The learned parametric mixture method is presented for a canonical cost fun
ction based ICA model on linear mixture, with several new findings. First,
its adaptive algorithm is further refined into a simple concise form. Secon
d, the separation ability of this method is shown to be qualitatively super
ior to its original model with prefixed nonlinearity. Third, a heuristic wa
y is suggested for selecting the number of densities in a Beamed parametric
mixture. Finally, experiments have been conducted to show the success of t
his method on the sources that can either be sub-Gaussian or super-Gaussian
, as well as a combination of both the types. (C) 1998 Elsevier Science B.V
. All rights reserved.