用神经元网络辨识非线性系统中的网络结构选择

Network Structure Selection in Neural Network Based Nonlinear System Identification

本文定义了神经元网络的权值拟熵，在对多层前馈网训练的常规目标函数中加入权值拟熵作为约束项以改变网络的权值分布从而修定网络结构．将此方法用于一类非线性系统的神经网络辨识中可以优化网络模型输入项数和隐节点数目．

In this paper we define the pseudo-entropy of network weights. By adding it into the normalobjective function we can obtain a rational network structure during training. Put this method into the neuralnetwork based nonlinear system identification we can acquire a proper number of input and hidden neurons.