双隐层BP网络的设计与逼近分析

Design and approximation of double hidden layer BP network

为了设计一个实用、高精度的BP网络算法,对双隐层BP网络进行了深入研究。从对双隐层BP网络的设计入手,给出了设计该网络的所有关键因素,并且推导了两输入单输出类型的双隐层BP网络的学习算法。为了分析双隐层BP网络逼近能力,选取了4个具有代表性的函数,同时使用双隐层BP网络和单隐层BP网络进行逼近实验,从而比较双隐层BP网络与单隐层BP网络。实验数据表明,双隐层BP网络具有更高的训练样本的输出精度和更强的泛化能力,为BP网络的使用提供了另一条路径。同时也表明了超过两个隐含层的BP网络的实用性可能不太大。

In order to design a practical algorithm of BP network with high precision,it makes a deep research on double hidden layer BP network（ DHL-BP）. It describes the concept of double hidden layer BP network,analyzes all key factors in design process of this kind of network,derives the learning algorithm for two inputs with single output DHL-BP. In order to analyze the approximation ability of DHL-BP,it selects 4 typical functions and tests with DHL-BP and SIL-BP. It compares the test results and shows that DHL-BP has higher training sample output precision and stronger generalization ability. This provides another way for the use of BP network.