How to turn all integer variables into reals?

[SYury]

Hi,

I have a code that creates a MILP model using linear_solver API. Sometimes I need to solve this model, and sometimes I need to solve its LP relaxation with integer variables turned into real variables. How to do this with as little additional code as possible?

Thank you.

[sschnug]

Disclaimer:

• I never used this API
• I think it's depreated

That being said:

• A: In 99.9% of use-cases it's much more clever (and predictable) to just add a is_model_relaxed boolean-flag to your model-building code
  • This can imply that you do model your problem twice which might seem strange, but this approach is still the correct one in many use-cases (most safe and most explicit)

Howewer:

• According to linear_solver.proto#L59 there are two more options:
  • B: Just switch / mutate solver-type to one only supporting continuous models
    • Might lead to use of a different solver which in some corner-cases can lead to problems (e.g. different behaviour in different numercial-trouble situations; but the whole MILP -> LP alone might have some effects)
  • C: Just transform / mutate all variable-messages:
    • optional bool is_integer = 4 [default = false]; -> false for every variable!
      • Either directly on protobufs
      • Or over your model-builder API:
        • void SetInteger(bool integer);
        • const std::vector<MPVariable*>& variables() const { return variables_; }

I'm in strong favor of A, then C.
(C is what happens in most solver-implementations afaik when they support: solve_relaxation like Cbc)

[SYury]

Thank you. If there's no one-liner to do this, I'll probably stick with A and make a solver wrapper that substitutes IntVar with NumVar when needed.

[lperron]

there is a MPVariable::unrounded_solution_value()

[SYury]

Thank you, but as far as I understand, in this case I'll still have to solve MILP model even if I need relaxation. The overhead seems too large for me :(

[lperron]

no, solve with glop or clp Laurent Perron | Operations Research | ***@***.*** | (33) 1 42 68 53 00 Le mar. 4 mars 2025 à 17:47, Yury Semenov ***@***.***> a écrit :

…

Thank you, but as far as I understand, in this case I'll still have to solve MILP model even if I need relaxation. The overhead seems too large for me :( — Reply to this email directly, view it on GitHub <#4578 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACUPL3PLH73ULQCIOIW767D2SXKL5AVCNFSM6AAAAABYGXCJXOVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTEMZZGEYDQMI> . You are receiving this because you commented.Message ID: ***@***.***>

[SYury]

Thank you, I'll try it. I didn't know LP solvers could accept integer variables.

Btw, is unrounded_solution_value available in Python wrapper? Reference only mentions solution_value.

[sschnug]

Probably explained by linear_solver.proto#L60 and i would argue this approach is the aforementioned "B".

[lperron]

I checked, it does not. I will add it. Laurent Perron | Operations Research | ***@***.*** | (33) 1 42 68 53 00 Le mar. 4 mars 2025 à 18:06, Yury Semenov ***@***.***> a écrit :

…

Thank you, I'll try it. I didn't know LP solvers could accept integer variables. Btw, is unrounded_solution_value available in Python wrapper? Reference only mentions solution_value. — Reply to this email directly, view it on GitHub <#4578 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACUPL3PRJDAWSNQLX2MZOZL2SXMR3AVCNFSM6AAAAABYGXCJXOVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTEMZZGEZDSOA> . You are receiving this because you commented.Message ID: ***@***.***>