多层结构神经网络的等误差范围逼近与收缩学习方法及其应用

The Equal-Error Range Approximation and Shrink Learning Rule for Multilayer Perceptions and Its Applications

本文提出了一种用于前馈型多层神经网络学习的等误差范围逼近与收缩学习方法,这种方法仅仅要求网络的实际输出落在理想模式输出的一个事先给定的误差范围之内,从而可以大大提高网络的学习速度,且运算量小,而且通过适当选择等误差范围,它还可以提高网络在模式识别中的推广性能.如果网络用于模式联想等方面时,通过误差范围的逐步收缩,这种方法还可以以很小的额外代价提高网络学习的逼近精度;另外,它还可以避免传统方法中经常出现的训练模式反转等局域极小状态和过学习现象的出现.最后,文中给出了以这种方法训练的网络用于脑电波癫痫信号识别中的实验结果及其分析.

In this paper, we propose an Equal-Error Range Approximation and Shrinking Learning Algorithm for multilayer perceptrons. It requires the error between each network output node activation and its target to fall into a given error range, thus it can learn faster in lower calculation cost and may avoid reversed target output and overlearning. hence itcan improve networks'generalization abilities in pattern recognitions. Through gradually Shrinking of the error range, it can also enable the networks to learn the targets more accurately in less training iterations. Finally, we apply this learning algorithm trained network to the EEG detection, and the experiment results have showed the above advantages of the proposed algorithm.