free	NULL	\(\varepsilon_i = 0\) (exact interpolation)	Clean computer experiments
Nugget	"nugget"	\(\varepsilon_i \sim \mathcal{N}(0, \eta^2)\), \(\eta^2\) estimated	Numerical noise, ill-conditioned data
Heteroskedastic noise	numeric vector	\(\varepsilon_i \sim \mathcal{N}(0, \sigma^2_i)\), \(\sigma^2_i\) known	Physical experiments with known measurement error

Note

NuggetKriging and NoiseKriging from earlier versions of libKriging have been merged into Kriging. Simply pass noise = "nugget" or noise = <your variance vector> to the unified constructor.