Login  Register

Re: Armijo optimization

Posted by Nicolas Di Césaré on Nov 28, 2002; 2:26pm
URL: http://quantlib.414.s1.nabble.com/Problems-in-Bermudan-Swaption-tp2297p2301.html

Le ven 22/11/2002 à 19:57, Nicolas Di Césaré a écrit :

> Le ven 22/11/2002 à 11:44, Marco Marchioro a écrit :
> > Hi,
> > I'm no optimization guru, but I found the following paper
> >
> > http://csep1.phy.ornl.gov/mo/node11.html
> >
> > which advices to use alpha=1.e-4 and beta=0.9 for Armijo and Goldstein
> > criteria.
> > At the moment QuantLib has default values alpha=0.5 and beta=0.65.
> > Should we change these default values?
> >
> > Did anybody ever tested this algorithm?
> >
>
> Hi Marco,
>
> I have extensively use this algorithm. I don't remember where I have
> found that default values. I will try to find more informations.
> The only thing, I remember is that beta > 0.5.
>
> I will have a look the article next week.
>

Hi Marco,

There is an error in the default value of alpha. I have probably done a
copy/paste error (alpha=0.05!). The theoric recommandations for alpha
and beta are the following :  0 < alpha < 0.5 < beta < 1.

Thus, the both (corrected ;-) couple of values are admissible!

We are dealing with numerical algorithm, so to know which one is
better depend on the smoothness of the minimized function.

Do you think that a curve fitting example could be helpfull in the
example section?



--
Nicolas Di Césaré <[hidden email]>