[Coin-discuss] OSI branchAndBound vs CBC

[Matthew Saltzman]

Matt-

Good thoughts.  Could you file tickets so I don't lose these ideas?

Thanks.

		Matt


On Sat, 2009-05-23 at 19:45 -0400, Matthew Galati wrote:
> Thanks All.
> 
> This 
>    const char *args[] = {"Cbc_Test", "-solve", "-quit"};
>    CbcModel model(*solver.getRealSolverPtr());
>    CbcMain0(model);
>    CbcMain1(3, args, model);
> 
> seems to work much better than OsiCbc::branchAndBound( ) !
> 
> This is a very odd construction. Maybe I asked this before, but why
> don't we just have a CbcModel::model.solve( ) method? Why is there not
> some simple API like this for solving a MILP with CBC?
> 
> Moreover - for OSI users, why don't we have an OsiCbc::solveMilp( )
> method? The base class can have solveMilp( ) virtual and for those
> OsiXxx's that can solve MILPs (OsiCpx, OsiXpr, OsiCbc, ...), it gets
> derived, else it returns an error. I understand the reluctance to add
> MILP support in OSI when OSI2 was being discussed -- but that was many
> years ago. As an "interim" fix, can we at least make OSI work for
> MILPs - since many people are already using them to define and (try
> to) solve MILPs.
> 
> Thanks,
> Matt
> 
> 
> 
> 
>         Hi, 
>         
>         > I have a similar problem, which you helped me with
>         previously:
>         >
>         http://list.coin-or.org/pipermail/cbc/2008-February/thread.html#99
>         >
>         > Just a quick summary: I have a model in OSI and need to get
>         solution
>         > there as well:
>         >   OsiCbcSolverInterface solver;
>         
>         
>         Better use "OsiClpSolverInterface solver;" here.
>         I think Cbc do some special stuff if the instance comes as an
>         OsiClp object. 
>         
>         >   const char *args[] = {"Cbc_Test", "-solve", "-quit"};
>         >
>         > One option is to create a new CbcModel and solve it:
>         >   CbcModel model(*solver.getRealSolverPtr());
>         >   CbcMain0(model);
>         >   CbcMain1(3, args, model);
>         > This works fine, but how do I get copy the solution to the
>         OSI object?
>         > (I tried solver->getModelPtr()->setBestSolution and
>         > solver->getModelPtr()->setObjValue, using the values from
>         model, but it
>         > does not work - it does not update the "best possible
>         solution", so even
>         > if solver knows about the optimal solution, it does not know
>         it is
>         > optimal..)
>         
>         
>         CbcModel creates a copy of the Osi interface that you give to
>         it and
>         works on this one. So maybe using the OsiSolverInterface* from
>         model.solver() will work?
>         Or you fix integer variables to the solution in your Osi, try
>         to copy
>         the basis from model.solver() as well, and do a resolve() in
>         your Osi. 
>         
>         > The other option is to work directly on the CbcModel of the
>         OSI object:
>         >   CbcMain0(*solver.getModelPtr());
>         >   CbcMain1(3, args, *solver.getModelPtr());
>         > In works fine in the sense that solver now has the optimal
>         solution (and
>         > knows it is optimal).
>         > The problem with this version is that the CbcMain1 function
>         takes more
>         > than 3 times longer and in the first option - any idea why
>         and how to
>         > fix it?
>         
>         
>         Maybe the way the CbcModel is setup in OsiCbc is different. 
>         
>         Stefan
>         
>         --
>         Stefan Vigerske
>         Humboldt University Berlin, Numerical Mathematics
>         http://www.math.hu-berlin.de/~stefan
>         
>         
>         
>         _______________________________________________
>         Coin-discuss mailing list
>         Coin-discuss at list.coin-or.org
>         http://list.coin-or.org/mailman/listinfo/coin-discuss
>         
> 
> _______________________________________________
> Coin-discuss mailing list
> Coin-discuss at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coin-discuss
-- 
                Matthew Saltzman

Clemson University Math Sciences
mjs AT clemson DOT edu
http://www.math.clemson.edu/~mjs