摘要
在基本反向传播(BP)算法中,学习速率往往固定不变,限制了网络的收敛速度和稳定性。因此,提出一种动态调整BP网络学习速率的算法,以BP网络输出层节点的实际输出值与期望输出值的平均绝对值误差及其变化率为自变量,找出学习速率与两个自变量之间的函数关系。根据网络的实际学习情况,对学习速率进行动态调整。实例仿真结果表明,改进的BP算法在保持网络稳定性的同时,具有更快的收敛速度。而且,该算法只需恰当地选取几个参数,不受条件限制,因此具有普遍的适用性。
The learning rate is usually invariable in basic Back Propagation （BP） algorithm, thus the constringency rate and stabilization of network are constrained. Therefore, a BP algorithm based on dynamically adjusting learning rate was proposed. The average absolute error between actual output value of the output layer node and the expected output value and its change ratio were regarded as independent variables, then the function relation of learning rate and two independent variables was found. According to the actual learning circumstance of network, the learning rate was adjusted dynamically. Through the instance simulation, the improved BP algorithm is of more fast constringency rate while keeping the good stabilization than basic BP algorithm. Further more, the algorithm can select the appropriate number of parameters without any condition, and it is therefore of general applicability.
出处
《计算机应用》
CSCD
北大核心
2009年第7期1894-1896,共3页
journal of Computer Applications
基金
陕西省自然科学基础研究计划项目(SJ08-ZT14)
西安市科技计划项目(CXY08017(1))
关键词
反向传播算法
学习速率
动态调整
平均绝对值误差
变化率
Back Propagation （BP） algorithm
learning rate
dynamic adjustment
average absolute error
change ratio
