基于网络聚类选择的神经网络集成方法及应用  

Selective approach for neural network ensemble based on network clustering technology and its application

作　　者：刘大有[1,2] 张冬威[1,2] 李妮娅[1,2] 刘杰[1,2] 金弟[1,2] 

机构地区：[1]吉林大学计算机科学与技术学院,长春130012 [2]吉林大学符号计算与知识工程教育部重点实验室,长春130012

出　　处：《吉林大学学报（工学版）》2011年第4期1034-1040,共7页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(60873149;60973088;60773099;60703022);'863'国家高技术研究发展计划项目(2006AA10Z245)

摘　　要：面向农作物精准施肥量确定问题,提出一种基于复杂网络聚类选择的神经网络集成方法。在该方法中,首先采用回放取样方法来生成多个神经网络个体,其次利用网络聚类算法FEC从这些神经网络个体集中选出部分个体,再次对选出的神经网络个体分别用线性加权方法和非线性方法进行集成,最后对两个集成结果进行融合得到预测结果。于2008年在吉林省榆树玉米试验田七号地对该方法进行了应用。实验结果表明:该方法不仅优于传统的施肥模型、神经网络线性加权集成和神经网络非线性集成方法,而且还具有较强的泛化能力。To solve the difficult issue of crop precise fertilization,a neural network ensemble approach is proposed,which is based on the complex network clustering algorithm.In this approach,first,the method of sampling with replacement is adopted to produce the number of neural network units.Then,Yang＇s network clustering Forward Error Correction（FEC） algorithm is used to select the networks with high precision and great diversity.Third,the selected networks are ensembled separately with linear weighted ensemble method and nonlinear ensemble method.Finally,the prediction result is achieved by amalgamating the results produced by the two ensemble methods.Test results,carried out at the No.7 corn plot in Yushu city in 2008,reveal that it proposed approach is better than the traditional fertilization models,linear weighted ensemble model and nonlinear ensemble model.In addition,the ability of generalization of the approach is strong.

关 键 词：人工智能 神经网络集成 复杂网络聚类 线性加权集成 非线性集成 精准施肥 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]