Dual auto-regressive modelling approach to Gaussian process identification

Abstract

By modelling sources as a multivariate auto-regressive (AR) process, we have recently presented a dual AR modelling approach to identify temporal sources in independent component analysis (ICA) (Cheung et al. 2000, Cheung and Xu 1999 & 2001). However, our proposed existing algorithms for this approach are only suitable for the case that the residual term of the AR source process is non-Gaussian white noise. In this paper, we further study the Gaussian case, whereby a maximum-likelihood based algorithm is presented and experimentally demonstrated.