Re: [Help-glpk] Problem with "Numerical problems with basis matrix"

[Andrew Makhorin]

>Hi, I have a MIP problem that I try to use with glpk, but glpk "gives up"
>with the message "Numerical problems with basis matrix" after a while.
>Is there any way to "improve" the problem to get rid of the problem?

Unfortunately, the basis recovery procedure that needs to be performed
in that case is not implemented yet.

You can try one of the followind "remedies":

a) if your problem is not scaled, scale it before solving using the
   routine lpx_scale;

b) if your problem is scaled, do not scale it;

c) relax the pivoting tolerance as follows:

      lpx_set_real_parm(lp, LPX_K_TOLPIV, 1e-8);

   and if this will not help then 1e-7 and then 1e-6;

d) relax the primal feasibility tolerance as follows:

      lpx_set_real_parm(lp, LPX_K_TOLBND, 1e-5);

c) decrease the factorization frequency; in order to do that, replace
   the line 97 in the file 'glpspx1.c':

      if (lp->inv == NULL) lp->inv = inv_create(lp->m, 100);

   by the following:

      if (lp->inv == NULL) lp->inv = inv_create(lp->m, 20);

If you are using glpsol, not glpk api, you need to insert these changes
after the comment line 367 in the file 'glpsol.c'.

If it is possible, could you please write your problem in mps format
using the routine lpx_write_mps, gzip the mps file, and send it to
<address@hidden>, or at least inform me how many rows, columns, and
non-zero constraint coefficients your problem has.