custom constraint or-tools

[Nassim Bb]

Hello , 
I am a Phd student and i would like to create my own global constraint using or-tools , can you please refer me to some documentations on how to do it since the one that exist does not provide much information on how to do it. By the way my constraint is for itemset mining.
Thank you 
Nassim

[Laurent Perron]

Look at ortools/constraint_solver/constraint_solver.i and makefiles/Makefile.python.mk

and of course swig.org

Please note this is a nightmare.

--Laurent

--
You received this message because you are subscribed to the Google Groups "or-tools-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to or-tools-discu...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

--Laurent

[Nassim Bb]

Thank you for your help 

[Laurent Perron]

I assumed you wanted to use the original constraint solver (ortools/constraint_solver).

We do recommend using the new CP-SAT solver (ortools/sat/cp_model.h).

The only (big) downside is that you cannot write new constraints with this solver.

[Nassim Bb]

Actually i only want to create a custom constraint for filtering, you told me there is an example in  or-tools/examples/cpp/dobble_ls.cc 

[Yahia Lebbah]

But, adding new constraints, is an essences of CP. Don't understand why such key feature is omitted ...
Anyway, is the old original constraint solver (where we can added propagators) still maintained ?

Best regards,

Yahia Lebbah

[Laurent Perron]

Yes, it is still maintained. But we will not work on it to improve performance, except for the routing library.

The essence of CP is to solve problems with a high level modeling language.

We can add new constraints to the CP-SAT solver, it is just too complex, and there are no tools to debug them :-)

[J. E. Marca]

Is the lack of tools to debug constraints a PhD level problem, or is it just a programming cleverness level problem?

[Laurent Perron]

PhD level, when writing a constraint for the CP-SAT solver, you need to explain all propagation.

The key to performance it to write a minimal explanation.

The problem is that if the explanation is wrong, the effect may be visible much later during search. You may find a optimal value too optimal, or not enough...

As everything is delayed, it is hard to poinpoint the culprit code.