System Identification Programs in JAX
The main goal of this project is implementing variational smoother-error method for system identification, using JAX. This is described in the paper Parameterizations for Large-Scale Variational System Identification (arXiv:2404.10137) as variational system identification.
The functionality of the paper was first implemented in dimasad/automatica-2024-code, but as a proof-of-concept the codebase was too clunky and hard to maintain. I rewrote the code in dimasad/visid using Flax Linen, but I think that was not the best approach for this. The main reason is that SDE discretizations, Kalman filters, and other transformations on model functions that require differentiation are clunky with Flax. It is more flexible, for my approach, to work with JAX directly.
I want this code to be used for system identification of aircraft as well, similar to what was done in dimasad/scitech-2025-code, hence the name Sidpax, a homage to SIDPAC.
@article{Dutra2025Parametrizations,
title = {Parameterizations for Large-Scale Variational System Identification Using Unconstrained Optimization},
author = {Dimas Abreu Archanjo Dutra},
journal = {Automatica},
volume = {173},
pages = {112086},
year = {2025},
doi = {10.1016/j.automatica.2024.112086}
}