Author

Junxu Li

Date of Award

1999

Level of Access

Open-Access Thesis

Degree Name

Master of Science (MS)

Department

Electrical and Computer Engineering

Advisor

Mohamad T. Musavi

Second Committee Member

Richard O. Easen

Abstract

The objective of this thesis is to present a methodology for fine-tuning the parameters of radial basis function (RBF) neural networks, thus improving their performance. Three main parameters affect the performance of an RBF network. They are the centers and widths of the RBF nodes and the weights associated with each node. A gridded center and orthogonal search algorithm have been used to initially determine the parameters of the RBF network. A parameter tuning algorithm has been developed to optimize these parameters and improve the performance of the RBF network. When necessary, the recursive least square solution may be used to include new nodes to the network architecture. To study the behavior of the proposed network, six months of real data at fifteen-minute intervals has been collected from a North American pulp and paper company. The data has been used to evaluate the performance of the proposed network in the approximation of the relationship between the optical properties of base sheet paper and the process variables. The experiments have been very successful and Pearson correlation coefficients of up to 0.98 have been obtained for the approximation. The objective of this thesis is to present a methodology for fine-tuning the parameters of radial basis function (RBF) neural networks, thus improving their performance. Three main parameters affect the performance of an RBF network. They are the centers and widths of the RBF nodes and the weights associated with each node. A gridded center and orthogonal search algorithm have been used to initially determine the parameters of the RBF network. A parameter tuning algorithm has been developed to optimize these parameters and improve the performance of the RBF network. When necessary, the recursive least square solution may be used to include new nodes to the network architecture. To study the behavior of the proposed network, six months of real data at fifteen-minute intervals has been collected from a North American pulp and paper company. The data has been used to evaluate the performance of the proposed network in the approximation of the relationship between the optical properties of base sheet paper and the process variables. The experiments have been very successful and Pearson correlation coefficients of up to 0.98 have been obtained for the approximation. The objective of this thesis is to present a methodology for fine-tuning the parameters of radial basis function (RBF) neural networks, thus improving their performance. Three main parameters affect the performance of an RBF network. They are the centers and widths of the RBF nodes and the weights associated with each node. A gridded center and orthogonal search algorithm have been used to initially determine the parameters of the RBF network. A parameter tuning algorithm has been developed to optimize these parameters and improve the performance of the RBF network. When necessary, the recursive least square solution may be used to include new nodes to the network architecture. To study the behavior of the proposed network, six months of real data at fifteen-minute intervals has been collected from a North American pulp and paper company. The data has been used to evaluate the performance of the proposed network in the approximation of the relationship between the optical properties of base sheet paper and the process variables. The experiments have been very successful and Pearson correlation coefficients of up to 0.98 have been obtained for the approximation.

Share