一个带重启步的改进PRP型谱共轭梯度法
An Improved PRP Type Spectral Conjugate Gradient Method with Restart Steps
通讯作者:
收稿日期: 2021-01-18
基金资助: |
|
Received: 2021-01-18
Fund supported: |
|
作者简介 About authors
江羡珍,E-mail:
廖伟,E-mail:
毋晓迪,E-mail:
The Polak-Ribière-Polak algorithm is considered one of the most efficient methods among classical conjugate gradient methods (CGMs). To generate new conjugate parameter, an improved PRP formula is proposed by combining the strong Wolfe line search condition. Furthermore, a new spectral parameter and a new restart direction are designed, and thus a new spectral conjugate gradient method with restart steps is established. Using the strong Wolfe line search condition to yield the step length, the sufficient descent property and global convergence of the new algorithm are obtained under the general assumptions. Finally, for the proposed algorithm, a medium-large scale numerical experiments is performed, and compared with some existing efficient CGMs, the numerical results show that the proposed algorithm is very promising.
Keywords:
本文引用格式
江羡珍, 廖伟, 简金宝, 毋晓迪.
Jiang Xianzhen, Liao Wei, Jian Jinbao, Wu Xiaodi.
1 引言
考虑无约束优化问题
其中
其中
PRP方法自带重启功能, 即算法在迭代过程中一旦出现小步长, 下一步迭代即以最速下降方向进行重启, 从而有效避免算法连续产生小步长, 进而提高算法计算效率. PRP方法这个性质的严格表述最早见文献[8], 并称之为性质(
本文继续研究PRP方法的改进, 目的是建立结构简单、计算效率高、收敛性好的PRP型谱共轭梯度法. 全文余下部分安排如下: 第2部分结合Wolfe线搜索准则对PRP公式进行改进, 并基于新共轭参数设计新的谱参数, 引入重启条件并构造新的重启方向, 建立新算法; 第3部分证明算法的下降性和全局收敛性; 第4部分对算法进行中大规模数值实验并分析报告测试结果.
本节最后给出本文用于产生步长的线搜索准则—强Wolfe线搜索准则, 即
其中
2 改进PRP参数公式和新算法
Dai和Yuan[19]提出一类单参数共轭梯度法簇, 其共轭参数为
显然, 当
受(2.1)式分母结构的启发, 结合强Wolfe线搜索准则(1.2), 本文考虑将PRP公式改进为
显然, 当
谱思想较早出现在Barzilai和Borwein[20]提出的两点步长梯度法(简称BB方法)中, 其每步线搜索仅使用负梯度方向, 即
由(2.3)式不难发现, 若
谱参数
另一方面, Kou和Dai[23]提出一个带新型重启方向的改进三项共轭梯度法(简称KD方法), 其重启方向为
其中
当
基于搜索方向(2.5)及强Wolfe线搜索准则(1.2), 下面给出本文的算法步骤(简记为
初始步 任取初始点
步骤1 若
步骤2 采用强Wolfe线搜索条件(1.2)计算步长
步骤3 按
步骤4 令
3 JLJW+ 算法的下降性和收敛性分析
为了获得
(H1) 目标函数
(H2) 目标函数
以下引理表明, 在强Wolf线搜索条件下,
引理3.1 算法
对
证 当
(i) 当搜索方向
(ii) 当
若
若
由归纳假设, 可知
结合(3.5)和(3.6)式可得
根据(3.4)和(3.7)式可知, 搜索方向
将上式两边同时除以
此结合归纳假设
综合(3.2)和(3.8)式可知充分下降性(3.1)式对
基于
定理3.1 若假设
证 用反证法. 假设
一方面, 由(1.2)式, (3.1)式和假设(H2)可得
进而
另一方面, 由
(i) 当搜索方向
其中
(ii) 当搜索方向
将
根据(3.10)式和(3.12)式可知
其中
将(3.13)式两边同时除以
其中
结合(3.9)和(3.15)式得
另外, 由于
这与调和级数
4 数值试验
本节, 通过三组数值实验测试新方法的有效性. 第一组, 为测试改进的PRP型方法的有效性, 将JLJW方法与HS方法[1], PRP方法[3, 4]和LS方法[6]进行对比; 第二组, 为测试新型重启方向的有效性, 将
其中
本文算法的终止准则为以下两种情形之一
(1)
(2) 迭代次数
另外, 终止准则(2) 出现时认为该方法对相应例子失效, 并记为"F".
