Thanks for the link, from now on I won't expect the same behavior for Windows and Linux when solving a linear model. But what about the difference in the number of digits?
2009/6/23 Andrew Makhorin <[email protected]> > > I am getting back to this subject because I am facing a small > > problem using GLPK both on Windows and Linux. > > As I said in my previous e-mail, I managed to compile the library > > to use GLPK in java for Linux, and it works fine. However, I get a > > difference when running an instance under Windows on the one hand and > > on Linux on the other hand: it is the same optimal solution (same > > primal values for variables), but the basis is different, as well as > > the dual values of constraints and variables. It is a problem as I > > need to use these dual values in a Branch-and-Price algorithm to get > > new columns. > > > During my investigation, I saw a difference that might be linked to > > this behaviour: when printing the solution into a file (using > > glp_print_sol), the Karush-Kuhn-Tucker optimality conditions (at the > > end) don #39;t provide exactly the same display. On the one hand > > (Linux), there are 2 digits for the powers of 10 (ex: 0.00e+00), and > > on the other hand (Windows), there are 3 (ex: 0.00e+000). > > Could the problem I am facing be linked to a rounding problem? If > > so, how can I make the result identical for Windows and Linux? > > If you need the model for which this problem occurs, feel free to ask. > > Please see: > http://list.coin-or.org/pipermail/coin-discuss/2009-March/003548.html > >
_______________________________________________ Help-glpk mailing list [email protected] http://lists.gnu.org/mailman/listinfo/help-glpk
