This contribution presents a set of original and innovative methodological procedures for the development and the verification of artificial neural networks (ANNs). The procedures are conceived to provide a practical support in the build-up and verification processes of ANNs and address the following open issues: (i) the possibility to a-priori know the minimum size and the best assortment of the training database to optimize the ANN performance; (ii) the verification of the adequacy and effectiveness of the ANN input parameters, in order to avoid redundant information that may cause overfitting; and (iii) the assessment of the reliability of the ANN predictions when applied to new data. For this purpose, the procedures are applied to an existing ANN tool that the authors have recently designed for the prediction of the main parameters characterizing wave interaction with coastal and harbor structures. The main outcomes of the analyses suggest that (i) the size of training database depends on the ANN degrees of freedom; (ii) the use of three literature indicators allows to verify the appropriate selection of the input parameters; and (iii) the Euclidean distance among training data and new inputs can provide a useful a-priori indication of the reliability of the ANN predictions.

A methodological approach for the development and verification of artificial neural networks based on an application to wave–structure interaction processes

Sara Mizar Formentin
;
Barbara Zanuttigh
2018

Abstract

This contribution presents a set of original and innovative methodological procedures for the development and the verification of artificial neural networks (ANNs). The procedures are conceived to provide a practical support in the build-up and verification processes of ANNs and address the following open issues: (i) the possibility to a-priori know the minimum size and the best assortment of the training database to optimize the ANN performance; (ii) the verification of the adequacy and effectiveness of the ANN input parameters, in order to avoid redundant information that may cause overfitting; and (iii) the assessment of the reliability of the ANN predictions when applied to new data. For this purpose, the procedures are applied to an existing ANN tool that the authors have recently designed for the prediction of the main parameters characterizing wave interaction with coastal and harbor structures. The main outcomes of the analyses suggest that (i) the size of training database depends on the ANN degrees of freedom; (ii) the use of three literature indicators allows to verify the appropriate selection of the input parameters; and (iii) the Euclidean distance among training data and new inputs can provide a useful a-priori indication of the reliability of the ANN predictions.
Sara Mizar Formentin; Barbara Zanuttigh
File in questo prodotto:
Eventuali allegati, non sono esposti

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: http://hdl.handle.net/11585/669158
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 4
  • ???jsp.display-item.citation.isi??? 3
social impact