LP model solves with CLP and PDLP but fails using GLOP

[kasvas-ee]

What version of OR-Tools and what language are you using?
Version: 9.7.2996
Language: Python

Which solver are you using (e.g. CP-SAT, Routing Solver, GLOP, BOP, Gurobi)
GLOP, CLP, PDLP

What operating system (Linux, Windows, ...) and version?
Windows

What did you do?
Steps to reproduce the behavior:

1. Created a solver instance: solver = pywraplp.Solver('Model', pywraplp.Solver.GLOP_LINEAR_PROGRAMMING)
2. Built the LP model (LP file exported and attached)
3. Set solver parameters to reduce tolerance:

    params = pywraplp.MPSolverParameters()
    params.SetDoubleParam(pywraplp.MPSolverParameters.PRIMAL_TOLERANCE, 0.001)
    params.SetDoubleParam(pywraplp.MPSolverParameters.DUAL_TOLERANCE, 0.001)

5. Called ret = solver.solve(params)

What did you expect to see
ret value as zero from solver.solve, indicating optimal solution.

What did you see instead?
ret as 4 from solver.solve, which indicates 'Abnormal / Error'

Make sure you include information that can help us debug (full error message, model Proto).
Following zipped file contains the LP model in LpFormat:
lp_file.zip

Anything else we should know about your project / environment
Using PDLP solver solves the LP model with an objective function value of 44116.22. CLP solver solves the model too.

[lperron]

glop is much stricter w.r.t numerical errors.

Have you enabled logging ?

[kasvas-ee]

I see. No, I have not enabled logging. I am using ortools on python, is there a way to enable logging in Python too? Could you please share some link I could refer to enable logging?

[kasvas-ee]

@lperron I tried searching in the other discussions for a way to enable logging in python, but could not find the right syntax. Could you please share the syntax or a thread that discusses this?

[lperron]

solver.EnableOutput()

[lperron]

https://or-tools.github.io/docs/pdoc/ortools/linear_solver/pywraplp.html#Solver.EnableOutput

[kasvas-ee]

Thank you @lperron .
I enabled the logs and noticed the cost perturbation was probably larger than the solution checker tolerance. Logs pasted below. Is there a syntax in Python to raise the 'solution_feasibility_tolerance' to 1e-5? I tried looking in 'pywraplp.MPSolverParameters' for the feasibility tolerance setting, but could only find 'PRIMAL_TOLERANCE' and 'DUAL_TOLERANCE'.

Final unscaled solution:
Primal objective (before moving primal/dual values) = 4.407207591129694E+04
Dual objective (before moving primal/dual values) = 4.407207538915282E+04
Max. primal values move = 9.09495e-13
Max. dual values move = 9.16677e-17
Primal objective (after moving primal/dual values) = 4.407207591129694E+04
Max. rhs perturbation = 8.18545e-12
Max. cost perturbation = 2.05641e-06
Max. primal infeasibility = 5.45697e-12
Max. dual infeasibility = 1.20905e-14
Objective error <= 0.060193
The needed cost perturbation is too large !!
status: IMPRECISE
objective: 44072.1
iterations: 2504
time: 0.0960571
deterministic_time: 0.220503

[lperron]

#1868 (comment) Laurent Perron | Operations Research | ***@***.*** | (33) 1 42 68 53 00 Le jeu. 2 mai 2024 à 14:12, kasvas-ee ***@***.***> a écrit :

…

Thank you @lperron <https://github.com/lperron> . I enabled the logs and noticed the cost perturbation was probably larger than the solution checker tolerance. Logs pasted below. Is there a syntax in Python to raise the 'solution_feasibility_tolerance' to 1e-5? I tried looking in 'pywraplp.MPSolverParameters' for the feasibility tolerance setting, but could only find 'PRIMAL_TOLERANCE' and 'DUAL_TOLERANCE'. Final unscaled solution: Primal objective (before moving primal/dual values) = 4.407207591129694E+04 Dual objective (before moving primal/dual values) = 4.407207538915282E+04 Max. primal values move = 9.09495e-13 Max. dual values move = 9.16677e-17 Primal objective (after moving primal/dual values) = 4.407207591129694E+04 Max. rhs perturbation = 8.18545e-12 Max. cost perturbation = 2.05641e-06 Max. primal infeasibility = 5.45697e-12 Max. dual infeasibility = 1.20905e-14 Objective error <= 0.060193 The needed cost perturbation is too large !! status: IMPRECISE objective: 44072.1 iterations: 2504 time: 0.0960571 deterministic_time: 0.220503 — Reply to this email directly, view it on GitHub <#4192 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACUPL3NDZT6EQQV2NOD2WY3ZAIUUDAVCNFSM6AAAAABHB5JTSSVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TEOJUHEZTE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[kasvas-ee]

