Optimization And Application Research Of Neural Network Based On PSO Algorithm

Artificial Neural Network is a kind of large-scale parallel distributing system, it is characterized by self-learning, self-organizing, self-adapting and non-linearity. Artificial Neural Network has mature application in many areas. Utilizing modern science and new theories to study it is a currently research hotspots.Particle swarm optimization (PSO) algorithm is a kind of optimization algorithm based on the theory of swarm intelligence. Because of its simple concept and easy implementation with only a few tuning parameters, it has been applied in lots of study as soon as it was proposed . In this paper, we utilize PSO algorithm to optimize RBF network and SOM network. Radial Basis Function (RBF) Network is a feedforward network with three layers. Different from other network, it is a nonlinear network. So, it has strong generalization ability and good approximation capability. Self-Organizing feature map (SOM) network has the characteristic of TopologicallyConsistency and Vector Quantization, any pattern in input layer can cast to single-dimensional or bidimensional discrete graph, so it is very useful in pattern recognition. Characteristics, work principle, and parameters of PSO algorithm are analyzed in detail in the paper, Stand PSO algorithm also. Finally, we selected the Stand PSO algorithm to improve the learning algorithm of neural network. The learning algorithm of RBF neural network is analyzed in detail in the third chapter. Then, using PSO algorithm improves the procedure of determining the number of hidden nodes in RBF network,and presents a new RBF network algorithm for function approximation. And the new algorithm is effective in function approximation.The main learning algorithms of SOM neural network are introduced in detail in the fourth chapter of the paper. Later, against the problem of dead cluster, PSO algorithm is applied to the adjustment algorithm of weights in SOM neural network. The new algorithm can effectively reduce waste of network resources. The new algorithm is applied to the algorithm of manifold learning and analyzed by experiment at the last part of the paper.