Generalization over different cellular automata rules learned by a deep feed-forward neural network

by Marcel Aach, Jens Henrik Goebbert, Jenia Jitsev https://arxiv.org/abs/2103.14886

Introduction

In this repository we provide the code for reproducing the experiments performend in the Paper "Generalization over different cellular automata rules learned by a deep feed-forward neural network".

Organization

├── README.md              <- README doc for reproducing experiments
├── data_generator.py      <- data generator and helper function to generate the different CA trajectories
├── ca_unet.py	           <- the encoder-decoder neural network model
├── run.py                 <- the script for running the training (and testing) the network 
├── requirements.txt       <- requirements file for running the script  
├── params
  ├── params_simple        <- the parameters for the simple generalization case
  ├── params_level_1       <- the parameters for the level 1 generalization case
  ├── params_level_2       <- the parameters for the level 2 generalization case
  ├── params_level_3_extra <- the parameters for the level 3 extrapolation generalization case
  ├── params_level_3_inter <- the parameters for the level 3 interpolation generalization case
├── juwels_run.sh          <- batch script to run the code on the JUWELS supercomputer at FZ Juelich

Installation

For running the scripts only TensorFlow 2.3.1 and Python 3.8.5 are required.

How to run

Run the run.py python file with one of the following flags:

simple: for the simple generalization using only 3x3 neighborhood in training and test set
1: for the level 1 generalization using 3x3, 5x5 and 7x7 neighborhood in training and test set
2: for the level 2 generalization using 3x3, 5x5 and 7x7 neighborhood in training but different rules in test set
extra: for the level 3 generalization using 3x3, 5x5 and 7x7 neighborhood in training but 9x9 in test set
inter: for the level 3 generalization using 3x3, 5x5 and 9x9 neighborhood in training but 7x7 in test set

Other hyperparameters (like number of epochs or batch size) can be set in the corresponding parameter file in the params folder.

The script will first train and validate the network on the training/validation datasets and afterwards evaluate on the test dataset.

Citation

If you find this work helpful, please cite our paper:

@misc{aach2021generalization,
      title={Generalization over different cellular automata rules learned by a deep feed-forward neural network}, 
      author={Marcel Aach and Jens Henrik Goebbert and Jenia Jitsev},
      year={2021},
      eprint={2103.14886},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

SLAMPAI / generalization-cellular-automata