fragile 0.0.42

Framework for developing FractalAI based algorithms.

Fragile

Fragile is a framework for developing optimization algorithms inspired by Fractal AI and running them at scale.

Features

• Provides classes and an API for easily developing planning algorithms
• Provides an classes and an API for function optimization
• Build in visualizations of the sampling process
• Fully documented and tested
• Support for parallelization and distributed search processes

About FractalAI

FractalAI is based on the framework of non-equilibrium thermodynamics, and can be used to derive new mathematical tools for efficiently exploring state spaces.

The principles of our work are accessible online:

• Arxiv manuscript describing the fundamental principles of our work.
• Blog that describes our early research process.
• Youtube channel with videos showing how different prototypes work.
• GitHub repository containing a prototype that solves most Atari games.

Getting started

Check out the getting started section of the docs, or the examples folder.

Running in docker

The fragile docker container will execute a Jupyter notebook accessible on port 8080 with password: fragile

You can pull a docker image from Docker Hub running:

    docker pull fragiletech/fragile:version-tag

Where version-tag corresponds to the fragile version that will be installed in the pulled image.

Installation

This framework has been tested in Ubuntu 18.04 and supports Python 3.6, 3.7 and 3.8. If you find any problems running it in a different OS or Python version please open an issue.

It can be installed with pip install fragile["all"].

You can find the pinned versions of the minimum requirements to install the core module in requirements.txt, and the pinned versions of all the optional requirements in requirements-all.txt.

Detailed installation instructions can be found in the docs.

Documentation

You can access the documentation on Read The Docs.

Roadmap

Upcoming features: (not necessarily in order)

• Fix documentation and add examples for the distributed module
• Upload Montezuma solver
• Add new algorithms to sample different state spaces.
• Add a benchmarking module
• Add deep learning API

Contributing

Contribution are welcome. Please take a look at contributining and respect the code of conduct.

Cite us

If you use this framework in your research please cite us as:

@misc{1803.05049,
    Author = {Sergio Hernández Cerezo and Guillem Duran Ballester},
    Title = {Fractal AI: A fragile theory of intelligence},
    Year = {2018},
    Eprint = {arXiv:1803.05049},
}

License

This project is MIT licensed. See LICENSE.md for the complete text.