提高前馈神经网络学习效率的学习算法探讨

Improvement on the efficiency of the learning algorithm for multilayer feedforward neural networks

简要分析了最速下降法（Ｓｔｅｅｐｅｓｔ Ｄｅｓｃｅｎｔ Ａｌｇｏｒｉｔｈｍ ，即ＳＤＡ 法）和正交校正共轭梯度法（ＣＧＭ－ＯＣ法）的优缺点，提出了一种进行多层前馈神经网络学习的新算法，即ＳＤ－ＣＧＭ－ＯＣ算法．该算法结合最速下降法与正交校正共轭梯度法的特点，在文中所述实际问题构造模型的基础上，论证了ＳＤ－ＣＧＭ－ＯＣ算法比传统的ＢＰ算法具有更高的学习效率和二次收敛率．实验结果验证了该学习算法的有效性．

This paper begins with a simple analysis of the advantages and shortcomings of the Steepest Descent Algorithm (SDA) and the CGM OC method, and proposes a new learning algorithm for multilayer feedforward neural networks. The new algorithm, that is, SD CGM OC Algorithm, combines the advantages of both SDA and CGM OC method, which make use of the construction of the practical problem. It gives higher learning efficiency compared with the BP algorithm, and what is more, its convergence is quadratic. The effectiveness of the SD CGM OC Algorithm is shown by the experimental results.