Towards Robust Bayesian Estimation for Linear Dynamical Systems

This paper presents a robust Bayesian approach for identification of linear dynamical systems (LDS) with non-Gaussian observation noises. The Bayesian treatment of the identification problem is formulated by introducing suitable likelihood functions and prior information. The posterior distributions over unknown parameters and hidden states are jointly estimated under variational Bayesian (VB) framework, where the augmentation technique is adopted to make inference of the LDS with random parameters, and the parameter uncertainties can be also quantified by the variance statistics. Numerical studies are performed to confirm the effectiveness of the proposed method.