Back away from inequality constraints that evaluate to NaN #38
Conversation
|
Oops, this change is not passing all of my tests yet. |
|
Actually these changes are good. The problem I was having was caused by use of ftol instead of xtol on a constrained problem. |
|
How does this look? |
|
Hm, do I need to save/restore slsqpb_state if I pick a point from the line search that is not the last point? |
jpritikin
force-pushed
the
master
branch
2 times, most recently
from
f232260 to
229e1ff
Compare
|
@jpritikin, regarding the SAVE/RESTORE_STATE: if I remember correctly, the original Fortran code was using the equivalent of |
|
Could you squash these (via |
This question was addressed with jpritikin@356bd27 |
|
Should the |
+ Preserve starting value if no evaluations were finite + Separately track the best minor and major estimate. This change ensures that we always move to the best point found during the line search even when the best point is not the last point. We also avoid 1 fit and gradient evaluation by moving the major iteration convergence check above the switch statement. + Treat failure to find a search direction as convergence + Only update minor during line search + Unconfuse code that returns the converged details + Rename 'x' to 'theSpot'. I loath single letter variable names. X marks the spot.
Sorry, I can't remember. Certainly the code doesn't hurt. It helps preserve an invariant condition that should be preserved. |
|
hello @jpritikin, do you think this has a chance to land someday or do you think we can close it ? |
|
I personally have given up. The changes are part of OpenMx. I believe the version of the code in OpenMx is superior to what I started with when I copied it from here, but I haven't kept track of nlopt and I have also ceased to contribute to OpenMx. |
|
thanks |
This permits slsqp to deal with the interaction between inequality constraints and soft feasibility constraints.