ART2神经网络聚类的改进研究

Improvement of Clustering of ART2 Neural Network

为进行快速动态层次聚类,通过分析自适应谐振理论(adaptive resonance theory,ART)神经网络的快速学习、主观设置警戒参数、输出无层次结构等优缺点以及自组织特征映射(self-or-ganizing feature map,SOFM)神经网络的侧反馈、不能动态聚类、输出无层次结构等优缺点的基础上,借鉴Hebb规则的思想,针对ART2神经网络的聚类算法进行了改进研究。通过结构描述、算法分析,该算法融合了ART2和SOFM的优点,克服其不足之处,以快速学习的方式形成可带有多层层次的动态聚类结构(不同的层次代表不同粒度的聚类),此外还降低了对警戒参数主观设置的要求,对于较粗粒度的聚类不再需要重新训练神经网络。并通过仿真实验证明该算法的有效性。

In order to achieve dynamic clustering with hierarchy structure, after analyzing the short-comings and advantages of adaptive resonance theory （ART） neural network, such as fast study, subjectively setting vigilance parameter and output without hierarchy structure; and after analyzing the shortcomings and advantages of Self-Organizing Feature Map （SOFM） , such as side-feedback, inability of dynamic clustering and output without hierarchy structure, improvement of clustering algorithm of ART2 neural network has been presented with the reference of Hebb Principle. By structure description and algorithmic analysis, this model incorporates the advantages of ART2 and SOFM and overcomes their shortcomings, obtains dynamic clustering structure with multilayer hierarchy structure by fast study （each layer denotes a category of different granularity ） ; this model also reduces the request of setting vigilance parameter and has no demand of retraining neural network of bigger granularity. Finally the effectiveness of the algorithm is demonstrated by simulation.