[Coin-discuss] modeling with GNU MathProg

[John Fuller]

It seems like
.a model expresses a problem that can be solved by a solver
.the interface to a solver is something different from a model
.a solver interface could use/collaborate with a model
"extracting" or adapting information from it.


On Tuesday, May 20, 2003, at 10:17 AM, Antonio Frangioni wrote:

>> A model class seems reasonable to me.  Are you imagining that an
>> OsiSolverInterface object would be contained in the model class, or
>> that the model class would inherit from the OsiSolverInterface class?
>>
>> Also, I'd really like to hear any other opinions out there.
>
> I think that a model class should be entirely different from OSI, for 
> several
> reasons:
>
> 1) it has nothing to do with algorithms, it is a description of an 
> instance
>
> 2) we could use some of the nice software that already exists, like 
> FLOPC++
>
> We have already done something like that, although rudimentary. One of 
> the things
> that a model class may help in doing is handling changes of the model 
> for those
> algorithms that use complex solution methods where different parts of 
> the same
> instance are handled by different solvers (think of a B&C with a 
> Lagrangian
> relaxation inside). What we did is a system where the model is 
> responsible to keep
> track of its changes for the various solvers that "register" with it 
> as being
> interested in knowing them. So the model object is not only a 
> repository of the
> data, but also of the dynamic changes of the data.
>
> 												Best
>
> 												Antonio
>
> _______________________________________________
> Coin-discuss mailing list
> Coin-discuss at www-124.ibm.com
> http://www-124.ibm.com/developerworks/oss/mailman/listinfo/coin-discuss
>