Source code for the AISTATS 2024 paper
"Looping in the Human: Collaborative and Explainable Bayesian Optimization" arXiv
We prepared an example of CoExBO with battery example.
- Demo1 human feedback for battery experiments.ipynb
- Demo2 synthetic human response.ipynb
Collaborative and Explainable BO (CoExBO)
- BO combines experimental results and expert preferences.
- BO generates pairwise candidates along with explanations.
- Human interprets the acquisitions and picks their preferred candidate
- Human conducts experiments and repeat step 1.
Utilising GP-SHAP, we can provide insights into the undergoing of the BO by attributing feature importance to the followings:
- Surrogate GP model
- Acquisition function (GP-UCB)
botorch 0.8.4 gpytorch 1.10 torch 1.13.0
Please cite this work as
@inproceedings{adachi2023looping,
title={Looping in the Human: Collaborative and Explainable Bayesian Optimization},
author={Adachi, Masaki and Planden, Brady and Howey, David A and Maundet, Krikamol and Osborne, Michael A and Chau, Siu Lun},
booktitle={Artificial intelligence and statistics},
year={2024}
}