一种有效脱离平坦区的改进BP网络  

Modified BP Network in Breaking Away from Flat Areas Effectively

作　　者：刘乐[1] 王洪国[1] 王鑫[1] 王宝伟[2] 

机构地区：[1]山东师范大学管理与经济学院,济南250014 [2]山东师范大学信息科学与工程学院,济南250014

出　　处：《计算机科学》2008年第7期96-98,共3页Computer Science

基　　金：山东省自然科学基金(Q2006G03)

摘　　要：本文提出了一种新型权值调整规则,当神经元输出值接近于0或1时,确保足够的权值调整幅度,解决了标准BP(Back Propagation)网络在训练进入平坦区后难以摆脱,训练速度很慢的问题。并给出了利用单隐层BP网络逼近非线性函数的仿真实验,证实了新规则在训练速度上的优越性,得出新规则中控制因子的理想取值为0.95。This paper proposes a new rule to adjust the weight value,which ensures enough weight adjusted range when the output of neuron is close to 0 or 1. The new rule solves a difficult problem for standard BP network to break away from the flat area when it has entered, here its training speed is very slow. A experiment which uses a BPnetwork with single hidden layer to approximate a nonlinear function is presented in the paper. It verifies the superiority of the new rule in the training speed,and draws a conclusion that the ideal value of the control factor in the new rule is 0. 95.

关 键 词：平坦区 S型函数 梯度下降 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TU433[自动化与计算机技术—控制科学与工程]