Experiments and supplementary derivations for the paper entitled
"Variational message passing for online polynomial NARMAX identification"
published at the 2022 American Control Conference.
The goal is infer parameters in a polynomial NARMAX model (see ForneyLab node code) and simulate outputs. Typically, (recursive) least-squares or another form of maximum-likelihood estimation is used. In this project, we employ variational Free Energy Minimisation in the form of variational message passing on a Forney-style factor graph.
We run a series of verification experiments on data generated from a NARMAX system, comparing performance as a function of sample size, polynomial order and simulation noise.
Questions, comments and general feedback can be directed to the issues tracker.