ART－2神经网络分类器的研究

Resarch on a Classifier Using the ART-2 Neural Network

ＡＲＴ神经网络是按照自适应谐振理论建立的一种自组织的人工神经网络，本文通过分析ＡＲＴ神经网络的结构，发现其用于模式识别中有很好的聚类特性，但在数据处理过程中有部分数据量丢失的现象，也就是说，非常重要的幅度信息没有被考虑到。本文提出一种新的结构和相应的新算法，并成功地把这种结构和算法用到分类器设计中。这种结构增加了一套幅度调整子系统对幅度信息先进行匹配，选出合适的几个类别来限定以后的分类；然后再根据相位信息来进一步从可选类中选出最相近的类来，这样系统进入谐振状态，开始网络系数的调整。如果其中任何一个步骤失败，即不能满足预先给出的阈值，我们就开劈新类或放弃这个样本，由于原先系统的运算量主要分布在网络的注意子系统的迭代过程和搜索过程、调整子系统的校验过程中，加入幅度调整子系统减小了搜索范围，因此运算量明显减少。实验表明新的网络结构用作模式分类时能适应更一般的情况，而且其分类速度也有明显提高。

ART Neural Network is based on the Adaptive Resonance Theory. After analyzing the traditional structure of the ART Neural Network, We find that this kind of structure has an inherent property of clustering, but some important information in the input sample is lost, that is, the amplitude of the sample is ignored. So an improved structure and its associated algorithm is presented, and is used successfully to implement a classifier. In the new structure, a new kind of orienting subsystme is added. When sample is inputted, calculate its amplitude first and select all possible suitable categories according to a pre - given threshold. Then from these categories, select the best according to the phase information, thus the system comes into a state of resonance - the long term coefficients are adjusted to make themselves more like the input vector. If any step failed, i. e. , the given threshold can not be satisfied, a new category is created by storing a pattern which is like the input sample or this sample is rejected. Experiments indicate that the new structure is more applicable and effective than the original one used for the classifier, more over, the classifying speed is significantly faster.