SPECFEM / specfem3d

SPECFEM3D_Cartesian simulates acoustic (fluid), elastic (solid), coupled acoustic/elastic, poroelastic or seismic wave propagation in any type of conforming mesh of hexahedra (structured or not).
GNU General Public License v3.0
398 stars 225 forks source link

USE_SEPARATE_CG_STEPLENGTHS deprecated #426

Closed rmodrak closed 9 years ago

rmodrak commented 9 years ago

Has anyone encountered a situation for which this feature is useful? I’m not sure it’s a good idea to include it.

In most cases, it is better to invert multiple parameters simultaneously rather than one at a time. For example, it is better to invert source and structure through a simultaneous inversion rather than through alternating decoupled iterations.

Put another way, treating each parameter one at a time is a kind of coordinate descent approach. There is some applied math literature showing that these methods tend to perform badly.

komatits commented 9 years ago

Indeed it could probably be removed. Vadim and Yi, could you confirm that you never use it and that Ryan can remove it?

Thanks, Dimitri.

On 02/19/2015 05:04 AM, rmodrak wrote:

In most cases, is better to invert multiple parameters simultaneously rather than one at a time. For example, it is better to invert source and structure through a simultaneous inversion rather than through alternating decoupled iterations. The underlying idea--to treat each parameter separately, one at a time--is also the basis for coordinate descent methods, which tend to perform quite poorly.

Has anyone encountered a situation for which this feature is useful? I’m not sure it’s a good idea to include it.

— Reply to this email directly or view it on GitHub https://github.com/geodynamics/specfem3d/issues/426.

Dimitri Komatitsch CNRS Research Director (DR CNRS), Laboratory of Mechanics and Acoustics, UPR 7051, Marseille, France http://komatitsch.free.fr

rmodrak commented 9 years ago

A comment about this feature has been added above the parameter definition in constants_tomography. If it's alright, I'll go ahead and close this issue.