OpenModelica
Improve the logic to select initial guesses in the hybrid solver
In #11010, an interesting case is discussed where the hybrid solver needs to handle a nonlinear system with initial guesses that are not good enough. In particular, there is one variable which has no start attribute (hence it starts from zero), a nominal value of 1 (which should be meant for scaling!) and min = 0, max = 0.05.
The solver first tries with 0, then it apparently uses the nominal attribute 1 as a second best candidate guess value. The problem is, that value is outside the min/max bound, and it causes a singularity.
I would suggest to improve the hybrid solver by always limiting initial guess values with the min and max attributes.
I would suggest to improve the hybrid solver by always limiting initial guess values with the min and max attributes.
#12201 changes the behavior so that iteration variables are always clamped between their min/max before every newton step, not only at the initial guess(es).
Excellent! I'm curious to see if this improves some simulation in the library testsuite. Thanks!
Currently only hybrid. But I will look at it again and try to find a better place to do the clamping. I'd prefer to do it at a single place in the code, but perhaps this needs to be copied to several places...
OK. In fact, this could be a double-edged sword, since in some cases min-max attributes could be a bit bogus, as demonstrated by the few regressions we got.
Ideally, we should try with clamping first, and if if fails, retry without.