基于改进的级联相关算法的系数定税预测  

The Coefficient Constant Tax Forecast Based on Improved Cascade-Correlation Algorithm

作　　者：李满玲[1] 

机构地区：[1]株洲职业技术学院工业设计系,湖南株洲412000

出　　处：《湘南学院学报》2013年第5期30-34,共5页Journal of Xiangnan University

基　　金：国家自然科学基金项目(61170283)

摘　　要：BP神经网络在非线性建模上具有强大的功能,但是其网络结构的选择需要经验,缺乏科学性.级联相关算法是一种优秀的构造方法,实现了网络结构的自动调整,但是其网络训练速度很慢,样本预测精度还有待提高.针对于此,提出一种改进的级联相关算法.在该算法中,网络结构起始于适当的隐结点数目,通过折半筛选算法动态调整每批新增最佳隐结点数目,还通过反向传播来修正各层结点的权值和阙值.最后,将该算法应用到系数定税建模实验中.实验证明,该算法比标准的级联相关算法网络训练速度大大提高,预测精度也更高,更适用于系数定税预测建模.BP neural network has powerful function in nonlinearmodeling,but the selection of network structure requires experience and lacks scientificness.Cascade correlation algorithm is a excellent method of construction and it can automatically adjust the network structure,but the network training speed is very slow and the forecast accuracy needs to be improved.To solve these problems,we put forward an improved cascade correlation algorithm,inwhich,the network structure started from the appropriate numberof hidden nodes and each batch of the new best hidden nodeswere dynamically adjusted by binary screening,and also by back propagation,each layer node weights and thresholdwere corrected.Finally,the algorithmwas applied to coefficients taxmodeling experiment.The results showed that,this algorithmgreatly improved the training speed of the standard correlation algorithm cascade network,the prediction accuracy and was more suitable for the coefficientof tax forecastmodeling.

关 键 词：级联相关 反向传播 路径优化 折半筛选 系数定税 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]