Offline Contextual Bayesian Optimization

Overview

In Bayesian Optimization (BO), many times there are several systems or "tasks" to simultaneously optimize. This repository contains Multi-task Thompson Sampling (MTS), a BO algorithm we developed to pick both tasks and actions to evaluate. Because some tasks are usually more difficult than others, MTS often significantly outperforms standard BO techniques.

Getting Set Up

The code is compatible with python 2.7. First, clone this repo and run

pip install -r requirements

By default the code leverages the Dragonfly library.

Reproducing Synthetic Experiments

The plots in the paper can be reproduced by running ocbo.py and cts_ocbo.py with the appropriate options file.

cd src
mkdir data
python ocbo.py --options <path_to_option_file>

or if continuous

python cts_ocbo.py --options <path_to_option_file>

After the simulation has finished, the plots can be reproduced by

cd scripts
python discrete_plotter.py --write_dir ../data --run_id <options_name>

or

python cts_plotter.py --write_dir ../data --run_id <options_name>

For discrete experiments, use the flag --risk_neutral 1 to show the risk neutral performance instead and use --plot_props 1 flag to show the proportion of resources given to different tasks.

With the exception of the experiment in Section 4, the table below shows the option file the corresponds to a given experiment.

Citing Work

If you use any code please cite the following:

@inproceedings{char2019offline,
  title={Offline contextual bayesian optimization},
  author={Char, Ian and Chung, Youngseog and Neiswanger, Willie and Kandasamy, Kirthevasan and Nelson, Andrew Oakleigh and Boyer, Mark and Kolemen, Egemen and Schneider, Jeff},
  booktitle={Advances in Neural Information Processing Systems},
  pages={4627--4638},
  year={2019}
}