基于FCM与KKT条件的增量学习方法

增量学习方法的思想是仅利用部分相关的样本集参与训练,即能够保留历史样本知识,又能够不断地吸收新的知识,提高机器学习效率和精度,解决了大量样本训练时间长和存储空间不足的问题。因此,如何有效地丢弃大量无效的样本点是增量学习算法研究的重点。文中提出了一种FCM(Fuzzy C-Means)和KKT(Karush-KuhnTucker)条件结合的增量学习方法,分别从历史样本集和新增样本集两个阶段对无效样本进行过滤,利用余下的样本进行训练。最后,利用UCI数据库中的4组数据进行实验分析,结果证明训练精度与全数据样本的训练精度几乎完全拟合。

An incremental learning approach based on FCM and KKT

The aim of the incremental learning approach is to be capable of not only reserving the historical sample knowledge, but also continuously absorbing new information, with only part of relevant sample set to train, which improves the learning efficiency and accuracy of machine and solves the problems of long training time for a large number of samples and not enough storage space. Therefore, the focus on the study of the incremental learning algorithm is how to effectively abandon plenty of useless sample points. This paper presents an incremental learning algorithm combining FCM （Fuzzy C-Means） term with KKT （Karnsh-Kuhn-Tucker） term, which filters useless samples from the stage of historical samples set to the stage of new samples set and uses the remaining samples to train. At last, the experimental analysis has been made based on the four groups of data in UCI database, and the results shows that the training accuracy using this method is almost completely fitting to that of using the total samples.