When @lgottwald wrote the MIP solver, the analytic centre LP was never presolved. For the Tulipa instances, the cost of the the analytic centre calculation without presolve is huge, particularly with IPX. Presolve reduces the problem size by factor of about 2.5 and with (also) using HiPO the instances are now solved vastly faster.
In general, though, does presolve impair the quality of the interior solution that's obtained? Clearly postsolve starts from the analytic centre of the presolved problem - even if that's empty, in the case of egout (see #2500), but do the postsolve operations tend to introduce variables and constraints on/near their bounds?
When @lgottwald wrote the MIP solver, the analytic centre LP was never presolved. For the Tulipa instances, the cost of the the analytic centre calculation without presolve is huge, particularly with IPX. Presolve reduces the problem size by factor of about 2.5 and with (also) using HiPO the instances are now solved vastly faster.
In general, though, does presolve impair the quality of the interior solution that's obtained? Clearly postsolve starts from the analytic centre of the presolved problem - even if that's empty, in the case of
egout(see #2500), but do the postsolve operations tend to introduce variables and constraints on/near their bounds?