零误差密度函数准则的BP神经网络学习研究

Learning of BP Neural Networks Based on Zero-Error Density Criterion Function

BP神经网络的学习通常以均方误差函数(MSE)为目标函数,当目标变量不满足高斯分布时,其结果可能偏离真正最优.零误差密度函数(ZED)利用非参数估计中的Parzen窗法得到误差在零点的概率密度函数.将零误差密度函数作为BP网络的目标函数时,通过对光滑参数的选择使新的目标函数能够适用于期望输出满足任意分布.仿真实验分别以零误差密度函数和均方误差函数为目标函数的BP网络学习在函数逼近方面进行比较,结果表明零误差密度函数要比均方误差函数的适用范围更广.

BP neural networks usually use mean squares error（MSE） function as the objective function,the results may deviate the optimal values in the condition that expected vectors don＇t follow Gaussian distribution.zero-error density （ZED） function uses Parzen window method of non-parameter estimation to get error density at origin,which can be used in the condition that expected output vector follow any density distribution by choosing an appropriate smooth parameter. Compared the BP networks with the new cost function with the BP networks with mean squared function in function approximation through the experiments,the simulation results show the zero-error density function has a larger range of application than mean squared error（MSE）function.