Research On Learning Algorithms Of Complex-Valued RBF Neural Networks

In recent years,the complex-valued signals are becoming ubiquitous in many areas.As an attractive option for dealing with complex-valued signals,the complex-valued neural network has become an important branch in the research field of the neural network.Radial basis function neural networks have the characteristics of simple structure,efficient network training process,excellent approximation performance and faster convergence speed,thus it becomes the one of the most commonly used neural network.This thesis carries out the research of complex-valued radial basis function and its algorithms,and several improved algorithms are proposed.First,we propose a distance-based hidden layer center initialization algorithm.This algorithm selects centers from all of the training samples,uses the distance between samples of different classes to express between-class scatter,takes the product of the average distance and variance between samples from the same class to represent within-class scatter,according to the ratio of between-class scatter and within-class scatter to measure whether the sample is chosen to be a center.At the same time,in order to make full use of the advantages of the complex-valued neural network,this paper applies the complex LM algorithm to train the full complex-valued radial basis function network.Second,this thesis also proposes an improved clustering center selection algorithm,all the training samples are taken as the candidate set of a hidden layer.When a new center is obtained,we find the samples could be covered by the newly added neuron,then the covered samples will be put back to the candidate set,can still participate in the center selection in the next epoch.This mechanism can make full use of the information contained in the training sample set.Once the center and width of the hidden neuron are determined,the weights of output layer can be calculated by the least square method.We apply the above algorithms to real-valued classification in the public database of UCI and gesture recognition in complex domain.Besides,thesis summarizes several methods of mapping real-valued inputs to the complex domain.Experimental results demonstrate the effectiveness of these proposed methods.