在试验中, 我们分别对迭代次数(Itr), CPU计算时间(Tcpu)及算法终止时梯度值
图 1
图 2
图 3
图 4
表 1 数值试验报告
Problems | JLJW+ | KD | DK | HZ | SPF2 |
Name/n | Itr/Tcpu/||g*|| | Itr/Tcpu/||g*|| | Itr/Tcpu/||g*|| | Itr/Tcpu/||g*|| | Itr/Tcpu/||g*|| |
bard 3 | 1430/4.93/7.6e-07 | 1565/5.19/8.0e-07 | F/F/3.2e-05 | 620/2.00/4.0e-07 | 235/0.64/6.6e-07 |
beale 2 | 631/0.96/3.6e-07 | 326/0.52/7.7e-07 | 86/0.11/1.6e-07 | 251/0.37/6.2e-07 | 142/0.20/4.3e-07 |
box 3 | 150/0.26/2.1e-07 | 330/0.57/6.0e-07 | 134/0.26/1.4e-08 | 475/0.89/8.9e-07 | 91/0.13/5.2e-07 |
cosine 300 | 19/0.03/5.2e-07 | 26/0.09/3.3e-07 | F/F/1.2e-04 | 32/0.06/4.9e-08 | F/F/3.0e+02 |
cosine 1500 | 1560/11.26/1.5e-07 | F/F/1.4e+03 | F/F/2.8e+03 | F/F/1.2e+02 | F/F/1.8e+03 |
cosine 4500 | F/F/2.9e+02 | F/F/5.8e-04 | F/F/1.2e+04 | 477/11.19/2.9e-07 | F/F/2.1e+03 |
dixmaana 3000 | 18/0.99/3.4e-07 | 17/0.90/3.1e-07 | 23/1.22/6.0e-07 | 26/1.75/1.3e-07 | 24/1.38/1.9e-07 |
dixmaana 12000 | 18/3.14/6.8e-07 | 17/2.93/6.2e-07 | 25/4.16/1.1e-07 | 21/3.70/1.6e-07 | 24/4.05/3.7e-07 |
dixmaanb 3000 | 11/0.39/7.4e-07 | 12/0.50/3.3e-07 | 34/1.76/1.7e-07 | 49/3.92/9.6e-08 | 14/0.61/1.5e-07 |
dixmaanb 12000 | 11/1.17/4.4e-07 | 12/1.51/5.8e-07 | 47/10.10/9.0e-08 | 40/10.30/6.5e-07 | 13/1.55/2.7e-07 |
dixmaanc 3000 | 24/1.48/3.8e-08 | 25/1.54/5.5e-07 | 29/1.62/2.4e-07 | 33/2.20/1.9e-07 | 35/2.18/8.0e-07 |
dixmaanc 12000 | 16/2.37/1.0e-06 | 25/4.38/4.5e-07 | 27/4.39/8.2e-07 | 73/18.86/5.6e-07 | 67/15.50/9.6e-07 |
dixmaand 3000 | 25/1.28/1.3e-07 | 25/1.32/3.6e-07 | 29/1.56/3.3e-07 | 54/4.45/2.3e-07 | 45/2.73/6.7e-07 |
dixmaand 12000 | 22/3.41/5.5e-07 | 23/3.67/6.5e-07 | 44/8.42/7.7e-07 | 29/5.73/7.2e-07 | 50/10.73/2.7e-07 |
dixmaane 6000 | 1308/199.45/9.0e-07 | 1229/192.04/9.3e-07 | 1941/331.01/8.5e-07 | 1716/274.31/9.3e-07 | 564/82.65/7.7e-07 |
dixmaanf 4500 | 775/94.63/8.3e-07 | 1258/161.65/7.9e-07 | 396/53.77/8.5e-07 | 1165/155.96/9.0e-07 | 470/59.77/7.1e-07 |
dixmaanf 9000 | 1080/237.42/9.3e-07 | 1078/246.31/8.8e-07 | 1396/340.53/9.4e-07 | F/F/1.8e-05 | 383/80.92/4.1e-07 |
dixmaang 7500 | 889/168.72/8.8e-07 | 1154/220.67/5.8e-07 | 1239/250.95/8.9e-07 | 1682/327.37/9.0e-07 | 614/112.05/9.5e-07 |
dixmaanh 4500 | 1417/179.40/9.2e-07 | 1324/169.60/9.8e-07 | 1006/137.94/7.2e-07 | 953/120.53/9.9e-07 | 704/81.78/9.8e-07 |
dixmaani 120 | 1930/9.09/8.7e-07 | 1595/7.68/6.7e-07 | F/F/1.2e-03 | F/F/8.4e-07 | 731/3.29/4.2e-07 |
