[Bug-glpk] [Fwd: Re: Simplex termination 32 vs. 64 bit]

[Andrew Makhorin]

-------- Forwarded Message --------
From: Horst Schirmeier <address@hidden>
To: Andrew Makhorin <address@hidden>
Cc: address@hidden
Subject: Re: Simplex termination 32 vs. 64 bit
Date: Mon, 9 Apr 2012 23:36:46 +0200

On Tue, 10 Apr 2012, Andrew Makhorin wrote:
> > [...]
> > The input data is attached to this email.  Is it somewhat malformed, or
> > did I indeed step on a GLPK bug?
> 
> No, this is not a bug.
> 
> The point is that glpk uses floating-point calculations that are
> sensitive to many platform-dependent factors (compiler options, etc.).
> That means that you may expect to obtain identical results only on
> identical platforms. In particular, you may note that 32-bit version
> performed 851 simplex iterations to solve the lp relaxation while 64-bit
> one performed 844 iterations. 
> 
> Your mip instance is generally hard for the branch-and-bound solver
> (used by default), and it was solved quickly with the 64-bit version for
> a happy chance. However, I think that the branch-and-cut solver with the
> mir cuts enabled (--mir) is able to solve your mip for a short time
> independently on the platform used:
> [...]
> 
> The Gomory cuts (--gomory) also seem to be efficient:
> [...]

Thanks for clearing this up.  Now I need to read up on MIR / Gomory cuts
to decide whether I can expect these to be the right choice for my whole
problem class; these ILPs are part of an automated process.

Thanks!

Kind regards,
 Horst Schirmeier