[SciPy-user] Solver with n-dimentional steps

Michael McNeil Forbes mforbes@physics.ubc...
Sun Jun 15 14:28:19 CDT 2008


Yosef,

You can also try using the Brodyen method (the b in bfgs) directly to  
solve your problem.  You could try some of the methods  
scipy.optimize.broyden*, or look at my attached code.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: broyden.py
Type: text/x-python-script
Size: 9221 bytes
Desc: not available
Url : http://projects.scipy.org/pipermail/scipy-user/attachments/20080615/06b19b44/attachment.bin 
-------------- next part --------------


The Broyden method is a generalized secant method, so the Jacobian is  
slowly updated as the iteration proceeds.  It is useful if you cannot  
compute the Jacobian.  It starts as an iteration, so if you can  
arrange your equations so that they are expressed as an iteration x->F 
(x) that sort of converges, then this will work best, but it often  
works even if the iteration does not converge.

Michael.

On 15 Jun 2008, at 6:55 AM, Hoyt Koepke wrote:

> Yosef,
>
> I'm suggesting this partly to satisfy my own curiosity, so ignore this
> if it is difficult.   Since you are running with so many dimensions,
> you might want to try minimizing the square of your function using
> fmin_l_bfgs_b.  lbfgs is written for higher dimensions, and I've heard
> it can give incredible speed improvements over many other methods.
> I'm curious if it would help in your case.
>
> --Hoyt
>
> On Sun, Jun 15, 2008 at 1:41 PM, Yosef Meller  
> <yosefmel@post.tau.ac.il> wrote:
>> On Sunday 15 June 2008 13:19:54 David Warde-Farley wrote:
>>> It's doing finite differences to estimate the gradient. This is  
>>> pretty
>>> much unavoidable if you're only giving it those three arguments.  
>>> Have
>>> you thought about supplying fprime as a function to analytically
>>> compute the gradient?
>>

----------------------
Mailing address:

Michael McNeil Forbes
UW Dept. of Physics
Box 351560
Seattle, WA, 98195-1560

For couriers:
Physics/Astronomy Building, Room C121
3910 15th Ave NE
Seattle, WA, 98195-1560
Font Desk: (206)-543-2770

If you would like to visit me personally:
Room B482 (Fourth floor)
(206) 543-9754


More information about the SciPy-user mailing list