BP神经网络的饱和分析及改进

为了改善BP神经网络易形成局部最小,收敛速度慢的缺点,从分析三个因子学习因子、惯性因子和形状因子对BP算法性能影响出发,提出了离线调整学习因子和惯性因子,在线调整形状因子的联合优化方法.这种方法使网络在训练时,不仅神经元的连接权在不断调整,而且其自身的输入输出关系也在变动,从而使网络脱离饱和区,提高了收敛速度.最后以最典型应用函数逼近和XOR分类为例进行验证,仿真结果显示,联合优化方法不仅提高了网络训练速度,还提高了收敛精度,而且比一般的改进方法效果更好,具有一定的实用价值.

Analysis and Improvement of Paralysis of BP Neural Networks

An improved method is expounded in order to ameliorate the disadvantage of BP neural network such as local minima,low training speed.The influence of three factors such as learning factor,momentum factor and shape factor on the capability of BP algorithm is assayed,meanwhile,a joint-optimized method for adjusting learning factor and momentum factor outline and adjusting shape factor online is proposed.When the network is training,not only the information between two neural elements is adjusted,but also the relation between input and output is varying.At last,examples such as function approximation and XOR classification show that the joint-optimized method not only improves the training speed but also ameliorates the convergence accuracy,at the same time,this method is superior to other methods,and has practical value.