dixmaanj 2700 | 1193/94.28/8.5e-07 | 1382/108.65/6.2e-07 | 1320/110.36/9.5e-07 | F/F/1.0e-05 | 1153/85.27/6.4e-07 |
dixmaank 3000 | 1142/98.31/6.5e-07 | 1158/100.76/7.8e-07 | F/F/4.0e-05 | F/F/2.7e-05 | 1799/154.65/6.6e-07 |
dixmaanl 300 | 1554/14.02/9.8e-07 | 1424/12.64/8.9e-07 | F/F/4.6e-04 | F/F/8.0e-05 | 826/7.12/7.4e-07 |
dixon3dq 50 | 1199/1.50/8.8e-07 | 1613/2.28/6.0e-07 | F/F/5.9e-05 | 1836/2.41/5.3e-07 | 629/0.82/6.0e-07 |
dixon3dq 88 | 1934/2.80/8.3e-07 | 1476/2.09/6.7e-07 | F/F/1.0e-02 | F/F/4.9e-04 | 1227/1.65/5.2e-07 |
dqdrtic 60000 | 496/35.92/7.8e-07 | 426/31.92/2.8e-07 | 197/13.41/4.1e-07 | 917/67.92/7.1e-07 | 312/21.78/6.2e-07 |
dqrtic 100 | 22/0.06/5.7e-07 | 23/0.04/6.4e-07 | 34/0.04/7.7e-07 | 28/0.03/1.8e-07 | 60/0.12/1.8e-07 |
dqrtic 450 | 34/0.13/2.8e-07 | 47/0.19/3.4e-07 | 33/0.15/7.4e-07 | 31/0.13/5.0e-07 | 60/0.36/5.7e-07 |
edensch 10000 | 46/4.59/9.5e-07 | 54/6.25/6.2e-07 | F/F/8.7e-06 | F/F/3.8e-06 | F/F/3.3e-06 |
edensch 50000 | 55/24.21/1.9e-07 | 65/29.17/8.8e-07 | F/F/1.0e-04 | F/F/1.8e-05 | F/F/1.2e-04 |
edensch 100000 | 126/93.46/9.7e-07 | F/F/4.8e-05 | F/F/2.3e-05 | 120/139.38/5.3e-07 | F/F/2.0e-04 |
eg2 30 | F/F/2.9e-06 | F/F/3.0e-01 | 313/0.43/8.0e-07 | F/F/3.2e-03 | F/F/1.1e-05 |
eg2 80 | F/F/2.5e-04 | F/F/5.8e-06 | F/F/4.0e-01 | F/F/2.8e-02 | F/F/1.9e-05 |
engval1 6 | 87/0.11/1.9e-07 | 95/0.14/2.4e-08 | F/F/4.7e-01 | F/F/8.8e-01 | F/F/1.8e-01 |
fletchcr 10000 | 145/3.81/1.3e-07 | F/F/1.1e-03 | F/F/3.1e-04 | F/F/9.5e-04 | F/F/6.9e-04 |
fletchcr 300000 | 211/118.78/3.9e-07 | F/F/9.2e-03 | F/F/6.2e-03 | F/F/9.5e-04 | F/F/2.6e-02 |
freuroth 20 | 1700/2.52/9.8e-07 | 476/0.66/9.2e-07 | F/F/1.5e-05 | 455/0.60/6.8e-07 | F/F/6.8e-05 |
freuroth 36 | F/F/2.2e-06 | F/F/6.3e-06 | F/F/8.2e-06 | F/F/6.5e-05 | F/F/7.5e-05 |
genrose 2000 | 574/2.28/2.8e-07 | 399/1.94/4.1e-07 | 395/1.77/6.1e-07 | 1259/5.68/9.6e-07 | 692/2.85/4.4e-07 |
genrose 47000 | 270/25.44/6.7e-07 | 336/31.54/1.3e-07 | 474/42.83/7.1e-07 | 854/76.03/9.0e-07 | 571/48.69/8.7e-07 |
gulf 3 | 2/0.00/0.0e+00 | 2/0.00/0.0e+00 | 2/0.00/0.0e+00 | 2/0.00/0.0e+00 | 2/0.00/0.0e+00 |
helix 3 | 479/1.02/1.6e-07 | 282/0.62/4.2e-07 | 316/0.60/7.0e-07 | 731/1.56/6.7e-07 | F/F/3.5e+03 |
himmelbg 1000 | 3/0.01/1.6e-28 | 3/0.00/1.3e-28 | 3/0.00/1.6e-28 | 3/0.00/1.2e-28 | 3/0.00/1.6e-28 |
himmelbg 10000 | 3/0.01/5.2e-28 | 3/0.01/4.2e-28 | 3/0.01/5.0e-28 | 3/0.01/3.8e-28 | 3/0.01/5.0e-28 |
