针对小样本数据集的鲁棒单隐层前馈网络建模方法  

Robust single hidden layer feed-forward neural networks modeling for small datasets

作　　者：张荣[1] 邓赵红[2] 王士同[2] 蔡及时[3] 钱鹏江[2] 

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]江南大学数字媒体学院,江苏无锡214122 [3]香港理工大学护理学院,香港999077

出　　处：《控制与决策》2012年第9期1308-1312,1319,共6页Control and Decision

基　　金：国家自然科学基金项目(60903100);江苏省自然科学基金项目(BK2009067);中央高校基本科研业务费专项资金项目(JUSRP21128);江苏省信息融合软件工程技术研究开发中心开放基金项目(SR-2011-01)

摘　　要：单隐层前馈神经网络是应用最广泛的智能建模模型之一,但该模型面对小样本集时传统的学习算法易陷入过拟合,尤其当数据集包含较大噪音时学习模型鲁棒性较差,对噪音很敏感,针对此不足,提出一种针对小样本数据集的鲁棒单隐层前馈神经网络学习算法,所提出的算法由于引入了ε-不敏感学习度量和结构风险项,能有效克服传统学习算法存在的缺陷,显示出较好的鲁棒性,在模拟和真实数据集上的实验亦证实了上述优点。Single hidden layer feed-forward neural network（SLFN） is one of the most widely used models for intelligent modeling. But the model faces that for small sample sets, the traditional learning algorithm may train a model to fall into the over-fitting sate. In particular, when the dataset contains a large amount of noise, the trained model has weak robustness and is very sensitive to noise. In order to overcome this shortcoming, a robust learning algorithm of SLFN is derived for small and noisy datasets. Due to the introduction of e-insensitive learning measure and the structural risk term, the proposed algorithm can effectively overcome the shortcoming of the traditional learning algorithm. The experimental results on simulated and real-world datasets also confirm the above advantages.

关 键 词：单隐层前馈神经网络 ε-不敏感学习 结构风险最小化 鲁棒性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]