复数前馈神经网络的全局最优和快速学习算法  

Algorithm with global minima and high learning speed for training complex-valued feedforward neural networks

作　　者：张代远[1] 

机构地区：[1]南京邮电大学计算机学院,江苏南京210003

出　　处：《电路与系统学报》2006年第4期9-15,共7页Journal of Circuits and Systems

摘　　要：提出了一种新型复数前馈神经网络的学习算法。当输入层和隐层之间的权值计算出来后,就可以通过求解线性方程组得到隐层和输出层之间的权值。这些权值是全局最小点。另一方面,本文算法很容易确定全局最小点时隐层神经元的个数。本文算法具有很高的训练精度和学习速度。A novel algorithm is proposed for training complex-valued neural networks. The complex-valued weights between hidden and output layer are updated by solving linear system based on finding the complex-valued weights between input and hidden layer. These weights are global minima. On the other hand, the number of neurons in hidden layer can be obtained ve.ry easily for finding the global minima. The new algorithm has good performance both in high training accuracy and learning speed.

关 键 词：复数神经网络 全局最小 高训练精度 快速学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP18[自动化与计算机技术—控制科学与工程]