A Rational Initialization of the Feed-Forward Neural Network Regression.

前馈神经网络回归的合理初始化。

• 批准号: 10680328
• 负责人: MAYEKAWA Shin-ichi
• 金额: $ 1.41万
• 依托单位: National Center Test for University Admissions
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (C)
• 财政年份: 1998
• 资助国家: 日本
• 起止时间: 1998 至 1999
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-10680328/
• 关键词: neural network regression; initial estimate; variable selection; spline regression; parameter estimation; ロジスティク展開; Gram-Schumit法

In this research, a method to derive a rational initial estimate of the parameters of neural network regression model is derived.Let the first layer of the 3-layer feed-forward neural network regression model be donated as X and the parameter (intercept and weight) matrices from the first layer to the second be θィイD12*ィエD1, WィイD12*ィエD1. Our method first define the values of the parameters so that the second layer output OィイD12*ィエD1 = ψ(1θィイD12*'ィエD1 + XWィイD12+ィエD1) (1) consists of a set of the monotone functions rich enough to cover the criterion space, Y. Then, a variable selection technique will be used to select the best fitting subset of the columns of OィイD12*ィエD1 matrix resulting in the (selected) subset OィイD12ィエD1 and the weight matrix θィイD13ィエD1, WィイD13ィエD1 using the following multivariate linear regression : Y 【approximately equal】 1θィイD13'ィエD1 + OィイD12ィエD1WィイD13ィエD1 (2)

本文提出了一种神经网络回归模型参数合理初始估计的方法。设3层前馈神经网络回归模型的第一层为X，从第一层到第二层的参数（截距和权值）矩阵为θ D12* D1， W D12* D1。我们的方法首先定义参数的值,这样第二层输出OィイD12 *ィエD1 =ψ(1θィイD12 *’ィエD1 + XWィイD12 +ィエD1)(1)由一组单调函数足够富有的标准空间,y,变量选择技术将被用于选择最佳拟合的子集的列OィイD12 *ィエD1矩阵导致(选择)子集OィイD12ィエD1和权重矩阵θィイD13ィエD1, WィイD13ィエD1使用如下多元线性回归:Y【约等于】1θィイD13”ィエD1 + OィイD12ィエD1WィイD13ィエD1 (2)