[Help-glpk] Re: About int lpx_simplex(LPX *lp) method

[Andrew Makhorin]

>My question is why GLPK is not perfect? Is it just for large size problem?
>or for rounding errors?
>or...?
>
>How much robust and reliable GLPK is?

Glpk simplex solver is not perfect in the sense that currently it does
not include some important features from the standard repertoire of any
state-of-the-art simplex-based lp solver. For example, it does not
include a feature, like epsilon perturbation, against cycling (however,
due to steepest edge pricing and two-pass ratio test in both primal and
dual versions the cycling happens seldomly), or it is not able to
repair the singular basis, because the corresponding procedure is not
implemented yet. Some termination tests are also needs to be improved,
etc. The user should not expect that glpk is able to solve any lp
instance, that is why the documentation states that the glpk lp solver
is not perfect.

For not very large (up to 100,000 constraints) and not very hard lp
instances the glpk simplex solver is sufficiently robust. You can look
at benchmarks for netlib lp instances included in the distribution (see
the file bench.txt). In order to make sure that results reported by the
solver are correct and to estimate the solution accuracy you can check
the KKT optimality conditions with the api routine lpx_check_kkt.

Andrew Makhorin