Prior Learning for Bayesian Optimisation (PLeBO)

This repository hosts the code used for Data-driven Prior Learning for Bayesian Optimisation.

requirements.txt has the package requirements for the optimisation step.

The core files for the optimisation are optimisation.py and do_BO.py. They also use the following source files

• synthetic.py
• jam.py
• hbo.py

The preprocessing for PLeBO happens in the mcmc folder.

• requirement-numpyro.txt has the requirements for this step (which are different to those for the optimisation step). The preprocessing files are stored in the candidates folder.
• prior_learning.py has source files.
• learn_candidates_numpyro.py is the script for learning candidate hyperparameters, as well as the Gamma prior.

Some of the baselines have preprocessing steps.

• learn_supermodel.py is used to learn the set of Shared GP hyperparameters.
• learn_initial_points.py and learn_initial_points_satellite.py is used to learn initial points for Initial.

The experiments were run on a cluster, using

• write_exps.sh
• cluster_script.sh
• cluster_wrapper.sh

The figures are stored in the Figures folder. In the case of plot_prior_fit.py, use the Python environment created with requirement-numpyro.txt. To plot and print the results reported, use

• plot_cluster_grouped_results.py
• plot_example_tasks.py
• plot_motivation_together.py
• plot_prior_fit.py
• plot_results_summary.py
• print_durations.py

The synthetic optimisation tasks were generated using generate_synth_problem_set.py. These are stored in the problems folder.