Recurrent Fuzzy Neural Networks and Their Performance Analysis
Rafik Aziz Aliev, Bijan Fazlollahi, Babek Ghalib Guirimov, Rashad Rafik Aliev · InTech eBooks · 2008
In spite of great importance of fuzzy neural networks for solving wide range of real-world problems, unfortunately, little progress has been made in their development. In this study we have discussed recurrent neural networks with fuzzy weights and biases as adjustable parameters and internal feedback loops, which allows capturing dynamic response of a system without using external feedback through delays. In this case all the nodes are able to process linguistic information. As the main problem regarding fuzzy and recurrent fuzzy neural networks that limits their application range is the difficulty of proper adjustment of fuzzy weights and biases, we put an emphasize on the RFNN training algorithm. We have proposed the standard DEO-based method for learning of recurrent fuzzy neural network. The optimization method, customized for RFNN training, compares favorably with the existing gradient-based error minimization method as it is less complex and is more likely to locate the global minimum of network error. As the method does not require derivative information, it is very effective in case when dealing with different distance functions. Also, the considered global optimization algorithm can provide high accuracy of fuzzy mapping with relatively smaller network size. The RFNN was tested on a number of benchmark identification and time-series forecasting problems well-known in the literature as well as on application problems. Experimental results demonstrated very good performance on all considered problems.