Legendre正交基前向神经网络的权值直接确定法

Weights immediate determination for feed-forward neural network with Legendre orthogonal basis function

为避免权值反复迭代修正的冗长BP训练过程,避免传统方法陷入局部极小点,根据多项式理论,构造了一种新型前向神经网络模型,推导了基于最速下降法的误差反传算法和基于伪逆的直接确定法.仿真结果显示,迭代方法和伪逆直接确定法都能达到比较高的工作精度(10-6).

To avoid lengthy iterative BP-training and its localminima problem, a feed-forward neural network was developed using Legendre orthogonal polynomial activation function based on polynomial interpolation and curve-fitting theory. The neural weights-updating formula was derived by adopting the standard BP training method, and then a pseudo-inverse based method was proposed. Simulation results show that one-step weights-determination method could efficiently achieve at least equally small computational error （10^-6） compared with iterative-training method.