人工神经网络D最优设计的一种方法

A Method of D Optimum Design for Artificial Neural Network

利用初始的少数实例求初步的权值矩阵,根据该矩阵的人工神经网络模型找出偏导数向量的模最大时对应的设计矩阵,又根据该设计矩阵的实例求出新的权值矩阵,寻找新一轮设计矩阵。通过用泰勒级数展开的神经网络输出公式,求该输出函数的方差,证明了偏导数的向量的模最大是相关矩阵行列式最小(D最优)的充分条件。

The few initial examples are used to calculate preliminary weight matrix, according to the artificial neural network model of this matrix, the design matrix is found out which corresponds to the maximum modulus of the partial derivative vector of the model. According to the example of this design matrix, a new weight matrix is computed and a new design matrix is sought. With output formula of the neural network that is spread out with Taylor progression, the variance of the output function is deduced, it is proved that the maximum modulus of the partial derivative vector is the sufficient condition of the minimum determinant of correlation matrix ( D optimum).