中文摘要：

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）．

英文摘要：

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.