Sains Ma1aysiana 27: 93-106 (1998)                                                                                            Pengajian Kuantitatif/

                                                                                                                                                Quantitative Studies

Robust Neural Network Predictors

Using Radial Basis Functions

Siti Hajar Salleh

Department of Mathematics, Faculty of Mathematical Sciences

Universiti Kebangsaan Malaysia

43600 UKM Bangi, Selangor D.E.

ABSTRACT

Artificial Neural Networks have frequently been considered as black-box approaches for the non-linear modelling of processes. This paper explores the development of robust neural network models using radial basis functions and statistical bagging techniques. Neuro-Predictor software is used to investigate radial basis function network models in order to produce a more robust and parsimonious approach to modelling. Each model generates output coefficients which can be observed as a regression type of relationship. The procedure to develop robust model involves averaging the coefficients, to obtain a more robust model. The approach combines the techniques of statistical bagging and stacked regression.

ABSTRAK

Rangkaian Neuron Buatan sering dianggap sebagai kaedah kotak hitam untuk pemodelan tak linear bagi proses. Dalam makalah ini ditinjau pembinaan suatu model rangkaian neuron yang teguh menggunakan fungsi asas jejari dan kaedah statistik pembungkusan. Perisian "Neuro-Peramal" digunakan untuk menyiasat model-model rangkaian fungsi asas jejari untuk rnendapatkan kaedah yang lebih teguh dan parsimoni dalam aktiviti pemodelan. Setiap model menjana pekali output yang mempunyai hubungan regresi. Tatacara membina model teguh melibatkan kaedah mempuratakan pekali. Pendekatan tersebut menggabungkan teknik pembungkusan statistik dan regresi bertingkat.

RUJUKAN/REFERENCES

Bishop, C. 1991. Improving the genera1isation properties of radial basis function neural networks. Neural Computation 3: 579-588.

Haykin, S. 1994. Neural Networks: A Comprehensive Foundation. New York: Macmillan.

Leonard, J.A. & Kramer, M.A. 1991. Radial basis function networks for classifying process faults. IEEE Control System; 31-37.

Low, D. 1994. Radial basis functions networks. IEEE Proceedings on Artificial Neural Networks; 779-782.

Monotype Corporation. 1995. Neuro Predictor: User’s Guide and Reference, (UK). Monotype Corporation PLC.