KNITRO solver interface

[eminyouskn]

I would like to begin implementing an interface for the Artelys KNITRO solver. According to the developer guidelines, such contributions should start with a discussion, so I’m opening this thread to initiate that conversation.

[metab0t]

I am glad to accept external contributions.

You can read existing interface code like Gurobi or COPT, then implement LP, QP, MIP firstly.

NLP support is more difficult and you can refer to the ipopt part to understand how POI provides the callbacks.

[eminyouskn]

Great, thank you! I’ll keep you updated on my progress and will add to this discussion if I run into any issues or have further questions.

[eminyouskn]

Quick question, since KNITRO supports quadratic objectives, is there a QuadraticObjectiveMixin<T> (similar to LinearObjectiveMixin<T>) that I should inherit from, or is that not necessary?

[metab0t]

It is not necessary. Implementing set_objective(const ScalarQuadraticFunction &function, ObjectiveSense sense) and set_objective(const ExprBuilder &function, ObjectiveSense sense) is enough.

[eminyouskn]

Thanks! That’s exactly what I understood.

[eminyouskn]

I managed to implement KNITRO and have all tests passed except one test_clnlbeam. KNITRO seems to find a different optimal solution other than the expected one. Is this an expected behaviour? Or is KNITRO just finding another local minimum?

Here is the output:

tests/test_nlp_clnlbeam.py::test_clnlbeam[knitro] FAILED                                                                                            [100%]

======================================================================== FAILURES =========================================================================
__________________________________________________________________ test_clnlbeam[knitro] __________________________________________________________________

nlp_model_ctor = <class 'pyoptinterface._src.knitro.Model'>

    def test_clnlbeam(nlp_model_ctor):
        model = nlp_model_ctor()
    
        N = 100
        h = 1 / N
        alpha = 350
    
        t = model.add_m_variables(N + 1, lb=-1.0, ub=1.0)
        x = model.add_m_variables(N + 1, lb=-0.05, ub=0.05)
        u = model.add_m_variables(N + 1)
    
        for i in range(N):
            with nl.graph():
                model.add_nl_objective(
                    0.5 * h * (u[i] * u[i] + u[i + 1] * u[i + 1])
                    + 0.5 * alpha * h * (nl.cos(t[i]) + nl.cos(t[i + 1]))
                )
                model.add_nl_constraint(
                    x[i + 1] - x[i] - 0.5 * h * (nl.sin(t[i]) + nl.sin(t[i + 1])) == 0.0
                )
                model.add_linear_constraint(
                    t[i + 1] - t[i] - 0.5 * h * u[i + 1] - 0.5 * h * u[i] == 0.0
                )
    
        model.optimize()
    
        objective_value = model.get_model_attribute(poi.ModelAttribute.ObjectiveValue)
    
>       assert objective_value == pytest.approx(350.0, abs=1e-8)
E       assert 328.0967067335499 == 350.0 ± 1.0e-08
E         
E         comparison failed
E         Obtained: 328.0967067335499
E         Expected: 350.0 ± 1.0e-08

tests/test_nlp_clnlbeam.py:35: AssertionError
------------------------------------------------------------------ Captured stdout call -------------------------------------------------------------------

=======================================
          Commercial License
         Artelys Knitro 16.0.0
=======================================

Knitro using 1 thread.
No start point provided -- Knitro computing one.

Knitro presolve eliminated 0 variables and 0 constraints.

feastol                  1e-06
feastol_abs              0.001
opttol                   1e-06
opttol_abs               0.001

