> > On Thu, 2013-03-07 at 18:25 -0600, David T. Price wrote: > > > I will look for it soon. The same bug occurs in check_kkt as well. Did > > > you fix it there? > > > > > I fixed the similar bug in lpx_check_kkt. Please see a patched version > of glpapi10.c attached. You need to replace that file and rebuild the > package as usual. (The patch is for v.4.48.) >
The bug in lpx_check_kkt was fixed incorrectly. Please DO NOT USE the patch. > > I don't understand your argument. Consider the problem: > > > > Maximize 5 * x1 > > > > Bounds > > 1 <= x1 <= x2. > > > > The trivial basis will put x1 on the boundary nearer 0, namely x1 = 1. > > This is clearly primal feasible. However, the optimal solution of the > > problem is x1 = 2. So the starting solution can't be dual feasible. > > However, glp_warm_up says that the starting solution is "dual feasible" > > and hence the status reported is "optimal". That is wrong. In this case the complementary slackness condition is violated, neither primal nor dual feasibility conditions checked by lpx_check_kkt. This condition is usually not checked for basic solutions, where it is satisfied automatically (unlike interior-point solutions). However, this is so until double-bounded variables appear. _______________________________________________ Bug-glpk mailing list [email protected] https://lists.gnu.org/mailman/listinfo/bug-glpk
