随机权分布对极限学习机性能影响的实验研究  

Experimental Research on Effects of Random Weight Distributions on Performance of Extreme Learning Machine

作　　者：翟俊海[1] 臧立光 张素芳[3] 

机构地区：[1]河北省机器学习与计算智能重点实验室河北大学数学与信息科学学院,保定071002 [2]河北大学计算机科学与技术学院,保定071002 [3]中国气象局气象干部培训学院河北分院,保定071000

出　　处：《计算机科学》2016年第12期125-129,145,共6页Computer Science

基　　金：国家自然科学基金项目(71371063);河北省自然科学基金项目(F2013201220);河北省高等学校科学技术研究重点项目(ZD20131028)资助

摘　　要：极限学习机是一种训练单隐含层前馈神经网络的算法,它随机初始化输入层的权值和隐含层结点的偏置,用分析的方法确定输出层的权值。极限学习机具有学习速度快、泛化能力强的特点。很多研究都用服从[-1,1]区间均匀分布的随机数初始化输入层权值和隐含层结点的偏置,但没有对这种随机初始化合理性的研究。用实验的方法对这一问题进行了研究,分别研究了随机权服从均匀分布、高斯分布和指数分布对极限学习机性能的影响。研究发现随机权的分布对极限学习机的性能的确有影响,对于不同的问题或不同的数据集,服从[-1,1]区间均匀分布的随机权不一定是最优的选择。研究结论对从事极限学习机研究的人员具有一定的借鉴作用。Extreme learning machine （ELM） is an algorithm for training single-hidden layer feed-forward neural net- works （SLFNs）. ELM firstly employs randomization method to generate the input weights and hidden nodes biases, and then determines the output weights analytically. ELM has fast learning speed and good generalization ability. All methods published in literatures usually initialize the weights of input layer and biases of hidden nodes with a uniform distribution over the interval [-1,1]. However, there are no studies on the rationality of this setting in literatures. This paper investigated this problem by experimental approach. Specifically, the effects of random weights with uniform distribution,Gaussian distribution and exponential distribution were studied. We found that the random weight distributions do have impact on the performance of the extreme learning machine. For different problems or different data sets, the random weights with uniform distribution in [-1,1] are not necessarily optimal. The results of this paper can be used for reference by the researchers engaged in the study of ELM.

关 键 词：随机权分布 极限学习机 均匀分布 高斯分布 指数发布 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]