This is a standalone version of the Neural Evolution through Augmenting Topologies (NEAT) [Stanley2002] learning algorithm that I used for gait learning in modular robots in the tol-revolve project. I am generalizing the algorithm and making it into its own python package.

Currently the algorithm uses a simplified fitness sharing process that does not maintain species.

Installation

pip install -e git+https://github.com/egdman/neat-lite.git@master#egg=neat-lite

The examples/ directory contains a simple implementation of the neural network and an example of learning a XOR network. The neural network implementation is not really a part of this package, it's only included as an example.

The XOR example has 2 phases: augmentation (developing solutions through augmenting their topologies) and reduction (trying to simplify topologies of the solutions while keeping their fitness values high).