【Numberical Optimization】3 Line Search Methods (zen学习笔记)


1.Step Length


1.1 Wolfe conditions:

  • Armijo condition(保证\alpha _{_{k}}p_{_{k}}方向上的下降区间内 ):f\left ( x_{k} +\alpha _{k}p_{k}\right )\leq f(x_{k})+c_{_{1}}\alpha _{_{k}}\bigtriangledown f_{k}^{T}p_{k}
  • curvature condition (保证步长\alpha _{_{k}}足够大):\bigtriangledown f(x_{k}+\alpha _{k}p_{k})^{T}p_{k}\geq c_{2}\bigtriangledown f_{k}^{T}p_{k}

 strong Wolfe conditions:

  • 在 Wolfe conditions 修正了曲率条件: \left | \bigtriangledown f(x_{k}+\alpha _{k}p_{k})^{T}p_{k}\right|\geq \left|c_{2}\bigtriangledown f_{k}^{T}p_{k}\right|

1.2 The Goldstein conditions 

 

1.3 BLS 


2. Convergence of Line search methods


 Zoutendijk condition:    

                                                       \sum_{k=0}^{\infty}cos^{2}\theta _{k}\left \| \bigtriangledown f_{k} \right \|^{2}< \infty

If the angle \theta_{k} is bounded away from 90^{\circ} ,then

                                                                                \lim_{k\rightarrow \infty}\left \| \bigtriangledown f_{k} \right \|=0

 


3.Rate of Convergence


3.1 Q-convergence:

                                                                                \lim_{k\rightarrow \infty}\frac{\left | x_{k+1}-x^{*} \right |}{\left|x_{k}-x^{*}\right|^{\alpha }}=c

 

  • \alpha =1,0< c< 1 : linear convergence
  • \alpha =1,c=0 : super linear convergence
  • \alpha =2,c=constant : Qudractic convergence
  • \alpha >2,0< c< \infty :  \alpha阶收敛

 3.2 几个常见方法的收敛率


4.Step-Length Selection Algotithms


 

猜你喜欢

转载自blog.csdn.net/weixin_38716567/article/details/82889845