پیوندهای مفید
آمار
| تعداد نشریات | 33 |
| تعداد شمارهها | 874 |
| تعداد مقالات | 8,497 |
| تعداد مشاهده مقاله | 53,679,820 |
| تعداد دریافت فایل اصل مقاله | 9,773,617 |
Learning the initial step size and backtracking reduction factor for the Armijo line search | ||
| Journal of Mathematical Modeling | ||
| مقالات آماده انتشار، اصلاح شده برای چاپ، انتشار آنلاین از تاریخ 27 مرداد 1405 اصل مقاله (295.3 K) | ||
| نوع مقاله: Research Article | ||
| شناسه دیجیتال (DOI): 10.22124/jmm.2026.33801.3107 | ||
| نویسنده | ||
| Ahmad Kamandi* | ||
| University of science and technology of Mazandaran, Behshahr, Iran | ||
| چکیده | ||
| In this paper, we study the sensitivity of the Armijo backtracking line search to its two important parameters, the initial step size and the reduction factor, and propose a GRU-based recurrent model to estimate them adaptively in the steepest descent method. In the structure of the GRU model, we use a low-dimensional feature vector and safeguard the output so as not to lose the convergence properties. The model is trained on a set of strongly convex quadratic problems with varying condition numbers, enabling it to learn patterns that govern effective step-size selection. Numerical experiments show that the learned methods consistently outperform the conventional baselines, namely a grid-search tuned fixed-parameter Armijo rule and two adaptive heuristic strategies, in terms of both iteration counts and function evaluations. | ||
| کلیدواژهها | ||
| Armijo line search؛ initial step size؛ backtracking reduction factor؛ GRU-based recurrent model | ||
|
آمار تعداد مشاهده مقاله: 5 تعداد دریافت فایل اصل مقاله: 2 |
||