pod.ScaledGaussianProcessRegressor
ScaledGaussianProcessRegressor()Gaussian Process regressor that standardises its inputs before fitting.
[UQLab] (Scaling = true, the default used by [M25-code] and [M26]) standardises every input with the experimental-design mean and standard deviation, u = (x - mean(x)) / std(x), and fits the GP in u. This makes the length scales dimensionless (in units of each input’s standard deviation), so the length-scale bounds and initial values mean the same thing whatever the physical units of the inputs.
The scaling is computed in fit unless it has been frozen with freeze_scaling (used for bootstrap refits, so every resample shares the full-data scaling, and therefore the full-data length scales keep their meaning). predict and sample_y accept inputs in physical units. Every other behaviour, including all constructor arguments, is inherited from scikit-learn’s GaussianProcessRegressor.
Methods
| Name | Description |
|---|---|
| freeze_scaling | Fixes the input scaling so that fit reuses it instead of recomputing it. |
| transform_X | Maps physical inputs to the standardised inputs the kernel operates on. |
freeze_scaling
freeze_scaling(x_mean, x_std)Fixes the input scaling so that fit reuses it instead of recomputing it.
transform_X
transform_X(X)Maps physical inputs to the standardised inputs the kernel operates on.