himmelbg 100000 | 3/0.10/1.6e-27 | 3/0.11/1.3e-27 | 3/0.12/1.6e-27 | 3/0.11/1.2e-27 | 3/0.12/1.6e-27 |
kowosb 4 | 774/1.40/7.0e-07 | 817/1.55/4.7e-07 | 1574/3.07/8.7e-07 | F/F/5.2e-04 | 239/0.43/7.5e-07 |
liarwhd 500 | 487/1.06/1.9e-07 | 750/1.63/9.4e-07 | 1168/2.52/8.5e-07 | F/F/2.9e-03 | 268/0.49/1.2e-07 |
liarwhd 1000 | 534/1.43/2.3e-07 | F/F/1.3e-04 | F/F/1.6e-02 | 1267/3.35/9.8e-07 | 294/0.75/4.7e-07 |
liarwhd 10000 | F/F/5.6e-02 | F/F/5.8e+00 | F/F/3.2e+03 | F/F/1.5e+03 | 409/8.96/8.9e-07 |
nondquar 4 | F/F/1.2e-04 | F/F/5.0e-05 | F/F/4.9e-03 | 395/0.66/2.2e-07 | 87/0.12/6.5e-07 |
penalty1 1000 | 15/0.82/1.2e-07 | 15/0.82/1.2e-07 | 15/0.86/1.2e-07 | 15/0.87/1.2e-07 | 15/0.89/1.2e-07 |
penalty1 10000 | 9/28.41/9.6e-07 | 9/28.32/9.6e-07 | 9/28.02/9.6e-07 | 9/28.38/9.6e-07 | 9/28.46/9.6e-07 |
quartc 100 | 22/0.04/5.7e-07 | 23/0.04/6.4e-07 | 34/0.05/7.7e-07 | 28/0.05/1.8e-07 | 60/0.13/1.8e-07 |
quartc 450 | 34/0.14/2.8e-07 | 47/0.20/3.4e-07 | 33/0.14/7.4e-07 | 31/0.12/5.0e-07 | 60/0.31/5.7e-07 |
表 2 数值试验报告(续)
Problems | JLJW+ | KD | DK | HZ | SPF2 |
Name/n | Itr/Tcpu/||g*|| | Itr/Tcpu/||g*|| | Itr/Tcpu/||g*|| | Itr/Tcpu/||g*|| | Itr/Tcpu/||g*|| |
tridia 200 | 1491/2.44/3.2e-07 | 1380/2.41/9.8e-07 | 1699/2.90/9.3e-07 | F/F/5.9e-06 | 593/0.94/5.5e-07 |
tridia 400 | 1905/3.90/5.7e-07 | F/F/1.1e-03 | F/F/8.8e-03 | F/F/3.8e-05 | 733/1.29/6.9e-07 |
sinquad 3 | 406/0.57/4.7e-07 | 272/0.34/8.9e-07 | F/F/3.1e-04 | 501/0.63/3.0e-07 | 366/0.45/9.2e-07 |
vardim 8 | 12/0.01/4.0e-08 | 12/0.01/4.0e-08 | 12/0.01/4.0e-08 | 12/0.01/4.0e-08 | 12/0.02/4.0e-08 |
watson 3 | 107/0.26/2.6e-07 | 147/0.36/6.3e-07 | 51/0.12/2.1e-07 | 121/0.32/5.8e-07 | 108/0.31/9.8e-07 |
woods 10000 | F/F/1.3e-02 | 792/20.82/9.5e-07 | 883/22.24/6.2e-07 | 922/22.70/8.8e-07 | 309/7.19/9.9e-07 |
bdexp 1000 | 3/0.01/5.9e-107 | 3/0.00/7.1e-108 | 3/0.00/4.4e-107 | 3/0.00/2.5e-108 | 3/0.00/4.5e-107 |
bdexp 10000 | 3/0.01/1.2e-109 | 3/0.01/9.4e-110 | 3/0.01/1.1e-109 | 3/0.01/8.4e-110 | 3/0.01/1.1e-109 |
bdexp 100000 | 3/0.12/1.7e-109 | 3/0.13/1.7e-109 | 3/0.15/1.7e-109 | 3/0.15/1.7e-109 | 3/0.12/1.7e-109 |
exdenschnf 1000 | 25/0.05/4.6e-07 | 68/0.18/2.4e-07 | 65/0.17/5.2e-07 | 91/0.35/3.8e-07 | 96/0.28/9.7e-07 |
exdenschnf 10000 | 26/0.56/3.1e-07 | 68/2.43/7.6e-07 | 73/2.21/2.5e-07 | 85/2.85/2.6e-07 | 843/33.85/8.8e-07 |
