Optimal convergence factor for Gauss-Newton algorithms and its application to an adaptive parallel realization

Abstract

An efficient approach for the calculation of the optimal convergence factor for Gauss-Newton algorithms is proposed. The method leads to variable step size (convergence factor) algorithms that yield fast convergence, with an acceptable added cost in computational complexity. The results are applied to a recently proposed frequency-domain parallel realization for an adaptive IIR (infinite impulse response) filter, and evaluated in systems identification applications. Simulations confirm the expected results.<>