RE: [Help-glpk] FW: trying to stop solver

[Michael Hennebry]

On Tue, 4 May 2004, Robert Horvath wrote:

> I agree that accepting a non feasible solution for phase 2 is not a very
> good idea. But if you consider the problem that is to be solved it has 56188
> auxiliary variables and 157793 structural ones, so it has about 213981
> variables total. GLPK calculates the infeasibilities by adding all the
> infeasibilities found at each variable. Two things follow:
>
> 1.    IF the infeasibilities are distributed evenly then .732/213981=3.42e-6
> per one variable, which is acceptable I think.(I would really appreciate
> some remarks on that from Brady Hunsaker)
>
> 2.    If the infeasibilities are not distributed evenly, for example only a
> few variables are responsible for the infeasibilities found, then there is a
> problem, and the problem needs to be rearranged, or it just takes more time
> to calculate phase II.

The most straightforward way to determine which is the case is
to fetch the infeasible solution and do the comparisons.

If you find that some variables are unacceptably out of bounds,
that would not necessarily be the whole story.
It might be because that GLPK's idea of the least
infeasible solution is not the same as yours.
It might be that increasing GLPK's idea of the total infeasibility by
10% would allow reducing the maximum infeasiblity of a variable by 99%.
You might be able to work around this by rescaling the bad variables.

In any case, it seems to me that the problem is almost certainly
overconstrained, i.e. all feasible solutions are primal degenerate.
A solution with all basic variables in the strict interiors of
their ranges should produce a total infeasibility of zero even
after allowing for the vagaries of floating point computation.

-- 
Mike   address@hidden
"Nothing says it like words if you know how to use them."
                        --  the Professional Organization of English Majors