Orthogonalisation Procedure for Training Radial Basis Functions Neural Networks

• Autorzy: Strumiłło P. , Kamiński W.
• Języki publikacji: EN

Abstrakty

EN

The concept of using orthogonalisation procedure for training Radial Basis Functions (RBF) neural network is proposed. The algorithm is simpler and numerically more stable than the standard Gram-Schmidt function orthonormalisation procedure (no division by the normalisation term required). The proposed method enables sequential network incrementation, by orthogonalised basis functions, until satisfactory data processing performance is achieved. A simple recurrent scheme for restoration of the original (Le., non-orthogonal) RBF structure is also given. The method outperforms the singular value decomposition (SVD) matrix inversion method routinely used for computation of RBF network weights. The proposed orthogonalisation procedure for RBF network training is benchmarked on two computation problems: function approximation and non-linear system modelling. .

Słowa kluczowe

EN

Radial Basis Functions neural networks; function orthogonalism; training algorithms

• Wydawca: Polska Akademia Nauk, Wydział IV Nauk Technicznych
• Czasopismo: Bulletin of the Polish Academy of Sciences. Technical Sciences
• Rocznik: 2001
• Tom: vol. 49, nr 3
• Strony: 479--491
• Opis fizyczny: 6 rys., 1 tabela, bibliogr. 23 poz.

Twórcy

autor

Strumiłło P.

autor

Kamiński W.

• Institute of Electronics, Technical University of Łódź, ul. B. Stefanowskiego 18/22, 90-924 Łódź, Poland (Instytut Elektroniki Politechniki Łódzkiej),