一种优选神经网络训练样本的混合聚类算法

神经网络训练集中含有大量相似样本不但增加了网络的训练时间还对网络泛化能力存在较大影响,合理的选择样本集训练神经网络模型影响着建模的效率。根据实际应用中神经网络学习样本具有的内在特征和规律性,提出了一种基于自组织映射（SOM）神经网络的K-均值聚类算法优选神经网络样本,算法的主要思想是通过对样本数据的聚类分析,剔除孤立样本后挑选出具有代表性的样本训练神经网络。实验结果表明,相对随机选择法而言,本算法能够有效地减少训练样本的数目,提高建模效率。

Combined Clustering Algorithm for Neural Network Training Data Sample Selection

The neural network training set containing lots of similar samples not only increases the training time but also reduces the network＇s generalization performance.Rational selections of training sample to train the neural network affect the efficiency of modeling.According to neural network training samples in the actual application embodying inherent characteristic and regularity,a hybrid algorithm of self-organization map（SOM） neural network,combined with K-means clustering algorithm was proposed to select training data.By clustering analysis of sample data,isolated samples were removed and representative samples were selected to train the neural network.The experiment results expatiate that the algorithm is better than random selection method on reducing the number of training samples effectively and improves the modeling efficiency.