Performance comparisons between RBF networks and multilayer backpropagation network

摘要

The application of neural network technology for pattern recognition and classification in the field of machine intelligence requires advanced computational procedures. The multilayer structure using conventional backpropagation (CBP) network has been used extensively. However, the drawback of slow convergence speed must be dealt with. We have proposed some new techniques to speed up its training process by properly initializing the connection weights. In addition, we have used radial basis function (RBF) networks to compare its performance with backpropagation networks and its variants. To maximize RBF's potentials, we have experimented with different procedures to improve the classification accuracy. By using multispectral data and artificially-created data, we are comparing the results of different models for performances in speed and generalization.