Re: [Help-glpk] [Fwd: Scaling: Which? and Suppressing output]

[GLENN RHOADS]

On Thu, 24 Feb 2011 21:59:23 +0100
 "glpk xypron" <address@hidden> wrote:
> Hello Glenn,
> 
> your e-mail did not reach the GLPK help list directly, because
> you did not register. See
> http://lists.gnu.org/mailman/listinfo/help-glpk

I did register but I think I know the problem.  Rutgers University has
a central mail server that handles all email on all Rutgers systems.
 When I email something, it first goes through this server which
changes the outgoing address from my real address "@cs.rutgers.edu" to
the mail server's address "@dragon.rutgers.edu"  I had the same problem
when subscribing to another email list.


> > A: min|aij| =  4.000e-01  max|aij| =  2.800e+00  ratio =  7.000e+00
> > Problem data seem to be well scaled
> The message is written in file glpk-4.45/src/glpscl.c
> You could replace the following lines:
> 
>       if (min_aij >= 0.10 && max_aij <= 10.0)
>       {  xprintf("Problem data seem to be well scaled\n");
>          /* skip scaling, if required */
>          if (flags & GLP_SF_SKIP) goto done;
>       }
> 
> Best regards
> 
> Xypron

Thanks.  That solves the excessive output problem.
Any suggestions on which scaling method might be preferred for my
application?  I think what is happening is that some of the
coefficients are for all practical purposes effectively 0.  These
"effective zeroes" are being represented in the matrix as some very
small number, e.g. 7.2134e-32, whereas the other entries are of a
common everyday size like 8.2, or 0.34.  Which scaling method would be
preferred for handling this situation?

-- Glenn





> > -------- Forwarded Message --------
> > Subject: Scaling: Which? and Suppressing output
> > Date: Thu, 24 Feb 2011 11:38:50 -0500
> > 
> > Hi,
> > 
> > I have a C++ application where calling glpk_simplex() sometimes
> > generates a "numerical instability" warning message.  To fix this
> > problem, I tried to scale the data by first calling
> glpk_scale_prob().
> >  However, calling glpk_scale_prob() generally prints out a message
> like
> > 
> > 
> > A: min|aij| =  4.000e-01  max|aij| =  2.800e+00  ratio =  7.000e+00
> > Problem data seem to be well scaled
> > 
> > This is especially annoying because in this application, I am
> solving
> > *millions* or even billions of very small dense LPs (these LPs come
> > from matrix games).  How can I suppress these messages and still
> get
> > any warning or error messages that I need to see?  Right now, any
> > warning messages are getting lost in a never-ending stream of
> useless
> > messages.
> > 
> > 
> > Also, there are a couple of different scaling options available
> such as
> > "geometric mean" scaling and "equilibration" scaling.  I don't know
> > what these are.  Are there any useful rules of thumb as to which
> > scaling method to use?
> > 
> > -- Glenn C. Rhoads