pyrff: Approximating Gaussian Process samples with Random Fourier Features

This project is a Python implementation of random fourier feature (RFF) approximations [1].

It is heavily inspired by the implementations from [2, 3] and generalizes the implementation to work with GP hyperparameters obtained from any GP library.

Examples are given as Jupyter notebooks for GPs fitted with PyMC3 and scikit-learn:

• Example_RFF_PyMC3_1D
• Example_RFF_PyMC3_2D
• Thompson sampling, 1D with sklearn

Installation

pyrff is released on PyPI:

pip install pyrff

Usage and Citing

pyrff is licensed under the GNU Affero General Public License v3.0.

When using robotools in your work, please cite the corresponding software version.

@software{pyrff,
  author       = {Michael Osthege and
                  Kobi Felton},
  title        = {michaelosthege/pyrff: v2.0.1},
  month        = dec,
  year         = 2020,
  publisher    = {Zenodo},
  version      = {v2.0.1},
  doi          = {10.5281/zenodo.4317685},
  url          = {https://doi.org/10.5281/zenodo.4317685}
}

Head over to Zenodo to generate a BibTeX citation for the latest release.

References

1. Hernández-Lobato, 2014 paper, code
2. PES implementation in Cornell-MOE code
3. Bradford, 2018 paper, code