一种改进BP网络学习算法

针对BP神经网络的原始算法收敛速率慢、学习精度低、训练过程易陷入局部极小值问题,为解决上述问题,提出一种以变学习率BP算法为基础的改进算法,通过区分隐层和输出层的学习率,并用交叉熵作性能函数,提高算法的学习精度和训练速度,并经过数学推导,得到改进箅法的实现公式.将改进算法应用于奇偶数判别问题进行仿真,仿真实验结果与其它类似的方法进行比较后,发现改进算法大大降低了网络迭代次数,缩短了网络的训练时间,提高了训练精度,验证了该算法的有效性.

An Improved Learning Algorithm for BP Network

The primitive BP algorithm has low convergence speed and low precision.The training process is easy to fall into the partial minimum.To solve this problem,based on a BP algorithm with variable learning rate,this paper gives an improved algorithm.Through distinguishing the learning rate of hidden layer from that of output layer,and using cross-entropy as the function of performance,the algorithm learning precision and the training speed are increased.The comparison with other similar methods finally indicate...