mkgp 3.1.3

Classes for Gaussian Process Regression fitting of ND data with errorbars

mkgp

These classes and routines were developed by Aaron Ho, and this project repository was started in 2017. The underlying mathematics was founded on the book, “Gaussian Process for Machine Learning”, C.E. Rasmussen, C.K.I. Williams (2006).

When using this package in any research work, please cite: A. Ho et al 2019 Nucl. Fusion 59 056007, DOI: 10.1088/1741-4326/ab065a

Note that the package has been renamed from GPR1D -> mkgp in v3.0.0.

Installing the mkgp program

Installation is mandatory for this package!

For first time users, it is strongly recommended to use the GUI developed for this Python package. To obtain the Python package dependencies needed to use this capability, install this package by using the following on the command line:

pip install [--user] mkgp[gui]

Use the --user flag if you do not have root access on the system that you are working on. If you have already cloned the repository, enter the top level of the repository directory and use the following instead:

pip install [--user] -e .[gui]

Removal of the [gui] portion will no longer check for the pyqt5 and matplotlib packages needed for this functionality. However, these packages are not crucial for the base classes and algorithms.

To test the installation, execute the command line script:

mkgp_1d_demo

This demonstration benefits from having matplotlib installed, but is not required.

Documentation

Documentation of the equations used in the algorithm, along with the available kernels and optimizers, can be found in docs/. Documentation of the mkgp module can be found on GitLab pages

Using the gpr1d program

For those who wish to include the functionality of this package into their own Python scripts, a sample script is provided in src/mkgp/scripts/demo.py. The basic syntax used to create kernels, select optimizers, and perform the GP regression fits are outlined there.

For any questions or to report bugs, please do so through the proper channels in the GitLab repository.

Important note for users!

The following runtime warnings are common within this routine:

RuntimeWarning: overflow encountered in double_scalars
RuntimeWarning: invalid value encountered in true_divide
RuntimeWarning: invalid value encountered in sqrt

They are filtered out by default but may reappear if verbosity settings are modified. They normally occur when using the kernel restarts option (as in the demo) and do not necessarily mean that the final returned fit is poor.

Plotting the returned fit and errors is the recommended way to check its quality. The log-marginal-likelihood metric can also be used, but is only valuable when comparing different fits of the same data, i.e. its absolute value is meaningless.

From v1.1.1, the adjusted R² and pseudo R² metrics are now available. The adjusted R² metric provides a measure of how close the fit is to the input data points. The pseudo R² provides a measure of this closeness accounting for the input data uncertainties.