Neural networks for system identification: A control industry perspective

Although most applications of neural networks to system identification focus on one or two standard methods, a considerably greater variety is available. This paper discusses several approaches, indicating their distinctive aspects and noting successful demonstrations. These approaches can be used to develop forward or inverse models, with external or internal dynamics, and with or without a priori information. The second part of this pager addresses some issues related to the practical application of neural identification in industry: the requirements for system data, ease of use considerations, and the problem of the opacity of most neural network models. Broad-based applications of neural networks to dynamic system modeling are problematic until these and other such issues are satisfactorily resolved.