基于三次样条函数拟合的过程神经元网络训练

过程神经元网络的提出为大样本识别问题开辟了新途径，但其训练方法目前主要基于权函数正交基展开。这种方法基函数个数选取目前尚无理论依据。提出了基于三次样条函数拟合的过程神经元网络训练方法。首先将样本函数、过程神经元权函数的离散化数据拟合成分段表示的三次样条函数，然后计算样本样条函数与权值样条函数乘积在给定采样区间上的积分，并将此积分值提交给网络的过程隐层神经元，输出层由普通神经元组成。三次样条函数具有很好的光滑性、可积性、阶数低、参数少等优点，有效地简化了网络的时空聚合运算。实验表明该方法是可行的。

Training of procedure neural network based on spline function

It is procedure neural network (PNN) that inaugurates new approach for enormous samples recognition. At present, the PNN study method is mostly based on weight function expanding by vertical base function. How to determine the count of base function has not theory to fellow. A training method of PNN based on spline function is presented. First the discrete data of both sample function and procedure neuron weight function is converted to spline function, and then spline function product integral of both sample and weight in the sampling time is calculated, and the integral as a result of aggregation is submitted to procedure neuron of PNN hide layer. The output layer of PNN consists of general neuron. The advantage of lubricity, calculability, lower exponential and less parameter that spline function is presented simplifies aggregation operation of PNN in both space and tine. Finally an experiment example is proposed to illustrate the availability of the study method.