This is the companion code for the Gaussian process training method reported in the paper Learning Gaussian Processes by Minimizing PAC-Bayesian Generalization Bounds by David Reeb et al., NIPS 2018. The code allows the users to experiment with the proposed GP training method. Please cite the above paper when reporting, reproducing or extending the results.

This software is a research prototype, solely developed for and published as part of the publication cited above. It will neither be maintained nor monitored in any way.

The PAC-GP core code depends on Tensorflow. For running the experiments, GPflow and sklearn are also required.

The PAC-GP interface is similar to the ones of GPy and GPFlow. Training and evaluation of PAC-GPs is demonstrated in pac_gp/pac_gp_example.py on a small synthetic dataset.

The experiments reported in the publication can be reproduced by executing

python epsilon_study.py  --run --plot
python sparseGP_study.py --run --plot

for the dataset boston. Results are saved in the folders epsilon and ind_points.

If a similiar experiment shall be repeated on a different dataset, the easiest way is to write a new loading function in experiments/load_dataset.py. If one wishes to use PAC-GP in a different context the wrapper function compare in experiments/helpers.py may be used. It takes care of splitting into train/test, initialization, training and evaluation.

PAC-GP is open-sourced under the MIT license. See the LICENSE file for details.

For a list of other open source components included in PAC_GP, see the file 3rd-party-licenses.txt.