AE+SINDY-C

Code for the paper "Interpretable and Efficient Data-Driven Discovery and Control of Distributed Systems" by Florian Wolf, Nicolò Botteghi, Urban Fasel, and Andrea Manzoni.

The preprint is available on arxiv and if you use our code, please cite:

@misc{wolf2024interpretableefficientdatadrivendiscovery,
      title={Interpretable and Efficient Data-driven Discovery and Control of Distributed Systems}, 
      author={Florian Wolf and Nicolò Botteghi and Urban Fasel and Andrea Manzoni},
      year={2024},
      eprint={2411.04098},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.04098}, 
}

TL;DR

To simply reproduce all experiments in the paper, execute

cd src
sh run_experiments.sh

in your console.

Structure of the code

• /SindyRL-AE: This sub-folder is based on the sindy-rl code and the corresponding paper "SINDy-RL: Interpretable and Efficient Model-Based Reinforcement Learning" by Zolman et al., see arxiv-link. Check /SindyRL-AE/documents/LICENSE.pdf.

  We extended the code by the following new functionalities to enable the support of PDE experiments and our auto encoder framework:

  • dynamics.py containing the AutoEncoderDynamicsModel(BaseDynamicsModel),
  • burgers.py wrapping the controlgym environment for the first experiment of the paper,
  • navier_stokes.py wrapping PDEControlGym environment for the second experiment of the paper,

• /src: This sub-folder contains the actual configuration files, training and evaluation scripts.

  • run_experiments.sh will run all experiments present in the paper.
  • /config_templates contains the configuration files for the experiments in the paper. Each of the files corresponds to exactly one of the experiments and models shown in the paper.
  • /autoencoder provides the implementation of the auto encoder model, loss function and a custom logger.
  • /analysis: Helper methods to read logs, count parameters and create the heatmaps presented in the paper.
  • sindyrl_<>.py are the main scripts to train, load and evaluate the models presented in the paper.

Installation guideline

The easiest solution is to directly use the provided conda environment, via:

conda env create --name sindyrl --file=env_sindyrl.yml
conda activate sindyrl

or via the docker installation instructions provided in the sindy-rl repo.

Alternatively, one can manually install the dependencies

• Create a virtual environment with python 3.10.14.
• Install requirements of sindyrl via

  cd sindy_rl
  pip install -r requirements.txt

• Installation of controlgym from https://github.com/Flo-Wo/controlgym via

  git clone git@github.com:Flo-Wo/controlgym.git
  cd controlgym
  pip install -r .

• Install PDEcontrolGym from https://github.com/lukebhan/PDEControlGym.
• Install additional requirements via pip install -r requirements.txt.