Allow for Obtaining Posterior Distributions for Hyperparameters Associated with Warping Observed Data for Gaussian Process Regression

[andrew-angus]

PyMC currently does not allow for fitting hyperparameters of warped Gaussian processes (GPs), such as are outlined in the following papers: Snelson et al. "Warped Gaussian processes." Advances in neural information processing systems 16 (2003); Rios et al. "Compositionally-warped Gaussian processes." Neural Networks 118 (2019): 235-246.

The method involves finding transformations of non-Gaussian observed data which best fit the assumption of Gaussian observed data inherent in marginal GPs. Attempting to implement such a scheme in PyMC is met with the following error: "TypeError: Variables that depend on other nodes cannot be used for observed data."

It would be nice to have this feature if possible.