稳健逆传播算法用于预测混合物辛烷值和表面张力

采用中位数对训练过程中个别样本的训练残差进行了剪裁，消除奇异点或数据污染点的影响，从而使得训练模型更真实，泛化能力更强。改造后的算法称为稳健逆传播算法（ＲＢＰ），将这种算法用于推测混合物的辛烷值和表面张力，并与ＢＰ算法进行了比较，证明改造后的ＲＢＰ算法在泛化能力方面优于传统ＢＰ算法。

OCTANE NUMBER AND SURFACE TENSION PREDICTIONS BY USING ROBUST BACKPROPAGATION ALGORITHM

Neural networks were often trained by using backpropagation learning algorithm. In this methodology, sum of squared errors was employed to direct the training phase, this stratagem may have good results both for training and predicting under the normal distribution of errors, but in the practical problems, the error distribution is unknown and in most cases the error distribution may deviate from normal distribution which are caused by outliers or contaminated data points. In such situations, the trained models may straw away from the real model and overtraining is ofter encountered, i.e. the training may have a small error but the prediction may have an error large enough to be accepted. To deal with this defect, residuals of some training samples were cut down by using median to dispel the influence of outliers or contaminated data points. This may lead to a more realistic model and generalization ability. The reformed algorithm is called robust backpropagation learning algorithm (RBP) and was used to predict the octane number and surface tension of mixtures, the results showed that the RBP algorithm is superior to BP algorithms.