A study of the generalization capability versus training in backpropagation neural networks

Author

Dalianis, P.J. ; Tzafestas, S.G. ; Anthopoulos, G.

Author_Institution

Div. of Comput. Sci., Nat. Tech. Univ. of Athens, Greece

fYear

1993

fDate

17-20 Oct 1993

Firstpage

485

Abstract

The phenomenon of overtraining in backpropagation neural networks is discussed. The relationships between network size, training set size and generalization capabilities are examined. An extension to an existing algorithm of backpropagation is described. The extended algorithm provides a new energy function and its advantages, such as improved plasticity and performance, along with its dynamic properties, are explained. The algorithm is applied to some common problems and simulation results are presented and discussed

Keywords

backpropagation; generalisation (artificial intelligence); inference mechanisms; multilayer perceptrons; BP algorithm; backpropagation neural networks; common problems; dynamic properties; energy function; generalization capabilities; generalization capability; multilayer feature-based neural networks; network size; overtraining; plasticity; simulation results; training; training set size; Backpropagation algorithms; Function approximation; Intelligent control; Intelligent networks; Intelligent robots; Multi-layer neural network; Neural networks; Robot control; Testing; Vectors;

fLanguage

English

Publisher

ieee

Conference_Titel

Systems, Man and Cybernetics, 1993. 'Systems Engineering in the Service of Humans', Conference Proceedings., International Conference on

Conference_Location

Le Touquet

Print_ISBN

0-7803-0911-1

Type

conf

DOI

10.1109/ICSMC.1993.390760

Filename

390760