Kalman based artificial neural network training algorithms for nonlinear system identification

The utility of artificial neural networks (ANNs) in nonlinear system identification and control is intimately linked with the ability to parameterize the ANN structure on the basis experimental observations. Four existing training algorithms are reviewed under a parameter estimation framework, and the method of target state backpropagation previously proposed by the authors is extended. The new algorithm follows a different approach to the generation of error signals in embedded layers by backpropagating target or desired states rather than partial derivatives. The target states are used in conjunction with a linear Kalman based update algorithm, and transients associated with initial conditions are eliminated through a time-varying method of covariance modification. Comparisons of the five algorithms are made through a system identification problem, and the error convergence associated with each algorithm versus actual training time is presented. The results demonstrate an increased rate of convergence in comparison with backpropagation.<
>