exdenschnf 100000 | 27/4.41/1.1e-07 | 69/15.80/6.0e-07 | 73/14.86/7.8e-07 | 59/12.45/2.2e-07 | 114/25.09/2.1e-07 |
mccormak 2 | 21/0.05/9.3e-08 | 19/0.01/8.1e-07 | 63/0.07/9.0e-08 | 39/0.06/1.2e-07 | F/F/2.9e+04 |
exdenschnb 1000 | 22/0.03/8.2e-08 | 27/0.05/8.0e-07 | 74/0.19/4.5e-07 | 32/0.05/3.5e-07 | 102/0.27/4.8e-07 |
exdenschnb 10000 | 22/0.23/2.6e-07 | 28/0.40/7.1e-07 | 82/1.38/3.2e-07 | 25/0.29/1.0e-06 | 153/2.80/8.0e-07 |
exdenschnb 100000 | 22/1.58/8.2e-07 | 30/2.59/2.6e-07 | 88/9.13/2.7e-07 | 29/2.19/4.5e-07 | 91/9.42/1.5e-07 |
genquartic 5000 | 24/0.20/4.7e-07 | 31/0.32/9.0e-08 | 108/1.30/2.6e-07 | 33/0.22/8.6e-07 | 115/1.37/2.6e-07 |
genquartic 30000 | 35/1.33/1.9e-07 | 38/1.37/2.0e-07 | 96/4.61/5.6e-07 | 36/1.43/7.0e-08 | 157/8.53/2.4e-07 |
genquartic 100000 | 27/2.41/7.1e-08 | 37/3.79/5.2e-07 | 98/13.02/3.0e-07 | 81/11.20/8.8e-07 | 182/26.69/4.3e-07 |
biggsb1 100 | 1240/1.82/6.9e-07 | 1381/1.97/9.8e-07 | F/F/2.9e-05 | F/F/7.2e-06 | 621/0.82/9.4e-07 |
sine 4000 | 64/1.65/7.7e-09 | 180/4.19/1.3e-08 | F/F/8.0e-02 | 91/2.17/3.0e-07 | F/F/1.1e+03 |
sine 20000 | 94/7.82/1.7e-09 | 181/14.39/6.0e-13 | F/F/9.7e-03 | 81/7.52/9.5e-07 | F/F/7.6e-04 |
sine 40000 | 87/13.53/2.3e-08 | 394/67.01/7.9e-07 | F/F/5.0e-01 | 80/13.64/3.2e-07 | F/F/2.1e+03 |
fletcbv3 8 | 43/0.04/5.3e-07 | 51/0.05/4.7e-07 | 47/0.05/7.7e-07 | 30/0.03/7.4e-07 | 52/0.05/3.9e-07 |
nonscomp 2000 | 176/0.65/1.3e-07 | F/F/3.0e-03 | 100/0.37/4.2e-07 | F/F/4.6e-03 | F/F/5.0e-05 |
nonscomp 10000 | 275/6.17/6.5e-07 | F/F/6.2e-02 | F/F/3.8e-02 | 413/9.78/4.7e-07 | F/F/1.2e-04 |
power1 55 | 1699/2.30/9.5e-07 | F/F/5.1e-06 | F/F/1.4e-04 | F/F/3.6e-04 | 737/0.84/9.5e-07 |
raydan1 2000 | 857/2.94/6.9e-07 | 1118/3.86/6.5e-07 | F/F/1.4e-04 | 1335/4.78/7.8e-07 | F/F/9.2e-05 |
raydan1 3000 | 1252/8.69/7.5e-07 | F/F/2.5e-04 | F/F/2.6e-04 | F/F/3.8e-04 | F/F/2.3e-04 |
raydan2 500 | 15/0.02/2.2e-07 | 18/0.02/1.2e-08 | 24/0.04/2.6e-08 | 21/0.03/8.5e-08 | 14/0.02/6.6e-08 |
raydan2 5000 | 18/0.21/1.6e-07 | 21/0.28/7.2e-08 | 46/0.71/6.6e-07 | 38/0.56/2.7e-07 | 12/0.10/3.6e-07 |
raydan2 50000 | 18/1.44/8.9e-07 | 19/1.65/5.7e-07 | 21/1.89/9.8e-07 | 22/1.90/9.3e-07 | 48/5.15/5.0e-07 |
diagonal1 60 | 164/0.23/7.3e-07 | 88/0.09/9.4e-07 | 84/0.08/8.2e-07 | F/F/2.4e-06 | 179/0.28/8.7e-07 |
diagonal1 100 | 242/0.34/1.7e-07 | F/F/2.6e-05 | F/F/3.3e-06 | 348/0.57/9.7e-07 | F/F/5.5e-06 |