Thanks for the links. This thread helped me better understand the log, and I tried to mimic the steps to modify the feasibility tolerance.
I could not figure out the Python equivalent of this java code from the thread:
solver.setSolverSpecificParametersAsString(operations_research.glop.Parameters.GlopParameters.newBuilder().setSolutionFeasibilityTolerance(1e-3).build().toString());

Eventually I tried this, which seemed to work:
solver.SetSolverSpecificParametersAsString("solution_feasibility_tolerance:1e-5")

[lperron]

and https://or-tools.github.io/docs/pdoc/ortools/linear_solver/pywraplp.html#Solver.SetSolverSpecificParametersAsString Laurent Perron | Operations Research | ***@***.*** | (33) 1 42 68 53 00 Le jeu. 2 mai 2024 à 14:14, Laurent Perron ***@***.***> a écrit :

…

#1868 (comment) Laurent Perron | Operations Research | ***@***.*** | (33) 1 42 68 53 00 Le jeu. 2 mai 2024 à 14:12, kasvas-ee ***@***.***> a écrit : > Thank you @lperron <https://github.com/lperron> . > I enabled the logs and noticed the cost perturbation was probably larger > than the solution checker tolerance. Logs pasted below. Is there a syntax > in Python to raise the 'solution_feasibility_tolerance' to 1e-5? I tried > looking in 'pywraplp.MPSolverParameters' for the feasibility tolerance > setting, but could only find 'PRIMAL_TOLERANCE' and 'DUAL_TOLERANCE'. > > Final unscaled solution: > Primal objective (before moving primal/dual values) = 4.407207591129694E+04 > Dual objective (before moving primal/dual values) = 4.407207538915282E+04 > Max. primal values move = 9.09495e-13 > Max. dual values move = 9.16677e-17 > Primal objective (after moving primal/dual values) = 4.407207591129694E+04 > Max. rhs perturbation = 8.18545e-12 > Max. cost perturbation = 2.05641e-06 > Max. primal infeasibility = 5.45697e-12 > Max. dual infeasibility = 1.20905e-14 > Objective error <= 0.060193 > The needed cost perturbation is too large !! > status: IMPRECISE > objective: 44072.1 > iterations: 2504 > time: 0.0960571 > deterministic_time: 0.220503 > > — > Reply to this email directly, view it on GitHub > <#4192 (reply in thread)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/ACUPL3NDZT6EQQV2NOD2WY3ZAIUUDAVCNFSM6AAAAABHB5JTSSVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TEOJUHEZTE> > . > You are receiving this because you were mentioned.Message ID: > ***@***.***> >

[lperron]

yes, this is the write syntax. The string uses a json syntax. The parameters are in ortools/glop/parameters.proto Laurent Perron | Operations Research | ***@***.*** | (33) 1 42 68 53 00 Le jeu. 2 mai 2024 à 14:57, kasvas-ee ***@***.***> a écrit :

…

Thanks for the links. This thread helped me better understand the log, and I tried to mimic the steps to modify the feasibility tolerance. I could not figure out the Python equivalent of this java code from the thread: solver.setSolverSpecificParametersAsString(operations_research.glop.Parameters.GlopParameters.newBuilder().setSolutionFeasibilityTolerance(1e-3).build().toString()); Eventually I tried this, which seemed to work: solver.SetSolverSpecificParametersAsString("solution_feasibility_tolerance:1e-5") — Reply to this email directly, view it on GitHub <#4192 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACUPL3IJWGDKULDJU5S53U3ZAIZ2DAVCNFSM6AAAAABHB5JTSSVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TEOJVGM2TE> . You are receiving this because you were mentioned.Message ID: ***@***.***>