Wavelet neural networks employing over-complete number of compactly supported non-orthogonal wavelets and their applications
T. Yamakawa, E. Uchino, Takashi Samatsu · 2002
This paper proposes two types of new neuron models, WS neuron (wavelet synapse neuron) and WA neuron (wavelet activation function neuron), which are obtained by modifying a traditional neuron model with non-orthogonal wavelet bases, while Boubez et al. (1993) employed orthonormal wavelets. Four types of typical wavelet neural networks employing WS and/or WA neurons are discussed. The simplest wavelet neural network exhibits much higher ability of generalization and much shorter time for learning rather than a three-layered feedforward neural network. Furthermore the wavelet neural network is guaranteed to give the global minimum. Other three wavelet neural networks are examined for predicting chaotic behaviour of a nonlinear dynamical system. The performance in learning speed and prediction of wavelet neural networks are more significant than a four-layered feedforward neural network.>