Título: Variação De Modelos Neurais Identificados A Partir De Um Sistema Caótico
Autores: Corrêa, Marcelo V.; Aguirre, Luis A.; Braga, Antônio de P.
Resumo: The present work describes some results obtained when modeling a chaotic system using Multilayer Perceptrons. Such a system is considered as a benchmark for problems of identification of non linear systems because it presents a great diversity of dynamic regimes including chaos. The results described here show that several nets that satisfied the same stop criterion have very different dynamic behaviors. Besides, such behaviors become evident when techniques such as Poincaré maps and bifurcation diagrams are used to verify Neural Network generalization. The results suggest that when the objective of the training of a Neural Network is to model non linear dynamic systems, other approaches, besides the minimization of the prediction error should be used in model validation.
Artigo em pdf: 4cbrn_034.pdf
Arquivo BibTex: 4cbrn_034.bib