On Wed, Oct 28, 2009 at 04:47:39PM +0100, François Galea wrote:
However, some results I found are different from those which are in the corresponding '.ll' solution files. For instance, in 'boulet.pip', I get the solution minimum {0 ; 0 ; 0}, since all constraints have a nonnegative constant term. PIPlib finds a solution with negative elements : "if (p>=-5) then {-3*p-15 ; -p-5 ; p+5}."
This may be valid if variables and parameters are not supposed nonnegative, but my current implementation of the solver does not support negative variables and parameters.
That's what the options "Urs_parms" and "Urs_unknowns" in boulet.pip mean. That is, both the parameters and the unknowns have unrestricted sign. So, piplib's answer is correct (see the piplib manual, where this example originates from).
A bit of history. The original version of pip would assume that all unknowns and parameters are non-negative, but allowed the user to specify a "big parameter" that is treated in a special way by piplib. By reformulating the problem a little bit and by applying yet another trick to deal with urs parameters, the user could trick piplib into handling urs unknowns and parameters. A couple of years ago, I made some changes to piplib that allow the user to simply _indicate_ that the unknowns and parameters have urs and then piplib will perform all these tricks internally. The default was not changed for reasons of backward compatibility.
Now, I actually think it is wrong to expose this implementation detail to the user, so, in isl, I make no assumption about the sign of the unknowns or the parameters. In the demo application isl_pip, I then add the appropriate ">= 0" constraints if the user did _not_ specify the Urs_parms or Urs_unknowns options.
skimo