- changed status to closed
Better default for DoglegParams.deltaInitial
I propose a better value for the initial trust region radius of the DoglegOptimizer, which is defined by DoglegParams.deltaInitial. To close issue #452, it was suggested to change the default value of deltaInitial from 1 to 10. Indeed, that worked for the posted problem, but it appears that deltaInitial is defined in an absolute spatial scale. This new default value of deltaInitial causes a failure of the dogleg optimizer if the scale of the problem is multiplied by 10 (i.e. the vehicles are 6000 units apart instead of 600, they travel 30 units per time step instead of 3, and the standard deviations on distance traveled and inter-vehicle range are set to 0.1 units instead of 0.01)
It seems to me that it would be better to define the deltaInitial relative to the size of the full Gauss-Newton step on the first iteration (which is what I thought it meant when the default was set to 1, which is why I never thought to increase it). As far as I know, it is generally recommended to try the full GN step on the first iteration. However, changing the meaning of this parameter may surprise some folks that have set it manually for their application.
Comments (1)
-
- Log in to comment
GTSAM moved to github - I re-opened the issue there as https://github.com/borglab/gtsam/issues/9