Re: [Help-glpk] Working with larger numbers

[Gabor Retvari]

On Monday 28 July 2008, Markus Pilz wrote:

> we are working on a tool that utilizes glpk to solve a set of maximum
> and minimum cost flows. So far, our approach looks promising. (This
> implies a big thank-you to glpk team.)
>
> Currently, we have some difficulties with larger decision variables. For
> example, if  the variables of a maximisation problem (max flow) are
> limited by values above 1e9, the solution tends to be inexact.
>
> We still can use the solution as a basis for further processing but
> maybe we lack some lp (or glpk) basics to obtain exact values in a wider
> range.

Wouldn't it be plausible to run GLPK over precise arithmetics? You can do 
this by calling lpx_exact instead of lpx_simplex. Network flow problems tend 
to be seriously overconstrained and thus highly degenerate, calling for 
numerical instability. An exact solver would solve this issue, for the prize 
of being somewhat slower. I mean, *really* slower.

Best regards, 
Gabor

----

http://qosip.tmit.bme.hu/~retvari