Scholarlink

<p>Recurrent neural networks (RNNs) are well established with comprehensive models capable of solving zero finding problems. Most of the conventional designs apply the infinite activation, but may not be practical for implementation. This paper presents model designs of finitely-activated zeroing neural networks (structure-like RNNs), possessing the finite-time convergence property as well, for solving time-variant convex quadratic programming. Two techniques for realizing finite activation are provided, based on which novel activation functions (AFs) are constructed, including the conic AFs and the finitely-valued power-rate AFs. Theoretical analyses of finite-time convergence are presented in detail and settling time is exactly established for each model. It is shown that finitely-valued AFs can approximate or even outperform the original power-rate AFs. The proposed neural network models are applied to solve an example of time-variant quadratic programming, and the repetitive motion planning of redundant robots with joint angle and joint velocity constraints, where the anti-disturbance capability of the finitely-activated integral neural network models has been examined and verified. The obtained numerical results demonstrate effectiveness of the computing schemes.</p>

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

Time-variant quadratic programming solving by using finitely-activated RNN models with exact settling time

  • 期刊论文

  • 发表日期：

    2025年1月8日

查看全文
  • Mingxuan Sun,
  • Yu Zhang,
  • Liming Wang,
  • Yuxin Wu,
  • Guomin Zhong

摘要

Recurrent neural networks (RNNs) are well established with comprehensive models capable of solving zero finding problems. Most of the conventional designs apply the infinite activation, but may not be practical for implementation. This paper presents model designs of finitely-activated zeroing neural networks (structure-like RNNs), possessing the finite-time convergence property as well, for solving time-variant convex quadratic programming. Two techniques for realizing finite activation are provided, based on which novel activation functions (AFs) are constructed, including the conic AFs and the finitely-valued power-rate AFs. Theoretical analyses of finite-time convergence are presented in detail and settling time is exactly established for each model. It is shown that finitely-valued AFs can approximate or even outperform the original power-rate AFs. The proposed neural network models are applied to solve an example of time-variant quadratic programming, and the repetitive motion planning of redundant robots with joint angle and joint velocity constraints, where the anti-disturbance capability of the finitely-activated integral neural network models has been examined and verified. The obtained numerical results demonstrate effectiveness of the computing schemes.