aichem-gloss 1.0.0

Global-Local-Unexplored Sampling Strategy for Bayesian Optimization

aichem-gloss

Global-Local-Unexplored Sampling Strategy (GLOSS)

Overview

aichem-gloss is a Python package that implements the GLOSS algorithm, a optimization recommender that combines three complementary search strategies:

• Global Search: Explores the parameter space using acquisition functions with uncertainty quantification (UCB/LCB)
• Local Search: Exploits promising regions around the best known points for refinement
• Unexplored Search: Discovers underexplored regions of the parameter space with distance-based filtering

This package is particularly useful for:

• 🧪 Chemical space exploration and molecular design
• 🔬 Materials discovery and optimization
• ⚙️ Hyperparameter optimization for machine learning
• 🎯 Any expensive black-box optimization problem
• 🤖 Robotic chemistry and automated experimentation

Features

• ✨ Three-strategy approach: Balanced exploration and exploitation with unexplored region discovery
• 🔄 Strict deduplication: Ensures diverse and non-redundant recommendations
• 🎯 Model-agnostic: Works with any scikit-learn compatible model (GP, RF, etc.)
• 🚀 Easy to use: Simple API with sensible defaults
• 📊 Well-documented: Comprehensive documentation and examples
• 🔧 Flexible: Customizable strategy distribution and acquisition parameters

Installation

From PyPI

pip install aichem-gloss

From Conda

conda install -c conda-forge aichem-gloss

From Source

git clone https://github.com/pic-ai-robotic-chemistry/aichem-gloss.git
cd aichem-gloss
pip install -e .

Quick Start

import numpy as np
from sklearn.gaussian_process import GaussianProcessRegressor
from aichem_gloss import GLOSSRecommender

# Define parameter bounds
bounds = [[-5, 5], [-5, 5]]

# Create and train a surrogate model
model = GaussianProcessRegressor()
X_train = np.random.uniform(-5, 5, (20, 2))
y_train = np.sum(X_train**2, axis=1)  # Example objective
model.fit(X_train, y_train)

# Create GLOSS recommender
recommender = GLOSSRecommender(
    model=model,
    bounds=bounds,
    y_better_is_smaller=True,  # Minimization problem
    kappa=1.96,  # Exploration parameter
    exploration_fraction=0.5
)

# Fit with historical data
recommender.fit(X_train, y_train)

# Get recommendations
suggestions = recommender.suggest(n_recommend=8)
print("Recommended points:", suggestions['mixed'])

Documentation

Full documentation is available at aichem-gloss.readthedocs.io

Examples

Check out the examples directory for more detailed usage examples:

• basic_usage.py: Basic optimization with Gaussian Process
• Using with Random Forest models
• Custom acquisition functions
• Multi-objective optimization scenarios

Algorithm Details

The GLOSS algorithm distributes recommendations across three strategies:

Strategy Distribution

• For n_recommend > 3:
  • Global: 50% of recommendations (exploration with UCB/LCB)
  • Local: 25% of recommendations (exploitation around best points)
  • Unexplored: 25% of recommendations (distance-based unexplored search)
• For n_recommend ≤ 3: All recommendations use global search

Key Features

1. Strict Sequential Deduplication: Each new candidate is checked against all previously generated points to ensure diversity
2. Dynamic History Injection: The unexplored search dynamically updates its reference points to avoid recently generated candidates
3. Adaptive Repulsion Radius: Automatically calculated based on the density of historical points
4. Fallback Mechanisms: Multiple fallback strategies ensure robust recommendation generation

API Overview

GLOSSRecommender

GLOSSRecommender(
    model=None,                    # Surrogate model (scikit-learn compatible)
    bounds=None,                   # Parameter bounds [[min, max], ...]
    y_better_is_smaller=False,     # True for minimization, False for maximization
    kappa=1.96,                    # Exploration-exploitation trade-off
    exploration_fraction=0.5       # Unexploredd region threshold (0.0-0.99)
)

Main Methods

• fit(X, y): Fit the recommender with historical data
• suggest(n_recommend): Generate new candidate points
• get_num_recommends(n_recommend): Calculate strategy distribution

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

Citation

If you use this package in your research, please cite:

@software{aichem_gloss,
  title = {aichem-gloss: Global-Local-Unexplored Sampling Strategy for Bayesian Optimization},
  author = {Zhang, B.},
  year = {2025},
  url = {https://github.com/pic-ai-robotic-chemistry/aichem-gloss}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Developed by the PIC AI Robotic Chemistry team
• Inspired by Bayesian optimization and active learning research
• Built with NumPy and SciPy

Contact

• Author: B. Zhang
• Email: [e-mail]
• Organization: PIC AI Robotic Chemistry
• GitHub: pic-ai-robotic-chemistry

Related Projects

• scikit-optimize: Bayesian optimization library
• GPyOpt: Gaussian Process optimization
• Ax: Adaptive experimentation platform