Re: [Help-glpk] Suggestions to solve the master problem faster in a Branch-and-Price model.

[Andrew Makhorin]

> Now my question is: should I solve the model from scratch in the case
> I have to remove a lot of variables? 

Generally, not.

> Or is there a parameter configuration I should use in my specific
> case?
> 

Glp_simplex always starts the search from the current basis which is
provided in glp_prob by means of the statuses of rows and columns. This
allows to continue the search (re-optimize) efficiently if you obtain an
optimal basis, change the lp, and then call glp_simplex once again.
However, in this case you need to make sure that the row/column statuses
define a "correct" basis--the number of basic auxiliary/structural
variables should be equal to the number of rows, and the basis matrix
consisting of columns of basic variables, including unity ones for
auxiliary variables, should be non-singular. For example, removing an
active (more precisely, binding) constraint, i.e. a non-basic row, makes
the basis incorrect. To avoid this you may either remove the row
"logically" by making it free (unbounded) rather to remove it
"physically", or make the row basic (non-binding) before removing it by
changing the row/column statuses in an appropriate way.