基于熵值法的BP网络输入变量加权分层方法研究

Weight analysis based on the information entropy research on the inputs of ANN

当BP网络模型的输入变量包括多个类别时,如果其中几类变量的个数远多于其它类别的变量,变量多的这几类会削弱其它类变量对输出变量的影响,导致模型预报误差增大。提出BP网络输入变量加权分层的改进方法。根据熵值法模型对每个类别包含的所有变量按其重要程度加权平均,得到代表各类的综合影响指标,将这些综合影响指标作为BP网络模型的输入变量得到模型预报结果。改进后的模型更全面合理地考虑了各类输入变量的变化对输出变量的影响,发展了神经网络的应用理论。实例计算表明,模型预报精度得到明显提高。

When the input includes several regimentations and the number of some variables in some regimentations is much more than that of the other regimentations, the former will weak the latter's effect on the output, which leads to the augment about the forecasting error of the model. The entropy based the self- accommodation back propagation neural model is introduced to solve this problem, in which the several variables of each regimentation are weighted according to their importance, so each regimentation is turned into one input respectively in the back-propagation (BP) net work model. The improved model can take the all kinds of inputs into account entirely and reasonably, and boost the forecast accuracy, which develops the applied theory of the neural network.