Initialization by selection for wavelet network training

Abstract

We present an original initialization procedure for the parameters of feedforward wavelet networks, prior to training by gradient-based techniques. It takes advantage of wavelet frames stemming from the discrete wavelet transform, and uses a selection method to determine a set of best wavelets whose centers and dilation parameters are used as initial values for subsequent training. Results obtained for the modeling of two simulated processes are compared to those obtained with a heuristic initialization procedure, and demonstrate the effectiveness of the proposed method.

Wavelet neural networks: A practical guide

“…Recent studies proposed more complex methods that utilize the information extracted by the WA, (Kan & Wong, 1998;Oussar & Dreyfus, 2000;Oussar, et al, 1998;Wong & Leung, 1998;Xu & Ho, 2002;Q. Zhang, 1997).…”

“…However, in order to generate an orthogonal wavelet basis, the wavelet function has to satisfy strong restrictions, (Daubechies, 1992;Mallat, 1999). In addition the fact that orthogonal wavelets cannot be expressed in closed form constitutes them inappropriate for applications of function approximation or process modeling, (Oussar & Dreyfus, 2000).…”

“…J. Yao, Song, Zhang, & Cheng, 2000), in time series prediction, (Cao, et al, 1995;Chen, Yang, & Dong, 2006;Cristea, Tuduce, & Cristea, 2000), signal classification and compression, (Kadambe & Srinivasan, 2006;Pittner, Kamarthi, & Gao, 1998;Subasi, Alkan, Koklukaya, & Kiymik, 2005), signal denoising, (Z. Zhang, 2007), static, dynamic (Allingham, West, & Mees, 1998;Oussar & Dreyfus, 2000;Oussar, Rivals, Presonnaz, & Dreyfus, 1998;Pati & Krishnaprasad, 1993;Postalcioglu & Becerikli, 2007; Q. Zhang & Benveniste, 1992), and nonlinear modeling, (Billings & Wei, 2005), nonlinear static function approximation, (Jiao, Pan, & Fang, 2001;Szu, Telfer, & Kadambe, 1992;Wong & Leung, 1998), to mention the most important.…”

Artificial Intelligence Applications and Innovations III

“…Moreover this method does not use any information that the wavelet decomposition can provide. In literature more complex initialization methods have been proposed, [17] [14] [18]. All methods can be summed in the following three steps.…”

“…In [21] the RBS algorithm is used for the synthesis of a WN while in [17] an algorithm similar to SSO is proposed. In [18] an orthogonalized residual based selection (ORBS) algorithm is proposed for the initialization of the WN.…”