摘要
针对基于梯度下降法的BP算法提出一种改进方法:AF法。即在权值修正中,引入一个调整函数优化梯度。调整函数作用在激活函数的导数上,用于平缓和优化激活函数的导数运算。并就4元校验码和Feigenbaum函数逼近问题进行算法比较。实验显示,应用本方法的单隐层BP网就能迅速有效地解决异或问题和4元校验码问题。
In this paper, a new improved method： AF algorithm is presented to improve BP algorithm based on gradient methods. During adjusting the weights, the AF algorithm uses adjustment function to optimize derivative of activation function that is obtained by gradient methods. Then simulating results for 4 b parity problem and Feigenbaum function approximation are given for comparison with the improved BP algorithm,and the comparison results show that this method is useful for classification and function approximation.
出处
《现代电子技术》
2006年第16期102-105,共4页
Modern Electronics Technique
关键词
调整函数
BP算法
激活函数
导数
adjustment function
BP algorithm
activation function
derivative
