摘要
提出了一种新的神经计算的方法,其最大的特点是在进行特定的权值调整时无须加入学习率。计算开始时先给出一组随机的权值作为被拟合多项式的系数,由这组权值得到的拟合点上被拟合多项式值与期望值的各误差绝对值之和。依据其小分数的负值直接来调整权值,使各误差之绝对值不断减小,通过反复迭代计算,最终当各误差都在规定范围内时便得到期望的权值。在实际的计算中利用这种方法进行曲线拟合能方便地求出拟合多项式的系数,便于编程,简化了计算过程。将这种方法应用到范德堡函数的多项式拟合中,得到了范德堡函数f=1+0.03237714η-0.04037678η2+0.00857881η3-0.00077693η4+0.00002604η5并画出了曲线。
A new neurocomputing method is presented. Its major characteristic is not need to use the learning rate, when it carries on the weight adjustment in the iteration process. In the beginning of computing, an arbitrary set of weights as coefficients of polynomial were first imputed. The absolute summation of the errors between the computed polynomial and the expectations on the matched points were obtained. A small and negative fraction of it was taken as the weight adjustment , so as to minimize each absolute error. After a lot of iteration computations, each error was converged to a specified very litter one, then the expected weights were finally obtained. Using this method, the optimum weights can be able to be extracted conveniently. As like this simplified computation process, the advantages is easy to compile the computer programming. We apply this method to the polynomial fitting for Van der Pauw function,a quite satisfactory result is finally obtained, plotting a corresponding curve.
出处
《电子器件》
CAS
2008年第3期1015-1018,共4页
Chinese Journal of Electron Devices
关键词
神经网络
学习率
多项式拟合
范德堡函数
neural network
learning rate
polynomial fitting
Van der Pauws function