diagonal2 2000 | 1118/6.78/7.2e-07 | 1049/6.59/9.3e-07 | 334/1.97/6.9e-07 | 1988/12.48/7.6e-07 | 444/2.32/3.9e-07 |
diagonal3 150 | F/F/2.3e-05 | 194/0.40/7.9e-07 | F/F/8.4e-06 | F/F/6.0e-06 | F/F/4.3e-05 |
bv 2000 | 13/5.24/3.3e-07 | 14/5.79/3.1e-07 | 22/9.65/2.3e-07 | 9/3.23/4.6e-07 | 123/53.66/5.8e-07 |
bv 20000 | 1/0.14/1.2e-08 | 1/0.00/1.2e-08 | 1/0.00/1.2e-08 | 1/0.00/1.2e-08 | 1/0.00/1.2e-08 |
ie 100 | 15/1.33/2.8e-07 | 13/1.12/6.2e-07 | 20/1.94/5.0e-07 | 14/1.34/7.5e-07 | 38/4.96/6.6e-07 |
ie 200 | 15/4.47/3.9e-07 | 13/3.56/7.2e-07 | 18/5.35/3.1e-07 | 17/5.68/9.3e-08 | 40/17.36/3.2e-07 |
singx 200 | 348/2.28/7.4e-07 | 1218/7.82/6.3e-07 | F/F/3.5e-03 | F/F/7.0e-06 | 339/1.98/9.9e-07 |
singx 1500 | 1105/318.57/6.4e-07 | 1940/559.01/9.8e-07 | F/F/6.3e-03 | F/F/1.8e-05 | 292/75.74/9.0e-07 |
band 3 | 111/0.23/1.3e-07 | 67/0.12/6.3e-07 | 38/0.06/4.2e-07 | 77/0.11/6.9e-07 | 46/0.05/8.9e-08 |
gauss 3 | 59/0.16/5.6e-07 | 47/0.10/6.4e-07 | 15/0.02/1.7e-07 | 14/0.03/8.1e-07 | 30/0.08/7.4e-07 |
jensam 2 | 201/0.35/7.7e-07 | 119/0.19/6.7e-07 | 153/0.24/5.3e-07 | 211/0.39/7.0e-07 | 173/0.29/9.7e-07 |
lin 100 | 2/0.01/2.0e-14 | 2/0.00/2.0e-14 | 2/0.00/2.0e-14 | 2/0.00/2.0e-14 | 2/0.00/2.0e-14 |
lin 1000 | 12/83.01/1.3e-07 | 12/83.63/1.3e-07 | 12/82.60/1.3e-07 | 12/81.73/1.3e-07 | 12/82.49/1.3e-07 |
osb2 11 | 1717/6.76/8.4e-07 | F/F/2.5e-03 | F/F/8.2e-04 | F/F/6.4e-04 | F/F/4.9e-03 |
pen1 55 | 626/1.47/6.2e-07 | 448/1.02/9.2e-07 | 1475/3.47/9.5e-07 | 839/1.99/4.7e-07 | 106/0.18/1.5e-08 |
pen2 100 | F/F/4.1e-05 | 274/2.25/6.2e-07 | F/F/1.1e-05 | 430/3.57/9.0e-07 | F/F/1.6e-05 |
rosex 40 | F/F/7.3e-04 | 1077/2.00/5.3e-07 | 1496/2.58/9.4e-07 | 828/1.45/5.5e-07 | 141/0.18/7.1e-07 |
sing 4 | 545/0.97/4.7e-07 | 380/0.60/6.7e-07 | F/F/5.0e-04 | F/F/5.5e-06 | 254/0.50/3.4e-08 |
trid 500 | F/F/7.2e-05 | 899/55.29/1.0e-06 | F/F/1.5e-02 | 410/24.09/8.6e-07 | F/F/1.4e-04 |
trid 1000 | 500/76.37/7.1e-07 | 211/29.37/3.9e-07 | F/F/4.5e-04 | 754/111.54/8.3e-07 | 391/55.48/9.2e-07 |
图 5
图 6
参考文献
Method of conjugate gradient for solving linear equations
,DOI:10.6028/jres.049.044 [本文引用: 2]
Function minimization by conjugate gradients
,DOI:10.1093/comjnl/7.2.149 [本文引用: 1]
Note surla convergence de directions conjugées
,
The conjugate gradient method in extreme problems
,DOI:10.1016/0041-5553(69)90035-4 [本文引用: 2]
A nonlinear conjugate gradient method with a strong global convergence property
