宋跃忠

摘要：ART2是一种基于自适应谐振理论的无监督神经网络，由于其快速响应、实时学习等特点，被广泛的应用在实时聚类问题中。传统的ART2存在幅值信息丢失、容易产生模式漂移的问题，本文针对此不足提出了一种基于广义相似度和置信度的GSC-ART2网络。它通过引入广义相似度检测和竞争机制，解决了幅值信息丢失的问题。置信度结合广义相似度的权值调整方式抑制了模式漂移并使网络的连接权值更加准确。通过实验表明，GSC-ART2网络在处理幅值相关、样本渐变分类问题上的识别性能均优于传统ART2网络，从而证明了此GSC-ART2网络的有效性，也为解决模式识别中普遍存在的模式漂移问题找到了一种优良的解决方法。

关键词：GSC-ART2；模式漂移；幅值丢失；广义相似度；置信度；

中图分类号： TP391 文献标识码： A 文章编号：2095-2163（2014）04-

文章编号：

An Improved ART2 Neural Network： Resisting Pattern Drifting through General Similarity and Confidence Measures

SONG Yuezhong，LI Haifeng，GAO Chang

（ School of Computer Science and Technology，Harbin Institute of Technology，Harbin 150001，China）

Abstract： ART2 is a kind of non-supervised neural network based on the Adaptive Resonance Theory， and due to such advantages as rapid response and real-time learning abilities， ART2 has been widely used in real-time clustering problems. In traditional ART2 models， the amplitude information is usually ignored and the problem of pattern drifting often occurred. To solve such problems， general similarity and confidence measures are introduced into ART2 to form an improved model --- GSC-ART2. Using a vigilance-testing and a competition mechanism based on the general similarity， the problem of amplitude information losing is solved in GSC-ART2. The weights adjustment is modified to consider both the general similarity and the confidence measures. In such a way， ourthe designed GSC-ART2 inhibits pattern drifting and obtains more accurate network connections. Experiments showeds that the GSC-ART2 performeds better than traditional ART2 in cases where the data possess magnitude information and the data grading or pattern drifting exists. OurThe proposed GSC-ART2 network would become an universal solution to the pattern drifting problem in various applications.

Keywords： GSC-ART2； ART2； Ppattern dDrifting； aAmplitude Iinformation Llosing； General Similarity； Confidence Measure

0引言

抑制模式漂移和防止网络震荡是提高模式识别准确率关键因素，尤其对人工神经网络ART2学习过程中的抑制模式漂移、进行权值合理准确性研究，更是提升网络分类性能的重要工作。自Stephen Grossberg教授提出一种无教师竞争型神经网络的学习机制即自适应谐振理论（Adaptive resonance theory，ART）[1]，并解决了神经网络学习中稳定性与可塑性之间的矛盾，并后，虽然促进了一系列实时无教师聚类学习的模式识别方法的发展。，但它却容易产生模式漂移、网络震荡的问题，本文即针对这个问题进行了深入的研究。

ART网络是基于自适应谐振理论的神经网络，它在其在实现原理上借鉴了人脑对信息的处理机制，既能够保持旧的知识，还可以对新的知识进行学习。……