Problem Characteristics                     |           Presolved
-----------------------
Problem type: NLP
Objective: minimize / general
Number of variables:                    303 |                           303
  bounds:         lower     upper     range |     lower     upper     range
                      0         0       202 |         0         0       202
                             free     fixed |                free     fixed
                              101         0 |                 101         0
Number of constraints:                  200 |                           200
                    eq.     ineq.     range |       eq.     ineq.     range
  linear:           100         0         0 |       100         0         0
  quadratic:          0         0         0 |         0         0         0
  nonlinear:        100         0         0 |       100         0         0
Number of nonzeros:
              objective  Jacobian   Hessian | objective  Jacobian   Hessian
  linear:             0       400           |         0       400          
  quadratic:          0         0         0 |         0         0         0
  nonlinear:        202       400       302 |       202       400       302
  total:            202       800       302 |       202       800       302
Coefficient range:
  linear objective:          [0e+00, 0e+00] |                [0e+00, 0e+00]
  linear constraints:        [5e-03, 1e+00] |                [5e-03, 1e+00]
  quadratic objective:       [0e+00, 0e+00] |                [0e+00, 0e+00]
  quadratic constraints:     [0e+00, 0e+00] |                [0e+00, 0e+00]
  variable bounds:           [5e-02, 1e+00] |                [5e-02, 1e+00]
  constraint bounds:         [0e+00, 0e+00] |                [0e+00, 0e+00]

Knitro using the Interior-Point/Barrier Direct algorithm.

    Iter       Objective  FeasError   OptError   ||Step||      Time 
--------  --------------  ---------  ---------  ---------  --------
       0    3.211072e+02   1.35e+00
      10    3.336790e+02   2.36e-06   1.52e+00   4.25e-01      0.01
      20    3.287653e+02   9.40e-05   5.57e-01   1.89e+00      0.01
      29    3.280967e+02   9.35e-12   4.51e-09   8.58e-03      0.02

EXIT: Locally optimal solution found.

Final Statistics
----------------
Final objective value               =   3.28096706733550e+02
Final feasibility error (abs / rel) =   9.35e-12 / 6.94e-12
Final optimality error  (abs / rel) =   4.51e-09 / 1.53e-09
# of iterations                     =         29 
# of CG iterations                  =          6 
# of function evaluations           =         33
# of gradient evaluations           =         32
# of Hessian evaluations            =         29
Total program time (secs)           =       0.01752 (     0.114 CPU time)
Time spent in evaluations (secs)    =       0.00499

================================================================================

================================================================= short test summary info =================================================================
FAILED tests/test_nlp_clnlbeam.py::test_clnlbeam[knitro] - assert 328.0967067335499 == 350.0 ± 1.0e-08
  
  comparison failed
  Obtained: 328.0967067335499
  Expected: 350.0 ± 1.0e-08
==================================================================== 1 failed in 0.12s ====================================================================

[metab0t]

Can knitro export the nl file and test it with ipopt to ensure the model is correct?

[eminyouskn]

Unfortunately, I can not export it to nl. However I checked that the constraint are satisfied and the objective value is correct and it seems another local minimum that KNITRO obtains.

[metab0t]

That is OK. You can add this new objective value to the test like assert obj == pytest.approx(350.0, abs=1e-8) or obj == pytest.approx(328.0967, abs=1e-4)

[eminyouskn]

Perfect thank you.

[eminyouskn]

I just created a PR for adding support for KNITRO. I think to be able to run tests you need probably to have a license. Please let me know how can I send it to you?

[metab0t]

Thanks for contribution.

Please send e-mail to yue.yang#hfut.edu.cn (replacing # to @)

[eminyouskn]

I just sent by email the license file.

[eminyouskn]

Hi @metab0t

Thanks for your quick reactions on my PRs. I just want to ask you if you want me to add testing KNITRO in the CI?

[metab0t]

Yes, I have added the secret variable KNITRO_LICENSE to include the content of license you sent to me.

[eminyouskn]

I just added KNITRO to the CI.

[eminyouskn]

Currently, we only verify whether KNITRO is installed before triggering the KNITRO-related tests. We do not check whether a valid license is available or accessible.

Do you think it would be worthwhile to add functionality that also validates license availability before running these tests?

[metab0t]

I am not opposed to adding this feature. It would be a quality-of-life improvement to avoid test failures caused by license issues, but checking the license state is a highly solver-specific task.

Feel free to submit a PR.

[eminyouskn]

I will work on it and add a PR.