Separable nonlinear least-squares methods for on-line estimation of neural nets Hammerstein models

Abstract

When the estimation of Hammerstein models is based on a minimization of least-squares error criterion, the minimization problem becomes separable with respect to the linear parameters. Therefore, the original minimization problem can be reduced to a minimization problem only in the nonlinear parameters. The proposed recursive algorithms are resulted from this separated minimization problem. They have similar computational loads to the algorithms that result from the original unseparated problem, but they can converge faster and track better than them. In a system identification example, the proposed algorithms are shown to have better convergence and tracking properties than alternative algorithms.