[Help-glpk] numerical instability

[Robbie Morrison]

Hello Akhil

Just some thoughts.

How badly scaled is your problem?  Do you employ
scaling?  This is off by default under API usage and
needs an explicit call.  You can set the type of
scaling using bitwise or-able constants.  Please see:

  http://en.wikibooks.org/wiki/GLPK/Scaling

Also note the comments on stability here:

  http://en.wikibooks.org/wiki/GLPK/Troubleshooting#Numerical_instability

Perhaps you can get an "advance copy" of 4.46 from
Andrew?

Note the warning is just that.  Its presence does not
preclude an optimal solution.  So does the solver run
to completion?

Finally, my work involves the modeling a network of
processes, described via characteristic curves (that's
the simple explanation anyhow).  I found that whenever
I hit stability issues, my characteristic curves were
buggy, aphysicial, and in need of fixing.  Conversely,
correct curves have never (fortunately) given stability
problems.  But maybe the notion of physicality does not
apply in your case.

HTH, Robbie
---

------------------------------------------------------------
To:          Andrew Makhorin <address@hidden>
Subject:     Re: [Help-glpk] numerical instability
From:        Akhil langer <address@hidden>
Date:        Fri, 8 Jul 2011 11:42:34 -0500
------------------------------------------------------------

> Andrew,
>
> Thanks, Andrew. The NaN problem has been
> resolved. However, I keep getting the following
> warning:
>
> Warning: numerical instability (primal simplex, phase I)
>
> Please note that I have been using INT_MAX as the upper
> bound for some columns. Could that be the cause of the
> instability? I can change these columns also to be free
> but that would require some changes in the code.  Just
> wanted to check with you if changing the code will
> remove the instability?
>
> Thanks,
> Akhil

---
Robbie Morrison
PhD student -- policy-oriented energy system simulation
Technical University of Berlin (TU-Berlin), Germany
University email (redirected) : address@hidden
Webmail (preferred)           : address@hidden
[from Webmail client]