Research On Learning Algorithm Of Neural Network Optimized With PSO And Its Application

The traditional neural network training algorithms have many drawbacks, for example, it converges slowly and is easy to fall into local optimum. In recent years, some novel swarm intelligence optimization algorithms such as particle swarm optimization (PSO) were proposed. Generally, they have good global convergence performance and can be used to train the parameters and architecture of neural network. Also, some improved PSO algorithms were proposed by many researchers, but they still have different defects, respectively. The thesis firstly summarized the features of the existing methods and then proposed several modified algorithms in order to maintain the diversity of particle swarm and mine the prior information implied in the objects. Meanwhile, the proposed algorithms were applied to process DNA microarray data.The main work of the thesis includes four aspects as below:1) For many existing PSO algorithms, the particle swarm diversity can not be maintained well in the iteration process. To address this problem, an improved staging mutation particle swarm optimization algorithm (SMPSO) was proposed and used to train the parameters and architecture of neural network. SMPSO mutates particles that have too low fitness at early stage and later mutates individual extreme and global extreme that stagnate in excessive iteration. The diversity of particle swarm is always kept within reasonable range. The experiment results show that neural network optimized by SMPSO is more efficient than that optimized by traditional methods. It has great improvement in the convergence velocity of training and the accuracy of classification.2) The prior information implicated in the problem object can be abstracted and coupled into the training algorithm of neural network. It may speed up the convergence and improve the processing accuracy of model. Considering PSO algorithm coupled with prior information is rare until now, some research work for function approximation is first done in this paper.For function approximation problems, two neural network training algorithms optimized by PSO coupling with prior information were proposed. After the two function characteristics were converted into the equivalent mathematical expression, they were coupled into PSO for training neural network. The experimental results show that the prior information can make the algorithm converge rapidly and improve the approximation accuracy. In addition, the approximation effects of different functions are unequal by two kinds of prior information.3) For classification problem, two kinds of neural network training algorithms optimized by PSO coupling with prior information were proposed according to the number of samples. First, the prior information is abstracted by the bayes method and coupled into PSO to train the parameters of BP neural network. Second, the support vector machine (SVM) is adept in dealing with small-scale samples, thus it can be combined with PSO. The prior information is abstracted according to SVM and coupled into PSO to train the parameters of RBF neural network. Simultaneously, PLS is used to optimize the architecture of network. The experimental results prove that the prior information can narrow the initial search space and guide the process of particles flight, thereby enhancing the convergence rate and classification accuracy.4) DNA microarray data have some features, such as high dimension and small samples. The neural network model optimized by PSO coupling with prior information abstracted from small sample was used as the classifier of microarray data. The test results on several data sets validate that the proposed algorithm has clear superiority for processing microarray data.