Warping Gallery

WarpKriging maps each input variable through a learned transform \(w_j\) before the GP kernel is evaluated:

\[ k(\mathbf{x}, \mathbf{x}') = \sigma^2 \cdot k_{\text{base}}\!\bigl(\Phi(\mathbf{x}),\Phi(\mathbf{x}')\bigr), \qquad \Phi(\mathbf{x}) = \bigl[w_1(x_1),\dots,w_d(x_d)\bigr]. \]

Choose the warping that best matches the expected regularity of your input variables. Each page below shows the transform shape and a 1-D regression example.

Warping	Description	Free params
none	Identity — no warping	0
affine	Linear rescaling \(w(x)=ax+b\)	2
boxcox	Box–Cox power transform	1
kumaraswamy	Kumaraswamy CDF on \([0,1]\)	2
knots	Piecewise-linear monotone (Xiong 2007)	\(K+1\)
neural_mono	Monotone neural network	\(3H+1\)
mlp	Unconstrained per-variable MLP	varies
categorical	Learned embedding for nominal levels	\(L \cdot q\)
ordinal	Learned ordered positions for ordinal levels	\(L-1\)