综合反向传播算法

A Synthetically Backpropagation Algorithm

提出一种用于多层前向神经网络的综合反向传播算法．该算法使用了综合考虑绝对误差和相对误差的广义指标函数，采用了在网络输出空间搜索的反传技术，具有动态自调整学习率和动量因子，有神经元激活特性自调整、减少平台现象和消除学习过程中不平衡现象的能力．对比实验表明该算法有比基本ＢＰ算法快得多的收敛速度，并能取得全局最优解．

This paper presents a synthetically backpropagation algorithm for multilayered forward neural networks.A new general index function that consider the effect of absolute error and relative error on NN learing and performance and the backpropagation technique based on searching output space are proposed and used in the algorithm.The algorithm has both a dynamical adaptive regulation learning rate and a variable momentum coefficient,and has ability of self regulation active characteristic,eliminating flat phenomenon and convergence no equilibrium phenomenon during training.The contrast experiments indicate that the algorithm has more fast convergence speed than BP algorithm and can achieve a global optimal solution.