[Trilinos-Users] Convergence criteria in Belos

[erhan turan]

I think I found it. In BelosStatusTestFactory.hpp you can define the
following options for" Explicit Residual Scaling"

"Norm of Initial Residual"
"Norm of Preconditioned Initial Residual"
"Norm of RHS" / "Norm of Right-Hand Side"
"None" (no scaling)

In my case,

belosList->set( "Explicit Residual Scaling", "Norm of RHS" );

should be given. I'll test this option now.  Hope this helps other users;
these keywords are not listed in the documentation.

Erhan


On Thu, Apr 12, 2012 at 2:43 PM, erhan turan <turane at gmail.com> wrote:

> Dear All,
>
> I have difficulty in defining a convergence criteria other than the
> relative residual test in Belos. In AztecOO, we have the option to select
> among various convergence options using "AZ_conv": relative, absolute
> residual test or any others. Specifically what I need is AZ_rhs where the
> comparison is made against the 2-norm of RHS (i.e norm of b in Ax=b).
> Normally, if you have a solver that works with relative residuals only,
> then you might multiply your convergence tolerance with the initial norm
> and use that modified tolerance as the new criteria. And even in my case,
> it is still possible to modify the tolerance further with norm of b. That's
> fine but I don't want that.
>
> If these options are somewhat available but hidden within Belos, I would
> like to hear that. If not, do you have any suggestions? I believe,
> statusTestNorm could be replaced somehow using the
> statusTestOutputFactory.hpp file but not sure how to implement it. I was
> not able to relate it to AztecOO_StatusTest class as suggested in
> DesignRequirements.txt in doc folder of Belos source. I also noticed that
> in some Belos examples, there are additional operations after Solve() to
> calculate the residual explicitly. Those cases gave me the impression that
> only relative residual is available in Belos even if there is this belos
> option  "Explicit Residual Test" which is handy in left preconditioned
> problems. It would be nice of you if the residual operations in Belos are
> clarified somehow.
>
> Regards.
>
> Erhan
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20120412/6cf8990a/attachment.html