摘要
为了比较训练人工神经网络的所需样本点的选取,分别采用随机遍历法、正交设计法和均匀设计方法产生样本点,用于训练神经网络.分析结果表明,在样本点个数相同情况下,均匀设计法的代表性最好,正交设计法次之,而随机遍历法较差.随机遍历法随着样本点个数的增多,同样可以提高其代表性.当函数随变量在区间内变化较小(因素水平可以取的较少)时,正交设计法也不失为一个好的选择.均匀设计法在多变量,且每个变量需要选取较多水平数的情况下,更能体现它的优越性.
Based on the theory of the NN in the biology,the ANN is a complicated,massive,nonlinear dynamic system that simulates the biologic neural structureLots of sample points are required to train the ANNIn this paper, three methods are presented to produce the sample points adopted in training the ANNIt is testified that under the same condition even design is the best method to produce sample points,followed by orthogonal design,and the third is ransackingRansacking will be better if the number of sample points is largeWhen the function changes little with the variables,fewer factor levels can be chosen and orthogonal design is a good methodIts advantages stand out for the occasion of multiple variables and many factor levels.
出处
《郑州大学学报（工学版）》
CAS
2003年第1期63-65,69,共4页
Journal of Zhengzhou University（Engineering Science）
基金
国家先进制造技术中心资助项目(AMTRC2002-3)
