Next: Minimizing the Huber function
Up: Algorithm for minimizing the
Previous: The limited memory BFGS
The solver works as follows:
- 1.
- Choose , l, . Set k=0.
- 2.
- Compute
| |
(79) |
| (80) |
where meets the Wolfe conditions.
- 3.
- Let =min, . Update times using the pairs
,
i.e., let
| |
|
| |
| (81) |
| |
| (82) |
- 4.
- Set k=k+1 and go to 2 if the residual power is not small enough.
The update is not formed explicitly; instead we compute
with an iterative formula
Nocedal (1980). Liu and Nocedal (1989) propose scaling the initial
symmetric positive definite at each iteration as follows:
| |
(83) |
This scaling greatly improves the performances of the method.
Liu and Nocedal (1989) show that the storage limit for large-scale problems
has little effects.
A common choice for l is l=5. In practice, the initial guess
for the Hessian is the identity matrix ;
then it might be scaled as proposed in equation (). The
nonlinear solver as detailed in the previous algorithm converges to
a local minimizer of .
Next: Minimizing the Huber function
Up: Algorithm for minimizing the
Previous: The limited memory BFGS
Stanford Exploration Project
5/5/2005