simbapy 0.4.1

System identification Methods leveraging Backpropagation

SIMBa

System Identification Methods with Backpropagation

SIMBa (System Identification Methods leveraging Backpropagation) is an open-source toolbox leveraging Pytorch's Automatic Differentiation framework for stable state-space linear SysID. It allows the user to incorporate prior knowledge (like sparsity patterns of the state-space matrices) during the identification procedure.

Intallation

SIMBa is available on pypi, it can be installed with pip install simbapy.
Note that this will NOT install SIPPY or matlabengine to avoid compatibility issues (typically dependent on your MATLAB version). You can install them separately to allow Simba's initialization properties.

Alternatively, you can clone this github repository to use simba locally.

Compatibility with matlab

If matlab is installed on your machine, you can install matlabengine. If you are on the latest version of MATLAB, pip install matlabengine works, otherwise you might need to install an older version of matlabengine. See here the supported version of MATLAB.
SIMBa needs access the System Identification Toolbox and Symbolic Math Toolbox in MATLAB.
You can disable the use of matlab by overwriting IS_MATLAB in simba.parameters

Project status

SIMBa was first presented in Stable Linear Subspace Identification: A Machine Learning Approach and subsequently extended in SIMBa: System Identification Methods leveraging Backpropagation.

Known issue

For SIPPY users

There seems to be a bug in the control library when the dimension of the control input is one, which raises an unwanted exception.
This does not affect SIMBa in general but can fail is SIPPY is required (e.g., for SIMBa's initialization).

To correct it, you'll need to go to timeresp.py in the library, typically located in your virtual environment at .venv/lib/python3.x/site-packages/control/ and add the following transpose at line 1002, before _check_convert_array:

if len(U.shape) > 1:
    U = U.T

Contact

This project is jointly led by Loris Di Natale and Muhammad Zakwan, with the participation of Bratislav Svetozarevic, Philipp Heer, Giancarlo Ferrari Trecate, and Colin N. Jones.
Urban Energy Systems Lab, Empa, Switzerland
Laboratoire d'Automatique, EPFL, Switzerland

For more information, please contact [e-mail]