Empirical training process selection for three layer feedforward neural network learning procedure

International Journal of Artificial Intelligence and Knowledge Discovery

Saša S Nikolić²  Dragan S Antić²  Marko T Milojković²  Staniša Lj Perić²  Miroslav B Milovanović²  Darko B Mitić¹ 

[1] University of NišFaculty of Electronic Engineering Serbia

关键词: neural network;    training process;    Levenberg-Marquardt method;    total mean squared error;    selection;   

DOI  :  

学科分类：建筑学

来源: RG Education Society

【 摘 要 】

Neural network training procedures are analyzed in this paper with the purpose to find training type which can obtain the best network performances (minimal total mean squared error). Thirteen different training procedures are observed and tested with a large training database. The database is formed by a real servo system experimental data. Each training procedure is tested with different neural network structures. All neural networks which are used in the paper are standard three layer feedforward networks with two inputs and two outputs. Neural network structures are modified by changing the number of neurons in the hidden layer. Training procedures are tested with 9 different network structures in the first testing phase. Levenberg-Marquardt learning algorithm showed best performances after a second testing phase, where additional 6 network structures are examined. It provided minimal error after training procedure and its efficiency is empirically proved when an algorithm is used for processing large databases.

【 授权许可】

Unknown