BP 算法执行过程中的平台现象及其减少方法的研究

Methods of Eliminating Flat Phenomenon during BP Working

基于对ＢＰ算法执行中存在的平台现象的分析，从神经元作用函数的总输入、作用函数的选择、作用函数的自适应调整和网络初始权值的选择几方面，分析和研究了减少平台现象的措施提出了神经元作用函数的选择方法，给出了具有作用函数自适应调整的ＢＰ算法。

In this paper,the reasons of the flat phenomenon during BP working are analysed.The measures for eliminating the flat phenomenon are studied from several aspects.These aspects are:the input of neural network,the total input of neurons,the choice of activation function,the adaptive learning of activate characteristic of neurons,and the choice of initial weight values.The choice method of activation function is proposed.The BP algorithm with dynamic adaptive variable active characteristic of neurons is derived.Four measures for eliminating the flat phenomenon are gained with theoretical analysis and test verification.The tests indicate that these measures are effective.