miguelgondu / bounce

Code for the paper "Bounce: Reliable High-Dimensional Bayesian Optimization for Combinatorial and Mixed Spaces"

Bounce

This is the implementation of the Bounce algorithm for the paper "Bounce: Reliable High-Dimensional Bayesian Optimization for Combinatorial and Mixed Spaces".

Citation

Please cite our paper if you use Bounce in your work:

@inproceedings{
papenmeier2023bounce,
    title = {Bounce: Reliable High-Dimensional Bayesian Optimization for Combinatorial and Mixed Spaces},
    author = {Leonard Papenmeier and Luigi Nardi and Matthias Poloczek},
    booktitle = {Advances in Neural Information Processing Systems},
    year = {2023},
    url = {https://arxiv.org/abs/2307.00618}
}

Installation

Bounce uses poetry for dependency management. Bounce requires python>=3.10. To install poetry, see here. To install the dependencies, run

poetry install

This installation will not install all benchmarks. Especially, all benchmarks inherited from SingularityBenchmark will not be installed. The easiest way to install all benchmarks is to install the singularity container (see below). To install all benchmarks manually, clone the benchmark repository parallel to this repository and follow the instructions in the README.

Singularity container

To install Singularity, see here. To build the singularity container, run

sudo singularity build bounce.sif singularity_container

Usage

We provide a script to run the experiments in the paper. The script assumes that you installed singularity and built the singularity container. The singularity container must be in the same directory as the script and be named bounce.sif. To run the experiments in the paper, run

mkdir results # create a directory for the results (necessary!)
sh ./reproduce_results.sh # bounce.sif and results/ must be in the same directory

The results will be stored in the results directory.