Are there any hidden differences in OR-Tools' default parameters for a MIP model in Python vs C++?

Question

or-tools-discuss cross post here.

As a disclaimer -- I am very aware that OR-Tools' Python code are just wrappers around their C++ source code. I'm just getting an odd discrepancy between a model setup in Python and one setup in C++ and am wracking my brain to figure out what I've messed up (because its most likely me not implementing something correctly or missing a "hidden-to-me" difference between the two).

Setup Context

OS

• Pop!OS 22.04 (basically Ubuntu)
• x86_64

Python Environment

• Python 3.10.5
• ortools 9.6.2534

C++ Environment

• CMake 3.26.4
• gcc 11.3.0
• or-tools FetchContent'd from stable Git branch

Problem Context

I've got two scripts, one in Python and the other in C++, that are setup to do (what I believe) is the same thing: read in an MPS file and solve the contained MIP using the SCIP solver.

The MPS file being fed into both scripts is the same, so I would have expected them to produce similar results. I might have even expected the C++ program to run faster than its Python counterpart. But what I'm seeing is that the Python script finished in ~20 seconds or so, while the C++ script errors out during the solve process after 5+ minutes.

The C++ error produced was (effectively): ERROR: (node 1) unresolved numerical troubles in LP 2 cannot be dealt with, which is mentioned in this question.

Upon inspecting the logs from both solvers, I see they are both coming to the same primal / dual bounds. But whereas the Python script reaches those bounds and exits with an Optimal solution, the C++ script continues to iterate until error-ing out with the above issue.

I'm wondering if it could be related to the way I've read in the MPS file and loaded it into a model? The process for each is laid out below:

Python

• use ortools.linear_solver.python.model_builder.ModelBuilderand ModelSolver
• instantiate objects model = ModelBuilder() and solver = ModelSolver("SCIP")
• model.import_from_mps_file to read MPS
• solver.solve(model) to optimize MIP

C++

• #include <ortools/lp_data/mps_reader.h> and #include <ortools/linear_solver/linear_solver.h>
• instantiate solver `solver = MPSolver::CreateSolver("SCIP")
• read MPS file to proto model_proto = glop::MpsFileToMPModelProto(mps_file)
• load model from proto solver->LoadModelFromProto(*model_proto, ...)
• solver->Solve()

So, in Python, I'm using the ModelHelpers and in C++ I'm interfacing directly with the MPSolver class. Perhaps this is where the difference lies?

Any thoughts / suggestions as to why I'm seeing this difference would be much appreciated!

Answer 1

As explained on or-tools-discuss, model_builer and MPSolver use 2 different sets of parameter s. In particular the tolerances ans the gaps.

I recommend using settings the same solver specific parameters as strings on both front ends.