LERND

Lernd stands for Learning Explanatory Rules from Noisy Data. It is my implementation of the algorithm in the linked paper.

Learning the concept of even numbers from scratch*

If you found this code useful for your research please cite in your work:

@software{ingvaras_merkys_2020_4294059,
  author       = {Ingvaras Merkys},
  title        = {crunchiness/lernd: LERND - implementation of $\partial$ILP},
  month        = nov,
  year         = 2020,
  publisher    = {Zenodo},
  version      = {v1.1-alpha},
  doi          = {10.5281/zenodo.4294059},
  url          = {https://github.com/crunchiness/lernd}
}

Demo Jupyter notebooks

Demo Jupyter notebooks are available online on Kaggle for a quick look into how it works:

1. https://www.kaggle.com/ingvaras/lernd-intro-predecessor
2. https://www.kaggle.com/ingvaras/lernd-even

Notebook files for local use can be found on https://github.com/crunchiness/lernd-notebooks

Set up and run

Step 1

Run Lernd on Python 3.8+

You may create a conda environment (here named "lernd"):

conda create -n lernd python=3.8

(Follow instructions at anaconda.com to get and install the conda package manager.)

Activate the environment:

conda activate lernd

Step 2

Install requirements:

pip install -r requirements.txt

Step 3

Run experiments.

Some benchmark problems are defined in file lernd/experiments.py.

You may run lernd on them:

conda activate lernd  # activate environment if using conda
python -m lernd.experiments <problem>  # problems: predecessor, even 

Do a 100 runs at once, saving the output:

python -m lernd.run_many <problem> 100

Tests

Unit tests are in lernd/test.py.

Run them:

conda activate lernd  # activate environment if using conda
python -m lernd.test