Robust Identification of Linear Dynamical Systems with Skew-Heavy-Tailed Mixture

This study proposes a robust Bayesian method to identify linear dynamical systems (LDS) which has non-Gaussian noise in observations. The variational Bayesian (VB) framework is utilized to jointly estimate the posterior distribution of model parameters and hidden states. Augmentation techniques are used to enhance the inference process of LDS with random parameters, and variance statistics are used to evaluate the uncertainty of parameters. The efficacy of this method is demonstrated through numerical simulations.