A Meta-Reinforcement Learning Algorithm for Causal Discovery

Getting Started

• Clone this repository.
• Download anaconda at https://www.anaconda.com/products/individual
• Create a new environment with python=3.7.0. You can do that from the terminal with conda create --name [name] python=3.7.0).
• Activate the new environment with conda activate [name].
• Install all other packages with pip install requirements.txt
• If you are planning to use the benchmarks, make install/clone the corresponding code into the "benchmarking/third_party" folder. For getting DCDI to work you need to go to its project files and add return model in the last line of dcdi.main.dcdimain.

Training the model

If you want to reproduce the training of our model, run the following for the 3-variable environments: python training.py --test-set data/3en_0ex_8g_lin/ --n-vars 3 --save-dir experiments/delme --total-steps 200000000

And the following for the 4-variable environments: python training.py --test-set data/4en_0ex_200g_lin/ --n-vars 4 --save-dir experiments/delme --total-steps 200000000

Make sure to also check the other parameters in train.py if you want more flexibility.

Running benchmarks

To run the benchmarks, run the following: python run_benchmarks.py --val-data data/3en_0ex_8g_lin/scms.pkl --benchmarks BENCH --save-path experiments/test Where BENCH is one of ["ENCO", "random", "NOTEARS", "DCDI"]. See parameters in run_benchmarks.py for more flexibility.

Generating the Graph Data and corresponding SCMs

If you want to generate environments yourself, try the following: Example for generating 1000 random DAGs with 5 endogenous and 0 exogenous variables and for each of these graphs 10 SCMs with random linear functional relations. python gen_data.py --n-graphs 1000 --scms-per-graph 10 --save-dir PATH\5en_0ex_1000g\ --n-endo 5 --n-exo 0 --seed 1

Referencing

This code implements the algorithm of "A Meta-Reinforcement Learning Algorithm for Causal Discovery" by Andreas Sauter, Erman Acar, Vincent François-Lavet, 2022. If you are using this code, please reference

@inproceedings{sauter2023meta,
  title={A meta-reinforcement learning algorithm for causal discovery},
  author={Sauter, Andreas WM and Acar, Erman and Fran{\c{c}}ois-Lavet, Vincent},
  booktitle={Conference on Causal Learning and Reasoning},
  pages={602--619},
  year={2023},
  organization={PMLR}
}