On the construction of wavelet network
Date of Issue2006
School of Computer Engineering
Centre for Computational Intelligence
By taking advantage of both the scaling properties of wavelets and the high learning ability of neural networks, a wavelet network exhibits a high approximation and prediction capability. It can well approximate data, and thus be a good mathematic tool to find the hidden nature of a sequence of data, and to visualize the behavior of the data. According to Ocam's Razor hypothesis, compact models are believed to have the best generalization ability, since they represent data with the lowest structural complexity. Therefore, this research aims to construct a compact wavelet network, which can achieve an ideal approximation performance with the least number of wavelets. There are three steps involved to construct a wavelet network. First, a wavelet candidate pool is built up with a pyramid type initialization method. After that, the orthogonal least square (OLS) algorithm is employed to find those most critical wavelets. Finally, a stochastic gradient algorithm is utilized to recursively adjust the parameters.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
Nanyang Technological University