,DOI:10.1137/S1052623497318992 [本文引用: 1]
Efficient generalized conjugate gradient algorithms, part 1:theory
,DOI:10.1007/BF00940464 [本文引用: 1]
Global convergence properties of conjugate gradient methods for optimization
,
Convergence properties of algorithms for nonlinear optimization
,
An improved Polak-Ribiére-Polyak conjugate gradient method with an efficient restart direction
,
An efficient modified PRP-FR hybrid conjugate gradient method for solving unconstrained optimization problems
,DOI:10.1016/j.apnum.2019.06.003 [本文引用: 1]
A modified descent Polak-Ribiŕe-Polyak conjugate gradient method with global convergence property for nonconvex functions
,
A spectral conjugate gradient method for solving large-scale unconstrained optimization
,DOI:10.1016/j.camwa.2018.10.002 [本文引用: 1]
A modified nonlinear Polak-Ribiére-Polyak conjugate gradient method with sufficient descent property
,
Global convergence of some modified PRP nonlinear conjugate gradient methods
,DOI:10.1007/s11590-010-0224-8 [本文引用: 3]
A three term Polak-Ribiére-Polyak conjugate gradient method close to the memoryless bfgs quasi-newton method
,DOI:10.3934/jimo.2018149 [本文引用: 1]
A new conjugate gradient method based on Quasi-Newton equation for unconstrained optimization
,DOI:10.1016/j.cam.2018.10.035 [本文引用: 2]
A class of globally convergent conjugate gradient methods
,DOI:10.1360/03ys9027 [本文引用: 2]
Two-point step size gradient methods
,DOI:10.1093/imanum/8.1.141 [本文引用: 2]
A spectral conjugate gradient method for unconstrained optimization
,DOI:10.1007/s00245-001-0003-0 [本文引用: 1]
A modified self-scaling memoryless Broyden-Fletcher-Goldfard-Shanno method for unconstrained optimization
,DOI:10.1007/s10957-014-0528-4 [本文引用: 2]
A nonlinear conjugate gradient algorithm with an optimal property and an improved Wolfe line search
,DOI:10.1137/100813026 [本文引用: 1]
A new conjugate gradient method with guaranteed descent and an efficient line search
,DOI:10.1137/030601880 [本文引用: 1]
CUTEr and SifDec: A constrained and unconstrained testing environment, revisited
,DOI:10.1145/962437.962439 [本文引用: 1]
Testing unconstrained optimization software
,DOI:10.1145/355934.355936 [本文引用: 1]
An unconstrained optimization test functions collection
,
Benchmarking optimization software with performance profiles
,DOI:10.1007/s101070100263 [本文引用: 1]
/
〈 | 〉 |