一种非线性的BP学习算法

提出了一种非线性学习规则，以非线性函数th(x)取代传统学习算法中的线性函数x，来调整BP网络的连接权值和阈值。与传统的BP学习算法相比，其连接权值与阈值的调整量不仅与误差函数对连接权与阈值梯度的一次幂有关，而且也与梯度的高次幂有关。因此，克服了传统的BP学习算法过程中难以跳出局部极小值与收敛速度慢的缺点。模拟实验表明，该算法比传统的BP网络学习算法在学习时间和迭代次数方面都具有显著优势。

A Nonlinear Learning Algorithm for BP Networks

In this paper, a nonlinear learning regulation that adjusts weights and thresholds of BP networks with th(x)which replaces linear function x is proposed. In comparison with traditional learning algorithm for BP networks, the adjusted value of weights and neural thresholds is not only related to gradient's power 1 of error function to weights and thresholds but also related to gradient's higher order of that. So it overcomes properties of traditional BP algorithm, such as falling easily into local optimal solution and slower convergence speed. Simulation experiments indicate that the algorithm is more perferable than traditional BP algorithm in learning time and iterative times.