Bayesian Optimization With Asynchronous Expert Feedback

Code to reproduce the paper:

@inproceedings{kristiadi2024asyncBO,
  title={How Useful is Intermittent, Asynchronous Expert Feedback for {B}ayesian Optimization?},
  author={Kristiadi, Agustinus and Strieth-Kalthoff, Felix and Subramanian, Sriram Ganapathi and Fortuin, Vincent and Poupart, Pascal and Pleiss, Geoff},
  booktitle={AABI Non Archival Track},
  year={2024}
}

Setup

Requires python >= 3.10 and pytorch >= 2.0. Then install the dependencies:

pip install git+https://git@github.com/aleximmer/laplace
pip install git+https://git@github.com/wiseodd/laplace-bayesopt
pip install lightning rdkit tqdm gauche botorch

Next, download the submodules (MolSkill, for expert simulator in chemistry problems).

git submodule init
git submodule update

Usage

The expert simulator for toy problems is trained using toy_train_reward.py. The experiment scripts are toy_bo.py, chem_bo.py. Run <SCRIPT_NAME>.py --help for available options.