摘要
本文将BP网络的输入错误分为死区错误和反相错误,提出了对输入容错的高可靠的BP网络的构造方法。对上述两种错误,分别讨论了在离可靠构造下的可靠性问题,指出当输入正确的概率在一定范围内变化时,通过增加输入重复数,可以使得网络正确实现规定功能的概率接近于1,或者说,使得网络对于输入错误不敏感。
The input errors of BP(Back-Propagation) network are classified into dead point error and phase reversal error in this paper. A highly reliable reconstruction method for a BP network is proposed. The reliability problem of the reconstructed BP network is discussed under the above mentioned input errors respectively. It is shown that, when the probabilities that the inputs are correct (input reliabilities) are varying within a determined region, by increasing the number of inputs, the probability that the reconstructed BP network realizes the giyen function correctly (network reliability) may approach 1. The reconstructed BP network is fault tolerant to input errors to some extent.
出处
《自动化学报》
EI
CSCD
北大核心
1992年第5期551-558,共8页
Acta Automatica Sinica
基金
国家自然科学基金资助项目
关键词
神经网络
BP网络
可靠度
BP network
input errors
dead point error
phase reversal error
reconstruction.
