Bayesian Optimization over Discrete and Mixed Spaces via Probabilistic Reparameterization

This is the code associated with the paper "Bayesian Optimization over Discrete and Mixed Spaces via Probabilistic Reparameterization."

Please cite our work if you find it useful.

@inproceedings{daulton2022pr,
      title={Bayesian Optimization over Discrete and Mixed Spaces via Probabilistic Reparameterization}, 
      author={Samuel Daulton and Xingchen Wan and and David Eriksson and Maximilian Balandat and Michael A. Osborne and Eytan Bakshy},
      booktitle={Advances in Neural Information Processing Systems 35},
      year={2022}
}

Getting started

From the base bo_pr directory run:

pip install -e .

Structure

The code is structured in three parts.

• The utilities for constructing the acquisition functions and other helper methods are defined in discrete_mixed_bo/.
• The experiments are found in and ran from within experiments/. The main.py is used to run the experiments, and the experiment configurations are found in the config.json file of each sub-directory.

The individual experiment outputs were left out to avoid inflating the file size.

Running Experiments

To run a basic benchmark based on the config.json file in experiments/<experiment_name> using <algorithm>:

cd experiments
python main.py <experiment_name> <algorithm> <seed>

The code refers to the algorithms using the following labels:

algorithms = [
    ("sobol", "Sobol"),
    ("cont_optim__round_after__ei", "Cont. Relax."),
    ("pr__ei", "PR"),
    ("exact_round__fin_diff__ei", "Exact Round"),
    ("exact_round__ste__ei", "Exact Round + STE"),
    ("cont_optim__round_after__ts", "Cont. Relax. + TS"),
    ("pr__ts", "PR + TS"),
    ("exact_round__fin_diff__ts", "Exact Round + TS"),
    ("exact_round__ste__ts", "Exact Round + TS + STE"),
    ("cont_optim__round_after__ucb", "Cont. Relax. + UCB"
    ("pr__ucb", "PR + UCB"),
    ("exact_round__fin_diff__ucb", "Exact Round + UCB"),
    ("exact_round__ste__ucb", "Exact Round + UCB + STE"),
    ("cont_optim__round_after__ehvi", "Cont. Relax."),
    ("pr__ehvi", "PR"),
    ("exact_round__fin_diff__ehvi", "Exact Round"),
    ("exact_round__ste__ehvi", "Exact Round + STE"),
    ("cont_optim__round_after__nehvi-1", ""Cont. Relax. + TS"),
    ("pr__nehvi-1", "PR + TS"),
    ("exact_round__fin_diff__nehvi-1", "Exact Round + TS"),
    ("exact_round__ste__nehvi-1","Exact Round + TS + STE"),
]

These algorithms can be modified by additional arguments in the config.json file such as use_trust_region and a "acqf_kwargs" dictionary containing the pair"pr_use_analytic": true.

Each folder under experiments/ corresponds to the experiments in the paper according to the following mapping:

experiments = {
    "ackley13": "Ackley",
    "cco": "Cellular Network",
    "chemistry": "Chemical Reaction",
    "mixed_int_f1": "Mixed Int F1",
    "mixed_oil": "Oil Sorbent",
    "rosenbrock10_scaled": "Rosenbrock",
    "svm": "SVM",
    "welded_beam": "Welded Beam",
}

Additional folder provide the configurations for the trust region variants (simply adding "use_trust_region": true) and analytic PR where feasible simply adding ("acqf_kwargs": {"pr_use_analytic": true}). Note: this code can heavily exploit a GPU if available.

License

This repository is MIT licensed, as found in the LICENSE file.