LASSO-enhanced simulation error minimization method for NARX model selection

This paper investigates the combination of a previously developed simulation error minimization (SEM) method for NARX model selection with the Least Absolute Shrinkage and Selection Operator (LASSO). The latter introduces a regularization term in the performance index that penalizes model size increase. In the context of SEM-based model selection it can be used both to trim the candidate regressor set and to provide model pruning in the model construction phase. It is shown that the LASSO-enhanced SEM method significantly reduces the computational effort and provides at least as accurate model selection as the plain SEM method.