nloptr SLSQP returning successful flag despite constraints being unmet

[ssmil]

Hello,

I'm having an issue where nlopt is returning a successful exit flag, but only 1 of my 6 constraints are met. I am running many optimizations at once and using the exit flags as automatic checks, so this is causing me some significant issues. I've attached a file with the code I'm running.

I'm using nloptr version 1.2.2.2, algorithm "NLOPT_LD_SLSQP".

iteration: 16
x = ( 1.000000, 1.000000, 1.552908, 1.855128 )
f(x) = -1.907547
h(x) = ( 15.437643, 5.305731 )
g(x) = ( -2.609576, 0.328356, 0.500000, 3.871769 )

rest.CI.GI$status
[1] 3

nloptr flag issue.zip

[jcmonnin]

I have also encountered this issue in the past, and it looks like others had that too:
https://www.mail-archive.com/nlopt-discuss@ab-initio.mit.edu/msg00708.html

Below are some links about this issue in the Python implementation (which is based on the same fortran code):
scipy/scipy#7282
scipy/scipy#7618
scipy/scipy#8986

I have not tried to understand the python fix, however I have been using a simple fix which seems to do the trick, altough it's probably far from optimal (pull request #465).

Alternatively to

       if (!nlopt_isinf(fprev) && feasible) {

I wonder if following is better

       if (!nlopt_isinf(fprev) && feasible_cur) {

In the latter, it waits until a new feasible solution that meets the contraint before deciding on the ftol/xtol termination criteria.

2f81804 from an old pull request is also vaguely related to this.