In this paper, by analyzing the Madaline network work principle , according to the network characteristic and the short-comings of the existing algorithms including the adjust weigh formula used in MRII algorithm with many parameters , but most of which come from experience in practice without theoretical reason , it made networks fall into "local recycle" with turning over the neuron based on confidence principle .We present an improved learning algorithm based on MRII , by establishing neuron '' s sen-sitivity as a tool for measuring the turn of each hidden neuron , which reduces the number of weight adjustment .Some experimen-tal results verify the effectiveness of the proposed algorithm .%根据Madaline网络工作原理,针对其网络特点和现有算法中存在的缺点,包括存在权值修改公式参数较多不容易协调,经验取值缺乏理论依据不够灵活,按照置信度原则进行翻转神经元会陷入"局部震荡". 提出改进的MRII学习算法,通过建立神经元敏感性替代置信度作为度量隐层神经元翻转的尺度,并采用感知机学习规则,减少权值调整次数,实验结果验证了该算法的优越性.
展开▼