Note: This discussion is about an older version of the COMSOL Multiphysics® software. The information provided may be out of date.

Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.

Newton or Dogleg?

Please login with a confirmed email address before reporting spam

Hallo,

does anyone of you have experience with the new double dogleg solver (which is an option for
nonlinear solving and can replace the Newton-method)?

I have a nonlinear stationary problem (electromagnetic) where the damped Newton solver takes
~100 steps and uses a small damping factor for most of the steps. The dogleg solver
on the other side has fewer steps and they are mainly undamped Newton-steps. So
I wonder why the Newton solver is so bad.

But this does not generalise to all my cases, sometimes the Newton works just fine.

What are your experiences here? Are there configuration settings to push the Newton-solver to
higher damping factors? Is there a good literature reference explaining the double-dogleg
algorithm?

Thanks in advance.

Regards

Jens

1 Reply Last Post 2 oct. 2012, 09:07 UTC−4
Nagi Elabbasi Facebook Reality Labs

Please login with a confirmed email address before reporting spam

Posted: 1 decade ago 2 oct. 2012, 09:07 UTC−4
Hi Jens,

The new Double Dogleg solver according to the COMSOL documentation is a “sophisticated combination of the Steepest descent and Newton-Raphson methods”. The basic idea is that away from the solution (especially of nonlinear constrained problems) the Newton method can easily fail, while the gradient methods guarantee progress (even if slow) towards a minimum (could be a local one). At the vicinity of the solution however the Newton method has the best convergence rate. The new Double Dogleg solver intelligently combines the two methods to generate a good overall solver. It belongs to a broader type of solvers called Trust Region solvers. The credit of developing this algorithm it seems goes to this reference: www.springerlink.com/content/m33j465v5r565215/.

Comparing the Newton and Double Dogleg solvers for the problems I work on, I found that in some cases the new solver works much better and in some cases it didn’t and I switched back to the Newton method. I could have gone to COMSOL Support to try to fine tune the Double Dogleg settings for these problems, but it’s not worth it for me when another solver is working!

Nagi Elabbasi
Veryst Engineering
Hi Jens, The new Double Dogleg solver according to the COMSOL documentation is a “sophisticated combination of the Steepest descent and Newton-Raphson methods”. The basic idea is that away from the solution (especially of nonlinear constrained problems) the Newton method can easily fail, while the gradient methods guarantee progress (even if slow) towards a minimum (could be a local one). At the vicinity of the solution however the Newton method has the best convergence rate. The new Double Dogleg solver intelligently combines the two methods to generate a good overall solver. It belongs to a broader type of solvers called Trust Region solvers. The credit of developing this algorithm it seems goes to this reference: http://www.springerlink.com/content/m33j465v5r565215/. Comparing the Newton and Double Dogleg solvers for the problems I work on, I found that in some cases the new solver works much better and in some cases it didn’t and I switched back to the Newton method. I could have gone to COMSOL Support to try to fine tune the Double Dogleg settings for these problems, but it’s not worth it for me when another solver is working! Nagi Elabbasi Veryst Engineering

Note that while COMSOL employees may participate in the discussion forum, COMSOL® software users who are on-subscription should submit their questions via the Support Center for a more comprehensive response from the Technical Support team.