NEAT-Python

The basic idea behind NEAT is to evolve both network topology (structural adaptation) and weights/biases (parametrical adaptation).

As a starting point, you can check the XOR experiment in the examples folder.

This repository is the 2007 academic implementation (version 0.1), now updated for Python 3.10+. It is not the PyPI package neat-python (that name belongs to CodeReclaimers/neat-python). This tree keeps the original CTNN / Izhikevich / integrate-and-fire phenotypes and pole-balancing experiments.

Requirements

• Python >= 3.10
• Optional: matplotlib and pydot for plots and network diagrams (uv sync --extra viz)

Install

uv sync

Or, after packaging:

uv pip install -e .

XOR example

cd examples/xor
uv run python xor2.py

Version 0.2

Version 0.1 was developed only for academic purposes in 2007. The Python code was based on the C++ version by Kenneth O. Stanley and his paper on NeuroEvolution of Augmented Topologies. A C++ neural module was included for speed; this release defaults to the pure-Python implementations so a compiler is not required.

The original C++ sources remain in the tree (neat/nn/nn_cpp, neat/iznn, neat/ifnn, examples/pole_balancing/double_pole/dpole.cpp) but are not built. The double-pole integrator is now a Python port (dpole.py).

Paper: Structural and Parametric Evolution of Continuous-Time Recurrent Neural Networks.

Things TODO

• Create a pip package (project name: neat-python-classic) once this port is stable.
• Probably use PyTorch or TensorFlow, with a slightly different genome encoding to accommodate these changes.
• Add more activation functions and expand the ability of each neuron to adapt its own activation function.