Scholarlink

<p>This work investigates two newly developed modified conjugate gradient (CG) algorithms, namely BBM1 and BBM2, designed to enhance the performance of unconstrained optimization problems and improve neural network training. These algorithms incorporate advanced derivative-based techniques while employing flexible line search strategies to ensure sufficient descent directions. The proposed methods are supported by a solid theoretical framework that guarantees global convergence under suitable assumptions. Extensive numerical experiments demonstrate that both BBM1 and BBM2 outperform classical CG algorithms, such as the Dai-Yuan (DY) method, in terms of computational efficiency, convergence speed, and robustness. Moreover, the applicability of these algorithms is validated through their successful deployment in training recurrent neural networks, showcasing their capability to address large-scale nonlinear optimization tasks. The promising results confirm the potential of new methods as effective tools for various optimization problems and artificial intelligence (AI) applications.</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

Adjusting Parameters for Conjugate Gradient Method for Training Neural Networks and Solving Unconstrained Minimization Problems

  • 期刊论文

  • 发表日期：

    2025年9月15日

查看全文
  • Basim A. Hassan,
  • Alaa Luqman Ibrahim,
  • Ali Ahmed A. Abdullah,
  • Abdulameer A. Saad

摘要

This work investigates two newly developed modified conjugate gradient (CG) algorithms, namely BBM1 and BBM2, designed to enhance the performance of unconstrained optimization problems and improve neural network training. These algorithms incorporate advanced derivative-based techniques while employing flexible line search strategies to ensure sufficient descent directions. The proposed methods are supported by a solid theoretical framework that guarantees global convergence under suitable assumptions. Extensive numerical experiments demonstrate that both BBM1 and BBM2 outperform classical CG algorithms, such as the Dai-Yuan (DY) method, in terms of computational efficiency, convergence speed, and robustness. Moreover, the applicability of these algorithms is validated through their successful deployment in training recurrent neural networks, showcasing their capability to address large-scale nonlinear optimization tasks. The promising results confirm the potential of new methods as effective tools for various optimization problems and artificial intelligence (AI) applications.