Hierarchical Optimization with Large Language Models (HOLLM)

This repository contains a minimal implementation of Hierarchical Optimization with Large Language Models (HOLLM)

Getting Started

1. Setup Environment

# Create the conda environment
conda create -n hollm python=3.11

# Activate the environment
conda activate hollm

# Install the dependencies
python -m pip install -r requirements.txt

2. LLM API Keys

In our experiments we used gemini-1.5-flash. If using the same, set up the environment variables first:

echo "export GEMINI_API_KEY={api_key}" >> ~/.zshrc
source ~/.zshrc

3. Generate Configuration Files

# Generate YAML config files for all test functions
python generate_configs.py

4. Submit Jobs

# Submit jobs with specified settings in the config files
python run_benchmarks.py --config configs/levy.yaml

Directory Structure

hollm/
├── README.md
├── requirements.txt
├── run_benchmarks.py          # Main script (now uses YAML configs)
├── configs/                       # YAML configuration files
│   ├── ackley.yaml
│   ├── hartmann.yaml
│   └── ...
├── src/                           # Core implementation
│   ├── benchmark_functions.py
│   ├── acquisition_strategies.py
│   └── ...
└── generate_configs.py            # Helper to generate YAML configs

Optimizers Implemented

• TuRBO-1 (Trust Region Bayesian Optimization)
• Expected Improvement (EI)
• Log Expected Improvement (LogEI)
• Thompson Sampling (TS)
• Random Search
• Sobol Sequence
• LLM-based Global Optimization
• HOLLM