Optimizing Neural Networks: A Novel Conjugate Gradient Method Based on Perry’s Parameter and Logistic Mapping
2
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
The conjugate gradient method is well-known for its very low memory requirements, as it does not need to compute or store large matrices. It remains a popular choice among researchers addressing large-scale optimization problems. In this paper, we propose a modified conjugate gradient approach based on Perry’s conjugate gradient parameter combined with a logistic mapping formula. Furthermore, we apply the new method to neural network training. The proposed algorithm satisfies key theoretical properties, including descent conditions, sufficient decrease, and global convergence guarantees. Numerical experiments demonstrate that our novel conjugate gradient method outperforms classical methods such as Perry’s method in terms of efficiency and competitiveness. The performance com parison is conducted using a set of common benchmark problems of varying dimensions, with evaluation metrics including the total number of function evaluations and iterations. Additionally, the effectiveness of the proposed method is validated through numerical results on neural network training for trigonometric function approximation.
Abstract
The conjugate gradient method is well-known for its very low memory requirements, as it does not need to compute or store large matrices. It remains a popular choice among researchers addressing large-scale optimization problems. In this paper, we propose a modified conjugate gradient approach based on Perry’s conjugate gradient parameter combined with a logistic mapping formula. Furthermore, we apply the new method to neural network training. The proposed algorithm satisfies key theoretical properties, including descent conditions, sufficient decrease, and global convergence guarantees. Numerical experiments demonstrate that our novel conjugate gradient method outperforms classical methods such as Perry’s method in terms of efficiency and competitiveness. The performance com parison is conducted using a set of common benchmark problems of varying dimensions, with evaluation metrics including the total number of function evaluations and iterations. Additionally, the effectiveness of the proposed method is validated through numerical results on neural network training for trigonometric function approximation.