改善BP网络检验效果的研究

本文针对ＢＰ人工神经网络在处理连续变量时，学习效果很好，而检验效果不理想的情况下，提出三条改进检验误差的措施。１）检验样本的选择规则；２）引入虚拟样本；３）改变网络的拓扑结构、并实际应用于催化裂化主分馏塔优化操作建模的神经网络，收到很好的效果。这三条措施简便易行，可广泛适用于其他网络，以改善其检验效果。

Study on improving testing results of BP neural networks

ood results can be achieved in training process of artificial neural networks I notin the testing process when using the BP artificial neural network deals with continuousvariable problems. In this case, three measures that improve testing results: the rule ofchoosing testing data, introducing fictitious training examples and changing topologystructure of network,are proposed in this paper. They have been applied to the modellingoptimizing operation of Fluid catalytic cracking (FCC) main fractionator. The training andtesting results both are better than that of the common network without accepting themeasures.