Structure selection for polynomial NARX models based on simulation error minimization

This paper investigates the problem of model structure selection for polynomial NARX models. In particular it discusses how classical identification approaches based on prediction error minimization may lead to an incorrect evaluation of the importance of the regressors within the model, with the consequent inclusion of spurious terms in the model. The paper suggests an alternative approach, in which the model structure is selected based on the minimization of the simulation error. The approach is shown to be particularly effective when the identification data are not adequately exciting or oversampled or when the model family is under-parameterized.