SSP'05 -- Proceedings

	IEEE/SP 13th workshop on Statistical Signal Processing July, 17-20, 2005 - Bordeaux - France

Information regarding the paper

Title

A Robust General Normalised Gradient Descent Algorithm

Author(s)

Danilo Mandic	Imperial College
Dragan Obradovic	Siemens Corporate Tehcnology
Anthony Kuh	University of Hawaii

Get the paper in PDF format

To obtain Acrobat Reader (version 5 minimum required) necessary to his read.

Abstract

A modification of the General Normalised Gradient Descent (GNGD) algorithm is proposed which caters for the steady state performance of this algorithm. This has been achieved by combining a cooling schedule within the ``search then converge'' framework, with the standard GNGD which uses a gradient adaptive regularisation parameter within its learning rate. This way, we combine a very fast convergence of GNGD with additional stabilisation introduced through annealing. Initial simulations on linear and linear becnhmark signals support the approach.