Neural-Based Identification of Continuous Nonlinear Systems

In the study presented in this paper, applications of a three-layer feedforward networks with Gaussian hidden units is used to provide the ability to learn nonlinear characteristics of continuous dynamical systems. A new training approach based on the recursive least squares is presented. Results of this expedited learning scheme are compared to those of the more traditional method of gradient descent. Convergence property of the resulting nonlinear identification scheme is derived by applying the Lyapunov stability analysis.