> In a specialized single-commodity network flow solver, > the arithmetic is normally exact even if done in floating point.
> That is usually not true with multi-commodity flows. > If glpk floating point gives you the correct basis, > glpk exact should give you the exact answer in fairly short order. > So long as the glpk floating point basis is near-optimal and feasible, > a subsequent call to glok exact should be fairly quick. I think that Markus obtains a solution, where some flows are non-integral while he expects them to be integral due to network structure and integral arc capacities (I mean max flow and min cost flow problems). On the other hand, it is unclear to me why this could happen, because the constraint matrix is an indicent matrix of the network, so the basis matrix being an incident matrix of the spanning tree is triangular and therefore its LU-factorization is trivial and should be exact even in floating-point arithmetic. I suspect that there are some additional constraints in the model. _______________________________________________ Help-glpk mailing list [email protected] http://lists.gnu.org/mailman/listinfo/help-glpk
