[SciPy-user] [SciPy-dev] question about scipy.optimize.line_search

[Matthieu Brucher]

I already told dmitrey, but I say it on list.
My optimizers have several choice for line searches, including strong Wolfe
Powell rules.

Matthieu

2007/6/28, Dominique Orban <dominique.orban at gmail.com>:
>
> Alan G Isaac wrote:
> > On Thu, 28 Jun 2007, Dmitrey apparently wrote:
> >
> >>help(line_search) yields
> >>--------------------------------------------------------------------
> >>line_search(f, myfprime, xk, pk, gfk, old_fval, old_old_fval, args=(),
> >>c1=0.0001, c2=0.90000000000000002, amax=50)
> >>    Find alpha that satisfies strong Wolfe conditions.
> >>    Uses the line search algorithm to enforce strong Wolfe conditions
> >>    Wright and Nocedal, 'Numerical Optimization', 1999, pg. 59-60
> >>    For the zoom phase it uses an algorithm by
> >>    Outputs: (alpha0, gc, fc)
> >>--------------------------------------------------------------------
> >>So I need to know what are other args, especially gfk (is it a gradient
> >>in point xk?), old_fval, old_old_fval (I guess I know what do c1 & c2
> mean)
>
> >
> > This is certainly lacking documentation!  A little is here:
> >
> http://docs.neuroinf.de/api/scipy/scipy.optimize.optimize-pysrc.html#line_search
> > Can anyone help Dmitrey more?
>
> Each iteration of a linesearch procedure to satisfy the strong Wolfe
> conditions requires an evaluation of f and of its gradient. I have no
> idea who coded this and I don't have the book handy this moment, but I
> would guess gk is the gradient of the objective at the current trial
> point. No clue about the old_val and old_old_val (doesn't look like my
> dream programming style).
>
> Enforcing the strong-Wolfe conditions is not an easy task, is a
> sensitive process, and the algorithm presented in the book is certainly
> simplified as much as possible for clarity of exposition. For more
> robust software, you would be better off using the implementation of
> More and Thuente
>
> Moré, J. J. and Thuente, D. J. 1994. Line search algorithms with
> guaranteed sufficient decrease. ACM Trans. Math. Softw. 20, 3 (Sep.
> 1994), 286-307. DOI= http://doi.acm.org/10.1145/192115.192132
>
> This is Fortran software which you could interface. I did the job in
> NLPy (http://nlpy.sf.net). You should be able to reuse my interface.
>
> Dominique
> _______________________________________________
> SciPy-user mailing list
> SciPy-user at scipy.org
> http://projects.scipy.org/mailman/listinfo/scipy-user
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.scipy.org/pipermail/scipy-user/attachments/20070629/39b1b3a3/attachment.html>