Cunha, João Marco Braga da; Aguiar, Alexandre Street de (Advisor). Estimating Artificial Neural Networks with Generalized Method of Moments. Rio de Janeiro, 2015. 90p. PhD Thesis -Departmento de Engenharia Elétrica, Pontifícia Universidade Católica do Rio de Janeiro.Artificial Neural Networks (ANN) started being developed in the decade of 1940. However, it was during the 1980's that the ANNs became relevant, pushed by the popularization and increasing power of computers. Also in the 1980's, there were two other two other academic events closely related to the present work: (i) a large increase of interest in nonlinear models from econometricians, culminating in the econometric approaches for ANN by the end of that decade; and (ii) the introduction of the Generalized Method of Moments (GMM) for parameter estimation in 1982. In econometric approaches for ANNs, the estimation by Quasi Maximum Likelihood (QML) always prevailed. Despite its good asymptotic properties, QML is very prone to an issue in finite sample estimations, known as overfitting. This thesis expands the state of the art in econometric approaches for ANNs by presenting an alternative to QML estimation that keeps its good asymptotic properties and has reduced leaning to overfitting. The presented approach relies on GMM estimation. As a byproduct, GMM estimation allows the use of the so-called J Test to verify the existence of neglected nonlinearity. The performed Monte Carlo studies indicate that the estimates from GMM are more accurate than those generated by QML in situations with high noise, especially in small samples. This result supports the hypothesis that GMM is susceptible to overfitting. Exchange rate forecasting experiments reinforced these findings. A second Monte Carlo study revealed satisfactory finite sample properties of the J Test applied to the neglected nonlinearity, compared with a reference test widely known and used. Overall, the results indicated that the estimation by GMM is a better alternative, especially for data with high noise level. *
KeywordsArtificial Neural Networks; Multilayer Perceptron; Overfitting; Neglected Nonlinearity; Quasi Maximum Likelihood; Generalized Method of Moments; J-Test; Global Optimization. Uma série de outros tipos de RNAs foram desenvolvidos desde então. Muitas delas possuem uma estrutura desenhada para um propósito específico e, consequentemente, são treinadas através de algoritmos próprios. Apenas como ordem de grandeza, o verbete da Wikipédia dedicado aos tipos de RNAs (em inglês) apresentava, em janeiro de 2015, sete tipos principais de RNAs (muitos dos quais com subtipos), além de uma categoria denominada "Other Types', com mais oito tipos aparentemente menos relevantes. Detalhes sobre a maioria destes tipos de RNAs podem ser encontrados no compêndio [1].Apesar da diversidade de algoritmos de aprendizagem desenvolvidos existentes, todos podem ser classificados dentro de três paradigmas básicos. Na chamada aprendizagem supervisionada, há um conjunto estático de dados e, para cada veto...