BP Neural Network Optimization Based On Improved Particle Swarm And Its Application

BP neural network has good performance on nonlinear mapping generalization and fault tolerance. Since it is based on gradient descent BP algorithm, it has some defects, such as slow convergence, easy to fall into local minima and so on. These defects seriously affect the BP network. PSOA is a new optimization method based on the population behavior of swarm intelligence optimization algorithm. PSOA is a probabilistic search and does not require any gradient information. PSOA posses the properties that are robust, distributed, adaptive fast, and the collaborative, and good global convergence. Therefore, PSOA is chose to optimize the BP neural network, which can improve the performance of net word.This paper takes some research on the modified PSOA and applies the improved PSOA into optimizing BP neural network. The main research contents and results are as follows:(1) A mathematical derivation of the standard BP algorithm is taken. The limitations of BP algorithm are pointed. Furthermore, the reasons causing the limitations are discussed. We also compare the improved algorithm with the existed BP algorithm.(2) The basic principles and limitations of PSOA are considered. By means of dynamic adjustment strategy of inertia weight and learning factors, the convergence speed and accuracy are improved. The approach also balances the algorithm’s global search ability and local search ability. In addition, by introducing locally optimal test strategy, replacing part of the particle, and increasing particle diversity, the feature of algorithm converging to the global optimum is guaranteed. Using benchmark functions to conduct simulation on the improved PSOA, the results show that the improved algorithm is effective.(3) The BP neural network topology is introduced in the improved PSOA. Instead of the gradient correction by particle swarm iteration, the global optimum particles is found by iterative algorithm, that is, the optimal weight threshold for the net word is found. Simulating the experiments by using XOR problem, the test results verify the correctness of the improved particle swarm optimization BP neural network. (4) Combining the factors and indicators system of the earthquake prediction and the selected sample data, an earthquake prediction model based on neural network is constructed by the improved particle swarm, which confirms the effectiveness of the improved particle swarm neural network prediction system.