An Algorithm of Wavelet Network Learning from Noisy Data

Noise often leads to bad generalization of network. Many of the typical algorithms can not be applied for the on-line identification of complex system since they are not robust to the variance of the energy of noise. A new algorithm is proposed to solve this problem based on the frequency band of wavelet network. It is shown that the wavelet network trained by the new algorithm is a low-pass filter, which has removed the noise out the frequency band of network. Since the frequency band of noise is usually higher than that of network, the variance of noise little influences the identification of wavelet network, so the new algorithm is robust to the variance of noise. The analysis of theory and the results of simulation show that the new algorithm has the capacity of avoidance of overfitting and the robustness of noise