| Summary: | 博士 === 元智大學 === 通訊工程學系 === 99 === The radial basis function (RBF) neural network is regarded as a good method in many kinds of applications, including function approximation, classification, and prediction. However, there still exist some unsolved problems, including the inability of flat function approximation, the weakness for noise, and the tradeoff between the network performance and the network size. In this thesis, RBF neural networks are improved by sigmoid function, M-estimator, and the growing and pruning algorithm (GAP). The proposed improved RBF networks adopt the sigmoid function as their kernel due to its increased flexibility over the Gaussian kernel. Furthermore, this thesis presents an M-estimator based RBF learning algorithm. The Welsch M-estimator and median scale estimator are employed to get rid of the influence from the noise. Finally, the GAP adjusts the network size dynamically according to the neuron’s significance. To evaluate the network performance, the improved RBF neural networks were applied to three applications, noisy time series prediction, liver mass classification, and vision-based handwriting recognition system. The experimental results show the proposed GAP algorithm is able to dynamically adjust the number of neurons to approach an appropriate size of the network. Moreover, for the noisy time series prediction, the M-estimator based RBF learning algorithm eliminates the influence of noise. For other applications, the experimental results show that the proposed sigmoid kernel is efficient for classification problems.
|
|---|
