Re: [Help-glpk] control parameters and round-off error

[Michael Hennebry]

On Tue, 15 Jan 2008, Ali Baharev wrote:

> I did try the exact solver but it is very-very slow, as i expected.
>
> My LP problems are generated by successive linearization of a
> nonlinear problem, and i need to automate the solution process. So my
> problem is not only for this particular LP problem, i need an error
> estimate on the objective function value for each solved LP problem.

Try using the exact solver only at the end,
dropping all the constraints with basic slacks.
If any constraint are violated, add them back in and resolve.

> I checked the KKT conditions and all the solution and feasibility
> (both for primal and dual) qualities turned out to be high without
> exception during the entire solution process.
>
> So my question boils down to the following: how accurate is "quite
> accurate"? Which parameter(s) determines the accuracy of the objective
> function value?

The was a paper mentioned earlier in this list on getting reliable
information about linear programs from floating point computation.
I can't seem to find my copy at the moment..
One can get bounds on errors even without directed rounding
if one knows how ones computer does arithmetic.
Some compilers try to be helpful in ways that tend to defeat such efforts.
To work around, you moght have to do your postprocessing in long double.

-- 
Michael   address@hidden
"Those parts of the system that you can hit with a hammer (not advised)
are called Hardware;  those program instructions that you can only
curse at are called Software."