一种综合改进的BP神经网络及其实现

从网络拓扑结构、增加动量项、增加学习速率自适应调整、增加权重初始化的N－W法等几个方面讨论了BP神经网络的改进，通过这几种方法的组合实现了一种新的改进型网络。详述了网络结构参数如学习速率初始值、隐含元个数、网络收敛精度等的确定原则和方法，并针对某一问题进行了计算分析。计算中重点分析了学习速率和隐含元个数的确定、各种单一改进方法和综合改进方法的比较、网络训练样本量的取舍等。分析表明改进方法是有效果的。

A Kind of Comprehensively Improved BP Neural Network and Its Realization

The improvement method of a neural network has been stated in aspects of network topology, momentum, adaptive learning rates and Nguyen Widrow method of the weights initialization. The new improved BP neural network has been realized by the combination of these methods above. The rule and way to determine the network parameters such as the initial data of the learning rate, number of neuron in hidden layer, the convergence accuracy of the network have been analyzed. In the numerical example the determination to the learning rate and number of neuron in hidden layer, the comparison to the different improved methods and the selection to quantity of the training samples have been discussed. By the analysis it states that the improvement is feasible.