This repository contains the code for the paper Amortized Bayesian Experimental Design for Decision-Making (Huang et al., NeurIPS 2024). The full paper can be found at arXiv. Our implementations is built based on the TNP-pytorch library.
git clone https://github.com/huangdaolang/amortized-decision-aware-bed.git
cd amortized-decision-aware-bed
conda create --name tndp python=3.9
conda activate tndp
pip install -r requirements.txt
We use Hydra to manage the configurations. See configs for all configurations and defaults.
To run toy experiments, you can run train_toy.py.
To run decision-aware active learning experiments, you can run train_tal.py.
To run top-k HPO experiments, please use train_topk.py. An example case to run on Ranger dataset is as follows:
python train_topk.py nn=tndp_topk dataset="hpo/ranger"
For HPOB dataset, you can download it from https://github.com/releaunifreiburg/HPO-B, and put the downloaded data in data/HPOB.
If you find this work useful, please cite our paper:
@inproceedings{
huang2024amortized,
title={Amortized Bayesian Experimental Design for Decision-Making},
author={Huang, Daolang and Guo, Yujia and Acerbi, Luigi and Kaski, Samuel},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
}