New criterion of identification in the multilayered perceptron modelling

摘要

This paper deals with multilayered perceptrons modelling for time series analysis. Based on recent results about the least-squares estimation for non-linear time series, we propose a complete and feasible statistical methodology for both parameter estimation (learning pro-cess) and model selection (architecture selection). In particular, we solve the parameter pruning problem for multilayer perceptron models with a stepwise search method by using a least squares BIC~* criterion which is proved to be consistent.