L 1 prediction error system identification: a modified AIC rule

Abstract In this paper we shall present our recent work on model estimation and model validation, An attractive expression of the AIC rule in the framework of L 1 prediction error is presented, This result based on an original formulation of the FPE criterion in the L 1 context, tends to improve either the task of model structure choice or the validation of the estimated models. Owing to the fact that the analytical properties of the LSAD criterion (least sum of absolute deviations) it was not evident at all that results comparable to the classical L 2 expressions would be found.