基于广义回归网络的动态权重回归型神经网络集成方法研究  

Dynamically Weighted Ensemble Neural Networks with Generalized Regression Neural Network for Solving Regression Problems

作　　者：沈掌泉[1] 孔繁胜[2] 

机构地区：[1]浙江大学农业遥感与信息技术应用研究所,浙江杭州310029 [2]浙江大学计算机科学与技术学院,浙江杭州310027

出　　处：《计算机应用研究》2005年第12期41-43,72,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(40201021)

摘　　要：神经网络集成技术能有效地提高神经网络的预测精度和泛化能力,已成为机器学习和神经计算领域的一个研究热点。针对回归分析问题提出了一种动态确定结果合成权重的神经网络集成构造方法,在训练出个体神经网络之后,根据各个体网络在输入空间上对训练样本的预测误差,应用广义回归网络来动态地确定各个体网络在特定输入空间上的权重。实验结果表明,与传统的简单平均和加权平均方法相比,本集成方法能取得更好的预测精度。Combining the outputs of several neural networks into an aggregate output often gives improved accuracy over any individual output. This paper presents an ensemble method for regression that has advantages over weighted average combining techniques. Generally, the output of an ensemble is a weighted sum which are weights fixed. The ensembles are weighted dynamically, the weights dynamically determined from the predicted accuracies of the trained networks with training dataset, the more accurate a network seems to be of its prediction, the higher the weight. This is implemented by generalized regression neural network. Empirical results show that this method improved on prediction accuracy.

关 键 词：神经网络集成 BP网络 动态权重 广义回归神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]