实现BP神经网络从理论到实践的跨越

Bridging up the neural network theory and the practical applications

BP神经网络理论研究集聚了大量优秀的研究人员，并取得了很大的成就，但由于BP网络的建模理论和实际应用问题建模条件之间存在实质性差异，有些理论研究成果不能直接应用于实际问题的建模研究中．本文主要针对如何确定合理的隐层节点数、避免出现“过训练”现象以及神经网络模型性能评价等问题进行分析与研究，构筑BP神经网络建理论与实际应用问题建模之间的桥梁，为建立可靠、合理的BP网络模型提供依据．

A large number of experts have clustered for the study of the BP neural network （NN） theory, and got great achievements. On other hand, the essential differences between the neural network theory and modeling conditions in the practical applications are obvious. Some modeling theories are not suitable for modeling on most practical applications directly. How to determine the reasonable number of neurons on hidden layer to escape from the local minimum and ＂over-training＂ and evaluate the performance of the NN-based model are discussed and explored in this paper. The principles and steps are established to set up robust and reliable NN-based model for practical applications by bridging up the neural network theory and the practical applications.