[Coin-bcpdiscuss] test_ful() question and BCP_DBL_MAX

[Laszlo Ladanyi]

Hello all, 

I have changed the code (stable and trunk) so that bound violations are not
tested and BCP just blindly trusts the LP solver. Francois is probably right
that that's the best we can do... 

I have also changed the code so that now there is a constant BCP_DBL_MAX
defined in BCP_math.hpp (right now as 1e100) that is used for the purposes of
infinity. In various places of the code different hard-coded values were used,
and I think this is a much better solution. Still, it's possible that
someone's code was dependent on those hard-coded values, so this is a heads up
call :-).

I also have a request: I'd like to make a new point release, since the 1.0.0
tarball does not compile properly on cygwin. Could someone grab the stable
version from svn and try to compile it on cygwin/msys and whatever is
available on windows? BIG thanks and lots of brownie points!

Thanks,
--Laci

On Fri, 19 Jan 2007 fmargot at andrew.cmu.edu wrote:

> 
> The problem is that BCP requires precision for the solution of the original 
> problem, whereas most solvers give an option for the precision on the 
> solution of the scaled and preprocessed problem. It is possible to construct
> problems where the returned solution for the original problem violates 
> a constraint by an arbitrary fixed amount.
> 
> While there is a relation between the two kinds of precision, it is not
> easy to determine how to set them so that they match.
> 
> Francois
> 
> On Fri, 19 Jan 2007, Matteo Salani wrote:
> 
> > Interesting thread,
> >
> > Isn't it enough for BCP to request a less strict precision than the LP 
> > solver?
> > So that if you need more precision in the solution BCP proposes you should 
> > increase also the LP one?
> >
> > M
> >
> > fmargot at andrew.cmu.edu wrote:
> >> The problem with checking bounds on continuous variables is that it
> >> might lead to a BCP failure, although neither BCP nor the LP solver
> >> is really at fault: BCP wants a certain precision on the solution, but 
> >> there
> >> is no way to request that from the solver.
> >
> >
> >
> _______________________________________________
> Coin-bcpdiscuss mailing list
> Coin-bcpdiscuss at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coin-bcpdiscuss
>