You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
@AmielMetier@HediChennoufi@ocots improve inits of those problems in the list below that have _s versions; if they are here (although they have three versions, including an _s one) it is that, more or less randomly, optimisation with adnlp fails with a sincos domain error (value: -Inf)
@ocots not satisfying to have that many tests removed because of instabilities. what about re-including them, even if there are some fails (e.g. while generating the doc). FTR see control-toolbox/CTBenchmarks.jl#114 (comment) it concerns (at least):
insurance
robot
space_shuttle
steering
and also glider that converges on CPU but not on GPU; would be nice to improve init to have a converging run also on GPU, despite the convergence issue is also informative
Activity
jbcaillau commentedon Dec 9, 2025
@ocots not satisfying to have that many tests removed because of instabilities. what about re-including them, even if there are some fails (e.g. while generating the doc). FTR see control-toolbox/CTBenchmarks.jl#114 (comment) it concerns (at least):
and also glider that converges on CPU but not on GPU; would be nice to improve init to have a converging run also on GPU, despite the convergence issue is also informative
AmielMetier commentedon Dec 9, 2025
@jbcaillau Regarding the problem with inf, isn't it due to the initial value?
Do we need to perform a custom initialization for each model?
jbcaillau commentedon Dec 10, 2025
@AmielMetier probably, yes. having a better init (= closer to solution) should stabilise the process. to be discussed during next meeting