提高 BP 网络性能的一种方法

Method of Improving the Properties of BP Artificial Neural Networks

分析了ＢＰ网络在目标量的极值区存在较大误差的原因，提出了目标量扩展的数据标度化方法·检验结果表明，该方法不但可以消除目标量极值区的大误差现象、提高网络训练的收敛速度，而且还可以使网络具有外推预测功能，即同时提高了ＢＰ网络的训练性能和应用性能·

The reason that there existed big training errors in the extremum areas of the target parameters was analyzed and a new algorithm with target parameters value expansion in scale conversion was proposed. The results of training with the new algorithm showed that the big error phenomena was eliminated, and the rate of convergence of the network was increased. In addition, the network was endowed with extrapolation capability, which indicated that both properties of training and application were improved.