Set aircraft-derived upper bounds on mass and distance states - #1055
Set aircraft-derived upper bounds on mass and distance states#1055bdphilli wants to merge 1 commit into
Conversation
|
@robfalck , I didn't realize dymos set any constraints if the user specifies there are no bounds to a state. Do you think we'd get easier convergence if we set the bound guesses as Ben suggests? If so, maybe we should we reevaluate all instances where we set bounds to None? |
|
The way some optmizers work is that having a bound value beyond +/- 1E21 or 1E30 (depends on the optimizer), they treat it as "unbounded" and handle it a bit differently internally. I think the default behavior in OpenMDAO should be that bounds of None. One of the real issues with this is that the scaling forces the upper and lower bounds to cross the "treat this as unbounded" threshold. This is fixed in the autoscaler pull request on OpenMDAO. |
When mass_bounds or distance_bounds have None as the upper bound (the default), Dymos sets it to ~1e21. With mass_ref=1e4 and distance_ref=1e6, this creates scaled upper bounds of ~1e17 and ~1e15 respectively, while other design variables have scaled ranges of O(1). In build_phase(), replace None upper bounds with values derived from aviary_inputs: 2x GROSS_MASS for mass, 2.5x DESIGN_RANGE for distance. This reduces scaled upper bounds to O(10), matching the scale of other design variables. User-specified bounds are preserved — the override only applies when the upper bound is None. Addresses OpenMDAO#607
8ee643b to
edb0443
Compare
|
@robfalck are the None upper bounds still an issue that requires manual fixes like in this PR? |
|
For the time being, we had some outside contributions come in that conflicted at the time. Let me see if I can get it PR'd today. |
4 participants
When mass_bounds or distance_bounds have None as the upper bound (the default), Dymos sets it to ~1e21. With mass_ref=1e4 and distance_ref=1e6, this creates scaled upper bounds of ~1e17 and ~1e15 respectively, while other design variables have scaled ranges of O(1).
In build_phase(), replace None upper bounds with values derived from aviary_inputs: 2x GROSS_MASS for mass, 2.5x DESIGN_RANGE for distance. This reduces scaled upper bounds to O(10), matching the scale of other design variables.
User-specified bounds are preserved — the override only applies when the upper bound is None.
Resolves #607