[Clp] (no subject)

[Michael Hennebry]

On Tue, 28 Jul 2009, Bo Jensen wrote:

> This is something we do see in our customer models, they can be very "close"
> to being infeasible. Which due to tolerances and finite precision can create
> different results from optimizer to optimizer. I think you should get the
> farkas ray ( i.e certificate of infeasibility from farkas lemma ) provided
> by Clp (which I assume you can get), check the quality (i.e dual
> infeasibility and objective of the ray). If it seems ok, then its not some
> numerical issue in Clp and your model can be declared infeasible within
> tolerances (which might be something to look into to improve the model
> stability ).

> On Tue, Jul 28, 2009 at 11:59 AM, Daniel Bienstock <dano at columbia.edu>wrote:

>>   I have a tough LP that the callable version of Clp declares infeasible;
>> however the command-line version eventually (slowly) solves it but
>> complaining a bit at the end about remaining infeasibilities.  Commercial
>> solvers do solve this problem. Question: how can I get the callable version
>> to also solve the problem to optimality.  Thanks.

If you are close to the border, it can be difficult or
impossible to know for sure which side you are one.
For those occasions when it is possible, you might want to read
"Safe bounds in linear and mixed-integer programming"
by Arnold Neumaier and Oleg Shcherbina.
http://www.optimization-online.org/494.ps.gz .

Note that it's possible that all the solvers were correct.
The problem as stated might have been feasible,
but the conversion of "1.9" to a double is not precisely defined.
The callable version of Clp might have solved a
problem significantly different from the others.

-- 
Michael   hennebry at web.cs.ndsu.NoDak.edu
"Pessimist: The glass is half empty.
Optimist:   The glass is half full.
Engineer:   The glass is twice as big as it needs to be."