| Autor: |
Kacprzyk, J., Castillo, Oscar, Melin, Patricia, Ross, Oscar Montiel, Sepúlveda Cruz, Roberto, Pedrycz, Witold, Kacprzyk, Janusz, Jilani, Tahseen A., Burney, Syed Muhammad Aqil |
| Zdroj: |
Theoretical Advances & Applications of Fuzzy Logic & Soft Computing; 2007, p730-740, 11p |
| Abstrakt: |
A new method for modeling and knowledge extraction at each neuron of a neural network using type-I fuzzy sets is presented. This approach of neuron modeling provides a new technique to adjust the fuzzy neural network (FNN) structure for feasible number of hidden neurons and efficient reduction in computation complexity. Through repeated simulations of a crisp neural network, we propose the idea that for each neuron in the network, we can obtain reduced model with high efficiency using wavelet based multiresolution analysis (MRA) to form wavelet based fuzzy weight sets (WBFWS). Triangular and Gaussian membership functions (MFs) are imposed on wavelet based crisp weight sets to form Wavelet Based Quasi Fuzzy Weight Sets (WBQFWS) and Wavelet Based Gaussian Fuzzy Weight Sets (WBGFWS). Such type of WBFWS provides good initial solution for training in type-I FNNs. Thus the possibility space for each synoptic connection is reduced significantly, resulting in fast and confident learning of FNNs. It is shown that propsed modeling approach hold low computational complexity as compared to existing type-I fuzzy neural network models. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Supplemental Index |
| Externí odkaz: |
|
