NATURAL GRADIENT APPROACH TO BLIND DECONVOLUTION OF DYNAMICAL SYSTEMS
L. Zhang and A. Cichocki
zha@open.brain.riken.go.jp
In this paper we present a natural gradient approach to
blind deconvolution of dynamical systems, described by the
state space model. First we formulate blind deconvolution
problem in the framework of the state space model. The
blind deconvolution is fulfilled in two stages: internal repre
sentation and output separation, which give a new insight
into blind deconvolution in the statespace framework. The
cost function for blind deconvolution is discussed and adap
tive natural gradient learning algorithms for updating ex
ternal parameters are developed by minimizing a certain
cost function, which is derived from mutual information of
output signals. Stability of the algorithm is also given. Fi
nally computer simulations are given to show the validity
and effectiveness of the statespace approach.