评价神经网络拟合精度的另一个指标  

Another Criterion to Judge Performance of Neural Networks

作　　者：徐力平[1] 张炎华[1] 

机构地区：[1]上海交通大学信息检测技术及仪器系,上海200030

出　　处：《上海交通大学学报》2002年第7期980-983,共4页Journal of Shanghai Jiaotong University

摘　　要：在基于神经网络的故障检测中 ,要求神经网络拟合某特定动态系统时均方预测误差的均值及其标准差均小 ,故在以均方预测误差的均值作为评价神经网络逼近特定动态系统性能的系统化交叉证实法的基础上进行改进 ,以均方预测误差的均值和标准差两个指标评价神经网络逼近特定动态系统性能 .仿真结果证明 ,同时用两个指标不仅是可行的 。In some applications of neural networks, such as failure detection based on neural networks, both the average and the standard deviation of mean squared prediction errors are needed to be small. This paper modified a model selection method called cross validation to select the number of hidden units in neural networks from with one criterion to with the both criteria. The simulation result shows that selecting the number of hidden units in neural networks with the both criteria not only is feasible but also improves the model selection method more or less.

关 键 词：神经网络 拟合精度 预测误差 均值 标准差 评价指标 系统化交叉证实法 网络结构 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]