gpder 1.0.1

Bayesian Experimental Design for Minimizing the Uncertainty of Gaussian Processes

GPder

This package offers an implementation of the Gaussian Process (GP) Regression algorithm with and without derivative information.

Description

The following kernels can be used:

• RegularKernel: Kernel for regular GP regression

  $k(x_i, x_j) = \alpha^2 \mathrm{exp} \left( -\frac{\mid \mid x_i - x_j \mid \mid^2 }{2 \ell^2} \right) + \sigma^2 I$

• DerivativeKernel: Kernel for GP regression with derivative observations. Has the same form as the regular kernel but the covariance term is expanded to include derivative observations. The added noise is also expanded with the derivative noise parameter $\sigma^2_{\nabla}$.

  $k({x}_i, {x}_j) = \alpha^2 \mathrm{exp} \left( -\frac{\mid \mid {x}_i - {x}_j \mid \mid^2 }{2{\ell}^2} \right) _{\mathrm{expanded}} + \sigma^2 _{\mathrm{expanded}} I$

See "Efficient Estimation of Unfactorizable Systematic Uncertainties".

Install

pip install gpder

References

Efficient Estimation of Unfactorizable Systematic Uncertainties