Scholarlink

This paper presents the equilibrium between accuracy and computational cost through artificial neural network (NN) in order to obtain models of nonlinear systems by Hammerstein Wiener (H-W) structure based technique of identification. The seek for optimal parameters is carried out by the back-propagation method using gradient descent. Emphasizing the enhance accuracy, a method to minimize the number of parameters of the H-W structure is provided. It is proved by minimizing the number of synaptic weights the result can be the same in the H-W structure. This method is substantiated in two leading assumptions: a) design (choose linear functions activation) and b) training (initialization of equal values block by block). Eventually, to decrease the number of parameters an algebraic operation is implemented.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Accuracy and Computational Cost of the Hammerstein-Wiener Neural Network Structure Applied to the Nonlinear Systems Identification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jesús Namigtle Jiménez,
  • David Lara Alabazares,
  • Víctor Manuel Alvarado Martínez,
  • Eddy Sánchez-DelaCruz,
  • Irahan Otoniel José Guzmán,
  • Pablo Colorado Posadas

摘要

This paper presents the equilibrium between accuracy and computational cost through artificial neural network (NN) in order to obtain models of nonlinear systems by Hammerstein Wiener (H-W) structure based technique of identification. The seek for optimal parameters is carried out by the back-propagation method using gradient descent. Emphasizing the enhance accuracy, a method to minimize the number of parameters of the H-W structure is provided. It is proved by minimizing the number of synaptic weights the result can be the same in the H-W structure. This method is substantiated in two leading assumptions: a) design (choose linear functions activation) and b) training (initialization of equal values block by block). Eventually, to decrease the number of parameters an algebraic operation is implemented.