From kde-core-devel  Fri Nov 19 13:01:06 1999
From: Leon Widdershoven <l.widdershoven () fz-juelich ! de>
Date: Fri, 19 Nov 1999 13:01:06 +0000
To: kde-core-devel
Subject: Re: Need help for KSpread Maths
X-MARC-Message: https://marc.info/?l=kde-core-devel&m=94301652329983

Don Sanders wrote:

[SNIP] 

> "Hooke-Jeeves is slow, It is "wasteful". It is also robust
> as all get-out. I prefer my program to run for 5 minutes
> and converge to the minimum 100% of the time, than for it to
> run for 30 seconds and converge to the minimum 75% of the time.
> But that's just _my opinion_. Yours may differ."
> 
> (I doubt it will take five minutes for normal problems, this was written in
> 1994 and assumes the function f to have a vector values domain, the hooke.c
> example runs in a fraction of a second here).
> 

Is it in any way possible to predict the time necessary for computations?
Because I would say that 5 minutes solving time is not a problem, but 
in that case a progress bar would be nice.
If it is not possible to predict the number of iterations necessary (I 
don't know if the size of the steps to convergence can be a measure)
then there could be at least some signal going out of the routine saying
"I already did 5 million steps!" or something like that. Which can be
put in the status bar to let the user see its computing.

BTW: I do *not* mean to say that I do not agree with your opinion 
you'd rather wait 5 minutes for a solution that 30 seconds for 
perhaps nothing. Computation time, in general, is neglegible compared
to the time users need to find & type the formula.
And nothing is as frustrating as knowing there is a solution but the 
damned program can't find it:)  
  

[SNIP]

> > > (Maybe getting a bit offtopic for kde-core?)
> >
> > Maybe, but at least it's geeky and may be useful for a KDE app ;-)

It's also interesting; it shows the broad field of knowledge necessary
(and available!) for creating an office suit/desktop environment. 

> > A quick search in www.microsoft.com for "solver" shows some intriguing
> > stuff!
> >
> > For example, it seems the solver in Office 2000 is modular and
> > replaceable (you can replace it with a VB module, just what all numerics
> > expert recommend[1] ;-)
> >
> > Anyway, MS is even kind enough to say what algorithm they use!

I may not always like MS (neither did the judge: 
http://usvms.gpo.gov/findfact.html) but I think it's very good they 
open up at least this information. 


> > http://support.microsoft.com/support/kb/articles/Q82/8/90.ASP
> >
> >
> > --------------
> >    SUMMARY
> >
> >    Microsoft Excel Solver uses the Generalized Reduced Gradient (GRG2)
> > Algorithm for optimizing nonlinear problems. This algorithm was developed by Leon Lasdon, of the
> > University of Texas at Austin, and Allan Waren, of Cleveland State University.
> >
> >    Linear and integer problems use the simplex method with bounds on the
> > variables and the branch and bound method, implemented by John Watson and
> > Dan Fylstra, of Frontline Systems, Inc.
> > --------------
> >
> > I must confess I had never heard of this algorithm before.
> >
> >
> > [1] I can imagine the screaming: "Why can't I do it in FORTRAN!!!!"

Well, the NAG library is very, very good (and expensive, unfortunately). 
And also available in C:) 

And well, why can't I do it in fortran? Because I don't speak fortran too well:(