CGPAX

Simple implementations of Cartesian Genetic Programming (CGP) and Linear Genetic Programming (LGP) in JAX. The main goal is to interface CGP and LGP with Brax (an implementation of several Mujoco environments in JAX) to leverage parallelization on GPU and speed up the optimization.

Set up

Clone the current repo, create the conda environment for the project, and activate it.

conda env create -f environment.yml
conda activate gpax

You will need to login to Weights&Biases for your experiments to be tracked online:

wandb login

The project also uses a Telegram bot to send you notifications about the status of experiments. For it to work you need a telegram/token.yaml file structured as follows:

token: abc
chat_id: 012

More info on telegram bots and their creation is available at https://core.telegram.org/bots/tutorial.

Non-JAX version

This project does not natively support non-Jax environments. A simple version of this project edited to work with non-Jax environments is available at https://github.com/giorgia-nadizar/MarioGP-T.

Citation

If you use this code in your research paper please cite:

@inproceedings{nadizar2024naturally,
  title={Naturally Interpretable Control Policies via Graph-Based Genetic Programming},
  author={Nadizar, Giorgia and Medvet, Eric and Wilson, Dennis G},
  booktitle={European Conference on Genetic Programming (Part of EvoStar)},
  pages={73--89},
  year={2024},
  organization={Springer}
}