【摘要】 从多层前向神经元网络的BP算法出发,通过计算机仿真发现含有噪声干扰的样本对BP网络的训练会产生不良的影响:①神经元网络的学习速度不易把握;②网络的学习过程不收敛或收敛值会偏移真实的期望输出。文中从本质上对产生这些影响的原因进行了分析,并且进一步揭示了能量函数的选取是使BP算法抗噪声能力差、鲁棒性不强的主要原因。更多还原

【Abstract】 Based on BP algorithm of multilayer forward neural networks, some worse influences on BP network’straining are found by simulating the neural network when the training data with noise disturbance are employed inthe course of network training. These influences are: ① the rate of the network learning is unstable; ② thecourse of training is not convergence or the output value will converge to other value instead of the desiredoutput. In the paper, some reasons that bring the influnces are analyzed essentially, and we further reveal that theprincipal reason lies in the choice of the Energy Function that make the BP algorithm have weak ability to resist noiseperturbation.更多还原