NOTE: As of BoTorch 0.12.0, the dimension-scaled prior is the default. As such, the priors used in the paper can be used by running a standard BoTorch SingleTaskGP, with no additional changes to either the covar_module or the likelihood. See this post for more information. So, if you simply want to run the priors from the paper for your high-dimensional problems, you can just use BoTorch as usual. Just don't forget to cite the paper if you do. =)

@InProceedings{pmlr-v235-hvarfner24a,
  title = 	 {Vanilla {B}ayesian Optimization Performs Great in High Dimensions},
  author =       {Hvarfner, Carl and Hellsten, Erik Orm and Nardi, Luigi},
  booktitle = 	 {Proceedings of the 41st International Conference on Machine Learning},
  pages = 	 {20793--20817},
  year = 	 {2024},
  editor = 	 {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix},
  volume = 	 {235},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {21--27 Jul},
  publisher =    {PMLR},
  url = 	 {https://proceedings.mlr.press/v235/hvarfner24a.html},
}

Official repository for "Vanilla Bayesian Optimization Performs Great in High Dimensions"

To run any experiment in the paper:

pip install ax-platform hydra-core

Then, any of the synthetic benchmarks are runnable.

Basic Usage

Run Vanilla BO (with the proposed lengthscale scaling) by entering

python main.py

this will run 25D-embedded Levy4 by default. Change the benchmark by adding benchmark=**some_benchmark_name**. The complete list of synthetic benchmarks is

levy4_25, levy4_100, levy4_300, levy4_1000, hartmann6_25, hartmann6_100, hartmann6_300, hartmann6_1000

and the real ones (which require some more installation) are mopta, lasso_dna, svm, ant, humanoid.

Moreover, there are 4 different model variantsmodel=**model_name**, reflecting the models used in the paper including the appendix. The specification for each can be found in benchmarking/gp_priors: default, with_ops, rbf_mle, gamma_3_6.

Changing the prior of the models can be done by entering the arguments: model.gp_params.ls.loc=2 or model.gp_params.ls.scale=5 to set $\mu_0=2$ and $\sigma_0 = 5$. If nothing is entered, the default paper settings are run.

Finally, a complete example: python main.py model=default benchmark=mopta model.gp_params.ls.loc=2 seed=13 experiment_group=vanilla_bo_testrun algorithm=qlognei

Then, the results will land in results/vanilla_bo_testrun/mopta/qLogNEI/mopta_qLogNEI_run13.csv.

Other modifiable parameters can be found by checking the various options in /configs, such as benchmark.num_iters, benchmark.noise_std, acq_opt.raw_samples, and acq_opt.num_restarts.

If no options are specified, the default algorithm (used in the main results in the paper) is run, with the default number of iterations, acquisition budget and noise level.

Running real-world tasks

Lasso-DNA

To run Lasso-DNA, clone LassoBench and add it to PYTHONPATH. Run the DNA task by entering benchmark=lasso_dna

MOPTA & SVM

To run MOPTA and SVM, add BenchSuite (modified version included in the repo) to PYTHONPATH: export PYTHONPATH=${PYTHONPATH}:$PWD/BenchSuite

Run either benchmark as benchmark=mopta or benchmark=svm.

MuJoCo Ant & Humanoid

Build the recipes/mujoco_container with Singularity, and add the container path:

sudo singularity build containers/mujoco recipes/mujoco_container
export MUJOCO=${PWD}/containers/mujoco     

Then, run either benchmark as: benchmark=ant or benchmark=humanoid