Optimazation And Applizaction Of Self-organizing RBF Neural Network Based On PSO

Radial basic function neural network, a feedback neural network, has been widely applied in the optimization technology, intelligent control, and pattern recognition and so on. However, the structure of RBF neural network is established by experience or trying, and the structure will no longer be adjusted when it was established. Thus, the adaptive ability of the neural network has been greatly reduced. Therefore, how to obtain better performance in the optimization of parameters, adaptive adjustment of network structure, comprehensive analysis on the influencing factors of generalization capability, and improving the overall performance of the RBF neural network, has been a further problem.In order to solve the structural dynamic optimization problems of the RBF neural network, this paper deeply analyzed the RBF neural network features and designed an adaptive particle swarm optimization(APSO) algorithm so as to obtain the APSORBF neural network structure adjustment method. In order to further improve the generalization ability, an adaptive gradient multi-objective particle swarm optimization(AGMOPSO) algorithm is improved and an AGMOPSO-SORBF neural network optimization algorithm is designed. At the same time, the convergence of RBF neural network structure in dynamic adjustment process is analyzed. Paper main research work as follows:1. The design of APSO algorithm. To solve the problem that particle swarm optimization(PSO) algorithm have premature convergence and fall into local convergence, this paper proposed an adaptive particle swarm optimization algorithm. According to the diversity of group and the flight information of individual particle, the flight parameters are adaptive adjusted, which greatly avoid that particles are trapped in local convergence, and balance the exportation and exploitation ability. The experimental results show that the proposed APSO has higher precision comparison with other improved PSO.2. The design and applied of APSO-SORBF. In view of the RBF neural network parameter optimization and structural dynamic adjustment problems, we design an APSO-SORBF neural network. The parameters of RBF neural network(center, width and connection weights) are considered as particle position, and the particle dimension is mapped to the number of hidden layer neurons, thus the structure and parameters optimization of RBF neural network can be achieved at the same time. The experimental results of nonlinear system modeling show that the proposed self-organizing mechanism can optimize the structure of RBF neural network, and obtain higher prediction accuracy. Finally, the APSO-SORBF is applied to establish total phosphorus soft measurement model, and higher prediction accuracy is achieved.3. MOPSO algorithm design. In order to eliminate the conflict between the factors influencing the generalization ability of RBF, a MOPSO algorithm is designed. Then, an AGMOPSO algorithm is designed to enhance the ability of local search based on multi-objective gradient(MOG) and balance the global and local search through the proposed adaptive flight parameters strategy according to the flight information of particle. Experimental results show that AMOPSO has better accuracy and convergence rate.4. AGMOPSO-SORBF neural network design and application. In order to improve the generalization performance of self-organizing RBF neural network, an AGMOPSO-SORBF neural network is proposed. In this AGMOPSO-SORBF neural network, not only the neural network parameters and structure are optimized, but also the complexity, training accuracy and the smoothness of connection weights as the optimization target are simultaneously optimized to improve the generalization performance of this neural network. Nonlinear system modeling results show that the proposed AGMOPSO-SORBF neural network can optimize the structure and parameters of RBF neural network and obtain higher prediction accuracy.