Study On Evolutionary RBF Neural Network Classifier

This dissertation is devoted to a study of classification technology based on evolutionary RBF neural network while laying emphasis on its application to modulation classification and channel equalization of communication signals, including feature selection of modulation signals for classification based on genetic algorithms, the initialization of RAN (Resource-Allocating network) based on genetic algorithm, the design and implementation of modulation classifier based on evolutionary RBF neural network(RBFNN), and the design and implementation of channel equalizer based on evolutionary RBF neural network. The work finished in this paper is a part of a large scale army engineering project undertaken by the laboratory the author works with.The main achievements and innovations are summarized as follows:1. The genetic algorithm, one of the most important components of evolutionary theory, and its application in feature selection for modulation classification is studied in depth. Two improved genetic algorithms based on family competition are proposed. One of which is based on the population median-value that can adaptively change the probability of the crossover and the mutation. And the other is based on the information of the infeasible solution and the information of unsubmissive degree of the individuals, which can adaptively adjust the penalty coefficients. Both approaches can effectively coordinate the contradiction between the population diversity and the selection pressure, and quickly converge to the global optimal solution. Because only the selected feature subset is taken as the inputs of the RBF neural network modulation classifier, the adverse affects of redundant features on the classifier accuracy is avoided and the computational complexity is also greatly reduced.2. Based on genetic algorithms, two improved initialization algorithms for RAN (Resource-Allocating network) are proposed, which pre-cluster the input data of the RBF neural network and set the clustering center as the initial center of RAN algorithm. One improved algorithm is the pipelined genetic algorithm (PLGA), which can accelerate the convergence of the RAN algorithm. The other is to encode the 4QAM signal by Diploid genotypes, which initialize the complex-value RAN algorithm and improve the performance of complex-value RBF neural network equalizers. The two algorithms decrease the sensitivity of RAN algorithm to abnormal data and noises, thus improve both the accuracy of RBFNN modulation classifier and the performance of complex-value RBFNN equalizer on severe channels.3. A new improved fitness function based hierarchy genetic algorithm is proposed for the design of RBF modulation classifier. In conventional design, single-object weighted sum function is chosen as the fitness function, and several parameters need to be properly valued at first. Since the fitness function takes the complexity and accuracy of the network into consideration, minor variations in value would affect the structure and performance of the classifier. The new fitness function has only one undetermined parameter, which can be set accurately according to the network applications. This lowers the difficulty and complexity of the design of hierarchy algorithm and improves the accuracy of the RBF modulation classifier.4. A new improved hierarchy genetic algorithm based RBF channel equalizer is given. This kind of algorithms usually converts the traditional equalization to a problem of classification. A novel method is proposed for the valuing of undetermined parameters in the fitness function of hierarchy genetic algorithm. By determining the minimal distance between the zero point and the unit circle, the value range of the undetermined coefficients in fitness function can be obtained, thus the design complexity of the hierarchy genetic algorithm is reduced. Besides, the algorithm isn't affected by empirical factors, making the evolutionary RBF equalizer applicable for various channels. A dynamic niche-area-based evolutionary genetic algorithm is proposed. Using the minimal Hamming distance of the controlling gene in encoding, the searching range of the sub-population in the niche can be adaptively adjusted, which overcomes the problems of high complexity and poor performance in the case of fixed niche area. The algorithm can provide several sub-optimal solutions in a single run. According to the clustering effectiveness decision, the optimal solution is chosen as the RBF nework equalizer structure.5. A new multi-objective evolutionary algorithm based on improved selection strategy is proposed and applied to design RBF equalizer. Only part of the elites is chosen to enter the next generation, the number of which gradually decreases as the evolutionary process goes on. This suppresses the fast convergence of the local solution of the evolutionary population and solves the premature problem in improved nondominated sorting genetic algorithm(NSGA-II). The obtained Pareto optimal solution set of the evolutionary would provide more straight and reliable choice for the design of RBFNN equalizer.Simulation results in each chapter validate the correctness, feasibility and the performance of the algorithms.