Towards Safe Multi-Task Bayesian Optimization

Reference implementation and supplementary material for the L4DC 2024 paper “Towards Safe Multi-Task Bayesian Optimization.” The experiments use related, inexpensive tasks to accelerate safe Bayesian optimization of an expensive primary task while accounting for uncertainty in the learned task correlations.

Published paper (PMLR) · arXiv preprint · Supplementary PDF

Repository layout

Path	Purpose
code/test_run.py	Small interactive example with online posterior plots
code/run_N2.py	Two-laser experiments corresponding to Figure 3a
code/run_N5.py	Five-laser SaMSBO experiment corresponding to Figure 3b
code/plot.py	Recreates the comparison figure from the bundled result files
code/bo/, code/model/, code/cov/	Optimization loop and multi-task GP implementation
code/plant/	Laser-chain models and control utilities
code/data_paper/	Initial conditions and result data used in the paper

Requirements

The code was developed with Python 3.10.12 on Ubuntu 22.04.3 LTS. A virtual environment is recommended.

python3.10 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

slycot requires a working BLAS/LAPACK and Fortran toolchain when a compatible binary wheel is unavailable. The plotting script also uses LaTeX text rendering, so a LaTeX installation is required to reproduce the publication typography.

Running the experiments

Run the scripts from the code directory so that their relative data paths resolve correctly:

cd code

# Short visual demonstration
python test_run.py

# Generate two- and five-laser experiment data
python run_N2.py
python run_N5.py

# Use a truncated confidence parameter in either experiment
python run_N2.py trunc
python run_N5.py trunc

The full optimization experiments are computationally expensive. Experiment settings such as the disturbance level, number of evaluations, and selected controller type are documented near the top of each entry-point script.

Generated result files are written to code/data/. To recreate the comparison figure from the data shipped with the repository, run:

python plot.py

The figure is saved as code/figures/comparison.pdf.

Citation

@inproceedings{lubsen2024towards,
  title     = {Towards Safe Multi-Task Bayesian Optimization},
  author    = {L\"{u}bsen, Jannis and Hespe, Christian and Eichler, Annika},
  booktitle = {Proceedings of the 6th Annual Learning for Dynamics and Control Conference},
  series    = {Proceedings of Machine Learning Research},
  volume    = {242},
  pages     = {839--851},
  year      = {2024},
  publisher = {PMLR},
  url       = {https://proceedings.mlr.press/v242/lubsen24a.html}
}

License

This project is licensed under the GNU General Public License v3.0. See LICENSE.