基于一类非线性特性的FNN训练算法

Training Algorithm for FNN Based on a Class of Nonlinear Property

针对ＢＰ算法收敛缓慢和易于陷入局部极小的缺点，将基于一类非线性特性的动量项引入ＢＰ算法的梯度搜索，提出前向神经网络（ＦＮＮ）的一种通用且简单的全局训练算法（ＩＢＰＭ 算法）。结合升温策略，算法在优化精度和训练速度方面有较大的改善。

To solve the slow convergence of BP algorithm and avoid getting stick in local minima, a general and simple training algorithm for feed-forward neural networks(FNN), named IBPM, which combines a class of nonlinear property of momentum term with the gradient descent of BP algorithm, is presented. Using “temperature-raising” strategy, convergence rate and training speed of the algorithm are greatly improved. Simulation results verify the efficiency of such algorithm, and some conclusions to choose parameters are provided.