论文部分内容阅读
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．
A comprehensive backpropagation algorithm for multilayer feedforward neural networks is proposed. The algorithm uses a generalized index function that takes into account the absolute error and relative error synthetically. It adopts the backtracking technology in the network output space search. It has the dynamic self-tuning learning rate and momentum factor, self-adjusting of the activation characteristics of neurons and reducing the platform phenomenon Ability to eliminate imbalances in the learning process. Comparative experiments show that this algorithm has much faster convergence speed than the basic BP algorithm and can obtain the global optimal solution.