LLM4AD is an open-source Python platform designed for Automatic Algorithm Design (AD) using Large Language Models (LLMs). It targets researchers and developers seeking to automate the creation and optimization of algorithms across various domains like optimization, machine learning, and scientific discovery, offering a unified interface and accelerated evaluation.

How It Works

LLM4AD leverages LLMs to generate and refine algorithmic solutions. It supports various search methods, including Evolution of Heuristics (EoH), FunSearch, and Hill Climbing, which guide the LLM's exploration of the algorithm design space. The platform provides a structured framework for defining tasks, integrating LLMs (via API or local deployment), and evaluating generated algorithms, aiming for efficient and secure experimentation.

Quick Start & Requirements

• Installation: pip install llm4ad or cd LLM4AD && pip install .
• Prerequisites: Python >= 3.9 and < 3.13. gym is required for GUI and ML tasks. numba, tensorboard, wandb, and pandas are optional for specific features.
• Setup: Requires LLM API key configuration.
• Resources: Colab notebook available for online use.
• Documentation: https://llm4ad-doc.readthedocs.io/en/latest/
• Examples: https://github.com/Optima-CityU/LLM4AD/tree/main/example

Highlighted Details

• Supports unified interfaces for methods, tasks, and LLMs.
• Features multiprocessing for evaluation acceleration and secure evaluation with process protection.
• Offers logging support (local, Wandb, Tensorboard) and a GUI for method/task selection.
• Includes implementations for various search methods (EoH, FunSearch, (1+1)-EPS) and algorithm design tasks (Optimization, ML, Science Discovery).

Maintenance & Community

The project is actively maintained and welcomes contributions and collaborations. Community engagement is facilitated through QQ groups and issue reporting. Key contributors include Fei Liu, Rui Zhang, and others from City University of Hong Kong and Southern University of Science and Technology.

Licensing & Compatibility

Licensed under the MIT License. Parts of the project utilize code licensed under the Apache License 2.0. Commercial use is permitted, but contact is requested for profit-making purposes.

Limitations & Caveats

Support for other programming languages and additional search methods are listed as "Coming soon." The gym dependency may require careful version management to avoid conflicts with existing Python environments.