Leniax is a Lenia simulation library powered by JAX. It can efficiently simulate, render Lenia worlds and compute statistics.
It can also be used to search for creatures, initial conditions, kernels, and growth functions using Quality-Diversity algorithms and/or gradient optimization as the simulation engine is also fully differentiable.
To install a version from sources, clone the repo
git clone https://github.com/morgangiraud/leniax
cd leniax
Install Leniax library with conda (make sure you have it before typing the following command): make install
Then activate the environment: conda activate leniax
Finally, install the lib itself: pip install .
You can make sure that everything is working fine by running the following command: make ci
We provide multiple commented examples that you can explore in the examples folder. See the folder README.md for more information.
You can also find multiple tutorials mirroring the examples as notebooks. Check the Tutorials section in the documentation.
See here for the documentation: https://leniax.readthedocs.io
To serve the documentation locally, run the following commands:
pip install -r docs/requirements.txt
make docs
Then use your favorite browser to access the localhost URL of the docs/_build/html folder.
Remarks This library intends to provide an optimized way to search for creatures either by using quality-diversity algorithms or gradient-based algorithms.
Those require different tradeoffs and that's why 2 APIs are built. Instead of cramming everything into a single complicated API, we decided to make 2 simpler APIs.
This also means it is not currently straightforward to mix the two.
The goal of the non-differentiable search API is to provide an optimized way to search among a lot of variations. In this setting, Lenia is considered as a black-box function and we want to query the simulation as fast as possible for as many as possible inputs.
For more information about how the experiments and limitations check the experiments folder.
The gradient API does not allow one to explore multiple solutions nor multiple initializations at the same time but it allows one to compute the gradient related to any function parameters.
To know more about past and present research, check the experiments folder and its README.
In Lenia we have 3 things interacting with each other:
- A state is usually represented as a multi-layer grid (A generic metric space with a notion of distance and direction is sufficient).
- A non-local rule is applied to the state with kernels used by convolutions.
- A continuous function is applied to compute the changes.
The original Lenia aims to be a generalization of Conway's game of life and so, is a 2d world with only 1 channel to which is applied 1 kernel. Rapidly, multi-channel, multi-kernel, and multi-dimensional Lenia as been introduced (C>=1, D>=2, K>=1).
Notes:
- A state contains
C > 0channels andD > 0dimensions(1D, 2D, ...) - Kernels are applied per channel, they don't have access to other channel information.
- The convolution is a depthwise convolution
feature_count=Cso the number of kernelsKmust satisfyK = n*C. - The continuous function is applied to all channels. This is where cross-channel interactions happen.
- The Lenia state is in
[0, 1]w For more information about Lenia, please check the open science Lenia website.
This library aims to be fully generic. For now, it can handle any kind of simulation but can't render more than 3 dimensions. How to perceive/understand and or interact with more dimensions is still an open challenge.
The API is not stabilized yet. This library is a work in progress.
As one can read on the main website about QD algorithms:
A fascinating aspect of natural evolution is its ability to
produce a diversity of organisms that are all high performing in
their niche.
Quality-Diversity optimization (or illumination) is a new type of
evolutionary algorithms that aim at generating large collections
of diverse solutions that are all high-performing.
QD algorithms are an absolutely perfect match to explore and discover stable patterns in the Lenia world. We can consider stability as a measurement of quality and use any other Lenia statistics to ensure diversity.
Check the website for more information about those fantastic algorithms!
- Differentiable Quality Diversity
- Quality-Diversity Optimization: a novel branch of stochastic optimization
- Lenia - Biology of Artificial Life
- Lenia and Expanded Universe
- Growing Neural Cellular Automata
To cite this repository:
@software{leniax2022github,
author = {Morgan Giraud},
title = {{Leniax}: efficient and differentiable Lenia simulator},
url = {http://github.com/morgangiraud/leniax},
version = {0.2.1},
year = {2022},
}
In the above BibTeX entry, the version number is intended to be that from leniax/setup.py, and the year corresponds to the project's open-source release.
Leniax has been built thanks to the lenia research initiative.
To contribute you can directly interact with issues in this repository and come discuss with us in our discord