Taking over GPy-ABCD

[T-Flet]

I was wondering whether the maintainers of GPy would be interested in taking over the following project, since you are in the best position to maintain and improve what is there (as well as critique my use of your back-end data structures, I am sure).

GPy-ABCD is a library built on GPy which implements an Automatic Bayesian Covariance Discovery system.

Briefly, ABCD is a GP modelling method which consists in exploring a space of modular kernels constructed by iteratively combining a small set of simple ones (which are selected for their interpretability), and returning the best fitting models using them.
Due to its modularity, ABCD is capable of generating simple text descriptions of the fits based on the identified functional shapes.

(The idea and literature are summarised in the first couple of pages of the GPy-ABCD paper, and the figure below may give an intuitive understanding of the model-space-search process)

Model-space-search algorithm (fundamentally a beam search; legend at the bottom):

Input/output Example:

Most of the library is around generating, manipulating and converting kernel expressions both symbolically and numerically (expressions are simplified and parameters are merged or removed when required).

Putting aside the kernel expressions and model-space-search process, the library contains a few new GPy kernels which might benefit from further fleshing out and inclusion in GPy itself (which was my original intention), for example an implementation of the purely-periodic component of the StdPeriodic kernel in GPy (i.e. MacKay's).

GPy-ABCD only has basic documentation in the README, which is meant for users rather than developers, therefore I would obviously make myself available to go over the codebase if there is interest in taking it over.

Tangentially, two (still relevant) GPy issues were opened in the process of developing GPy-ABCD: #814 and #879