MultiBgolearn 0.1.1

Python package designed for multi-objective Bayesian global optimization (MOBO)

MultiBgolearn

MultiBgolearn is a Python package for multi-objective Bayesian global optimization (MOBO), with a focus on materials design tasks where several properties must be optimized at the same time.

The package extends the idea of Bgolearn from single-objective optimization to multi-objective optimization. It is suitable for candidate recommendation problems that require balancing competing objectives, such as maximizing one material property while minimizing or constraining another.

Features

• Multi-objective Bayesian global optimization workflow.
• Support for common acquisition strategies, including EHVI, PI, UCB, and qNEHVI.
• Automatic model comparison using leave-one-out cross-validation.
• Surrogate models based on scikit-learn regressors.
• Candidate recommendation from a virtual search space.
• Prediction result export and model-performance visualization.

Basic Usage

from MultiBgolearn.bgo import fit

recommended_data, improvements, index = fit(
    dataset="./data/dataset.csv",
    VSdataset="./data/virtual_space.csv",
    object_num=3,
    max_search=True,
    method="EHVI",
    bootstrap=5,
)

Input Data

dataset should be a .csv, .xlsx, or .xls file. Feature columns should come first, followed by the objective columns. The number of objective columns is specified by object_num.

VSdataset should contain the candidate virtual search space. If the candidate space contains more than 20,000 rows, MultiBgolearn samples 20,000 candidates with a fixed random seed before recommendation.

Main Parameters

• dataset: path to the observed training dataset.
• VSdataset: path to the virtual search-space dataset.
• object_num: number of objective columns in dataset.
• max_search: True for maximization and False for minimization.
• method: acquisition method, such as EHVI, PI, UCB, or qNEHVI.
• assign_model: optional model name. If not provided, MultiBgolearn evaluates available models and recommends the best one.
• bootstrap: number of bootstrap rounds for uncertainty estimation.
• batch_size: number of candidates selected for batch acquisition methods.
• noise_std: observation-noise standard deviation for qNEHVI.

Author

Dr. Bin Cao
Personal homepage: https://bin-cao.github.io/
GitHub: https://github.com/Bin-Cao/MultiBgolearn
Email: [e-mail]

For questions, issues, or suggestions, please open an issue on GitHub or contact the author by email.