论文部分内容阅读
与现有网络结构设计方法不同,本文将RBF网络解释为解释变量和被解释变量之间的一个非线性函数,基于RBF网络的学习动态特性提出2种修剪模型WRBF和TRBF。这两种模型根据参数显著性增加和删减节点,为网络结构设计提供了理论依据。对中国信贷序列预测的结果表明,这些模型能够识别外移、萎缩和衰减等冗余核函数,得到的精简网络具有最好的预测精度,对于提高货币政策前瞻性具有很好应用价值。
Different from the existing network structure design methods, RBF network is interpreted as a nonlinear function between explanatory variables and explanatory variables. Based on learning dynamics of RBF networks, two kinds of pruning models WRBF and TRBF are proposed. These two models provide the theoretical basis for the network structure design according to the significant increase of parameters and the deletion of nodes. The results of China’s credit sequence prediction indicate that these models can identify redundant kernel functions such as excursion, shrinkage and attenuation, and the resulting reduced network has the best forecasting accuracy, which has good application value for improving the forward-looking monetary policy.