Abstract
A new error back propagation method which is called EBP learning algorithm has been proposed by the one of the authors for a supervised learning of multilayer neural networks. This method is free from a gradient method which is used in the well known BP method or its improved versions. The method is composed of the following two stages.
A) Error back propagation: Determination of fictitious teacher signals of hidden layers from the output error of a neural network in the backward direction.
B) Weight parameters determination: Weight parameters of a neural network should be determined to minimize the fictitious error using fictitious teacher signals of each hidden layer.
An exponential weighted least squares (EWLS) method which is well known in the field of signal processing is proposed in this paper to use for a determination of weight parameters of B). Simulation results using a proposed EBP-EWLS algorithm for a learning of a sinusoidal function are presented. The results show the advantge of learning numbers for success and 100% convergence rate for initial values of weight parameters of a neural network.
