Title :
On the regularisation-enhanced training of RBF networks
Author :
Ciftcioglu, Ö ; Durmisevic, S. ; Sariyildiz, S.
Author_Institution :
Dept. of Building Technol., Delft Univ. of Technol., Netherlands
fDate :
6/23/1905 12:00:00 AM
Abstract :
The radial basis functions (RBF) network is important in neuro-fuzzy systems. In particular, their common universal approximator properties make fuzzy systems as well as neural network systems excellent representations for system modelling. In data-based modelling it is important that the overfitting should be avoided to enhance the generalisation capability of the model since this is an ultimate performance measure for the validity of the model. In this respect, in the majority of the reported researches with RBF networks, the issue of overfitting is omitted and it is attempted to make modelling errors vanish at the price of hidden degradation in generalisation properties of the network. The work addresses this issue in the RBF neural networks for enhanced neuro-fuzzy system modelling
Keywords :
function approximation; fuzzy systems; generalisation (artificial intelligence); learning (artificial intelligence); least squares approximations; radial basis function networks; RBF networks; common universal approximator properties; data-based modelling; generalisation capability; neuro-fuzzy systems; overfitting; radial basis functions network; regularisation-enhanced training; system modelling; Buildings; Chromium; Context modeling; Degradation; Fuzzy neural networks; Fuzzy systems; Least squares approximation; Neural networks; Radial basis function networks; Supervised learning;
Conference_Titel :
Fuzzy Systems, 2001. The 10th IEEE International Conference on
Conference_Location :
Melbourne, Vic.
Print_ISBN :
0-7803-7293-X
DOI :
10.1109/FUZZ.2001.1007324
