[scilab-Users] Re: Linear programming with scilab 5

[Alberto Marin]

>
> The problem may be related to the fact that there are more equality
> constraints
> (i.e. 96) than unknowns (i.e. 93). But it seems that only 81 of the 96 rows
> of
> the equality matrix are numerically linearly independent :
>
> -->rank(eqMat)
>  ans  =
>    81.
>
> There are no such rank defficient tests neither for karmarkar, nor for
> linpro.
> Actually, there were no unit test at all for linpro until June 2011... But
> I
> expect that linpro is more robust in general, so I am not surprised by the
> result. But karmarkar does not seem to work here, and returns with the
> exitflag
> equals to -1, meaning that the initial guess x0 could not be found. More
> testing
> should be done to create a simpler rank defficient linear problem with, say
> 4-5
> unknowns and see what happens. Do you have such a simple example and its
> exact
> solution ?
>
This kind of rank deficiency is very frequent in this kind of problem
(biochemical networks) and has never been a problem using the simplex
method. Since the only thing I know about the karmarkar algorithm is that it
is an interior point method, I have no clue about what the error message
about rank deficiency means. I have tried eliminating the redundancies but
the reesult is the same so the error may refer to a completely different
issue.

>
> When you say that you have problems when you use linpro through the Java
> interface, what do you mean exactly ? Can you provide a script ?
>

If I  put this problem in an .sce file and execute it, I get the result you
mention. No problem so far. If I call linpro from the java interface with
the Siclab.exec() method it also works fine as long as I use a 32 bits
machine. But calling linpro through  the exec() method in any linux 64 bits
machine crashes the java virtual machine. This same problem happened
whenever I called the javasci engine with scilab 4.2 but in 5.3 it is
restricted to the linpro function, any other operation I've tried so far
(including karmarkar) seem to work fine.

Alberto

>
> Regards,
>
> Michaël Baudin
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.scilab.org/pipermail/users/attachments/20110720/c2f275ed/attachment.htm>