IPOPT termination condition

[ajoardar007]

Hello,

While trying to solve a NLP using IPOPT (Matlab), I am unable to get the solver to terminate naturally. Upon looking through the IPOPT docs, I see that there are 4 criteria's, including:

'tol' <1e-08
'dual_inf_tol' <1
'constr_viol_tol' <1e-04
'compl_inf_tol' < 1e-04
Going back to the output that IPOPT drops by default, I was trying to match as many of the above points as possible to understand why the solver has still not terminated.
Is it appropriate for me to draw the following conclusions:

Can 'tol' (NLP error) be inferred from the 'objective' output?
Can 'dual_inf_tol' be inferred as the 'inf_du' output (infinity-norm (max) of the internal dual infeasibility)?
Can 'compl_inf_tol' be inferred as the 'inf_pr' output (infinity-norm (max) of the (unscaled) constraints)?
I am getting these kinds of outputs.

iter        objective                 inf_pr          inf_du         lg(mu)       ||d||            lg(rg)   alpha_du        alpha_pr       ls
1020     9.2803977e+000    5.69e-010    4.86e-001      -5.7      5.71e-007    5.9       1.00e+000     1.00e+000h  1
1021     9.2803977e+000    2.45e-010    5.53e-001      -5.7      2.43e-007    6.4       1.00e+000     1.00e+000h  1
1022     9.2803975e+000    1.22e-009    9.39e-001      -5.7      1.24e-006    5.9       1.00e+000     1.00e+000h  1
1023     9.2803975e+000    5.47e-010    1.09e+000     -5.7      5.40e-007    6.3       1.00e+000     1.00e+000h  1

While there is hardly any change in the objective, and the constraints are well met, the solver is still not stopping. Is there some way to display the above 4 criterions at the end of every iteration?
I am using 'casadi-windows-matlabR2016a-v3.5.5'.

Thanks,
Arnab

[svigerske]

The log is explained in the docu: https://coin-or.github.io/Ipopt/OUTPUT.html

Ipopt doesn't stop because the dual infeasibility (inf_du) is still above dual_inf_tol.

You cannot see the NLP error or complementarity violation from the log.
If you set option print_level to 6, you should get for each iteration a print of all violations (NLP error, primal, dual, complementarity) for the scaled and unscaled problem, like the one you see when Ipopt terminates. That values can be better matched to the various termination tolerances.