改进的ART神经网络用于模式识别

IMPROVED ART NEURAL NETWORK FOR PATTERN RECOGNITION

分析了ART网络在模式识别应用中的特性.在原ART1网络中,由于子集对包集的重新编码过程使得ART1网络具有学习过程不稳定性,从而影响了在任意输入环境下网络的分类性能,为此本文对ART1网络的匹配度计算提出改进方法,使匹配度计算公式更合理地反映了输入模式与模板模式的相似程度,有效地克服了学习过程不稳定性。在ART2网络中,本文根据同一类别中最大、最小模式数提出了警戒线户值的自适应调整方法,避免在固定户值下可能引起的分类过粗或过细,使网络具有一定容错能力,又有一定敏感性。

The properties of ART neural network in the application of pattern recognition are analysised in this paper. We find that the learning process of ART1 is unstable because of the recoding of subset to its coded superset. Improvement for match degree r is proposed in this paper to make P reflect the similarity between input pattern and coded pattern more reasonably. According to the maximum and minimum pattern number in one class we present the self-adaptive regulation method of vigilance P to enable the ART2 both resistant to noise and sensitive to different pattern. The improved ART1 ART2 networks can deal with arbitary input sequence more effectively.