Neural network identification: a survey of gradient based methods

Researchers in the artificial intelligence community view system identification as a training task, while those with a control background see it as a parameter estimation problem. A third and more general perspective is to view it as an optimization problem in which a performance index is minimised with respect to the parameters being identified. While these diverse interpretations result in differing terminologies and representations, the algorithms involved are essentially equivalent. Here the optimization perspective will be adopted. From this perspective neural modelling structures (NARX or NARMAX) can be classified as linear, nonlinear or mixed linear-nonlinear (hybrid) in the parameters. Linear, nonlinear or hybrid optimization techniques are then used for identification