Re: [Help-glpk] numerical instability

[Robbie Morrison]

Hi Akhil

------------------------------------------------------------
To:          Robbie Morrison <address@hidden>
Subject:     Re: [Help-glpk] numerical instability
Message-ID:
<address@hidden>
From:        Akhil langer <address@hidden>
Date:        Mon, 11 Jul 2011 20:38:28 -0500
------------------------------------------------------------

> @Robbie: The problem was very badly scaled (according
> to the definition given in the wiki).

Those definitions come straight from the documentation
in the GLPK source code.

> I used glp_scale_prob() but the warnings persist. In
> the program, the model gets modified with addition and
> deletion of rows between optimizations. Can I expect
> that GLPK will take care of rescaling required because
> of the changes in the model.

Andrew gave this answer:

|    ------------------------------------------------------------
|    To:          Akhil langer <address@hidden>
|    Subject:     Re: [Help-glpk] numerical instability
|    Message-ID: <address@hidden>
|    From:        Andrew Makhorin <address@hidden>
|    Date:        Sun, 10 Jul 2011 20:34:45 +0400
|    ------------------------------------------------------------
|
|    Numerical instability may happen because of badly
|    scaled problems, so try to scale your initial lp
|    instance with glp_scale_prob (do not use glp_scale_prob
|    before reoptimization).

I do not really understand what Andrew means.
Perhaps that rescaling necessarily throws away
valuable solution information.  Ask him if you need
to know.

> Regarding your observation of buggy curves when
> instability is reported. I am also facing a similar
> issue. My program runs in parallel environment. I am
> doing stochastic optimization. An executable with same
> input every time sometimes converges to optimal values
> without any issues and at other times leads to bound
> violations in in my algorithm. Since, it is running in
> parallel environment glpk solves different lp's in
> different runs but they should not lead to violations.

What kind of parallel environment?  Strictly speaking
GLPK is not thread-safe, although several users have
written threaded programs.  Here is the official line:

  
http://en.wikibooks.org/wiki/GLPK/Using_the_GLPK_callable_library#Thread_safety

Are you using multi-threading?  And if so, which library
for support?  'Pthreads' perhaps?

HTH, Robbie
---
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]