Neural Network Learning

It is known that Neural Networks (NN) have been successfully used to solve many problems [117,149,150,2,4,7,50,128,10] that could not be solved by, say, mathematical algorithms. However, performance depends not only on NN properties but alos how well we are modelling our system. One of the motivations behind the use of NN is that they are very efficient in learning and predicting nonlinear and complex functions. In addition, we do not need to know the details of the mathematical model of the system under study, which in some cases is very difficult to do or completely unavailable. The aforementioned problems are complex, nonlinear and, the case of network traffic, nonstationary. NN has been found to be powerful method to efficiently describe a real, complex and unknown process with nonlinear and time-varying properties. Our study and work is based on NN. We have used two kinds of NN: Artificial NN (ANN) [117,118,149,150] and Random NN (RNN) [49,89,48,45,101,51]. We have compared the performance of both of them. We have found that RNN have several advantages over ANN. However, a major disadvantage of RNN is that the available training algorithm is very slow and not efficient. This motivated us to propose new training algorithms for RNN to overcome the problem.