Float objective with CP-SAT

[Yury Semenov]

Hi,

I have a model with purely integer variables and constraints, and a linear objective function with float coefficients (such objective functions are supported by CP-SAT). When I try to solve the model, I get "CP-SAT does not support float variables" error. As far as I understand, this happens because minizinc creates an auxiliary float variable for the objective function. Is there any way to change this behavior and feed the model with float objective function to CP-SAT without creating float variables?

[guido.tack]

No, that's not possible within the MiniZinc architecture, since the objective is always passed to the solver in the form of a variable. The CP-SAT backend would have to detect that all the float variables in the model are only used to define the objective, which shouldn't be hard to do. As far as I understand, the CP-SAT solver scales floating point objectives to integers internally, so it may be easier to change your MiniZinc model to use coefficients that are scaled up to integers.

Cheers,

Guido

[Yury Semenov]

Thank you, I'll try it.

вторник, 28 мая 2024 г. в 02:47:27 UTC+3, guido.tack:

[Max Ostrowski]

Just as a side remark, I think that there is some lost real estate here.
SAT based CP solvers could definitely benefit if they get a a weighted set of variables as input to their objective function.
It is much harder to undo this extra variable and it can hinder propagation.

Best,
Max

--
You received this message because you are subscribed to the Google Groups "MiniZinc" group.
To unsubscribe from this group and stop receiving emails from it, send an email to minizinc+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/minizinc/2b3323d4-9113-476f-8858-920eb9d98d37n%40googlegroups.com.

[Laurent Perron]

For CP-SAT, this is equivalent as presolve expand the objective. 

To view this discussion on the web visit https://groups.google.com/d/msgid/minizinc/ea201231-c486-4ad4-8b13-045ef7a74980%40potassco.com.