Study On Improved Algorithm And Application Of BP Neural Network

Artificial neural network is a nonlinear system which simulates cerebrum information processing algorithm. It has powerful distributed information storage, parallel processing and adaptive learning ability. BP network contains the most essence part of neural network theory. Owing to simple structure and technical mature, it has been widely applied in pattern recognition, intelligent control and other areas. However, BP algorithm uses the steepest descent algorithm, thus there are two main shortcomings of slow convergence rate and easy to fall into local minimum.Firstly, we expound back-propagation networks systematically in this paper. For the platform phenomenon of standard back-propagation algorithm, an adaptive magnified error signal is constructed. The derivative of the activation function is modified to make the weight adjustment avoid falling into the saturation areas. Then the theorem for the convergence of the proposed algorithm is presented and proved.Secondly, considering the comprehensive factor of influencing BP networks performance, this paper presents an adaptive learning rate adjustment and dynamic adjustment S-type activation functions combination of improved BP algorithm. The proposed algorithm connects the learning rate with the error function, and the slope of the activation function of each hidden and output unit is automatically adjusted.Finally, combining genetic algorithm which is good at overall search with BP algorithm which has much strong local optimizing ability, a new BP network training method based on improved genetic algorithm is designed. It uses the hierarchical code, adaptive crossover and mutation, pruning similar individuals, dynamic supply new individuals and other operations, so the network structure and weight are optimized at the same time.In view of the BP algorithm's flaw, this paper makes the improvement from three different perspectives to the standard BP algorithm. The simulation results show that the improved algorithms have quick convergence rate, strong optimization and generalization ability, and good application and utility.