[Gecode] design question

[Gustavo Gutierrez]

On Thu, Mar 03, 2005 at 04:07:41PM +0100, Guido Tack wrote:
> On Thursday 03 March 2005 15:51, Christian Schulte wrote:
> > I am sorry but from what you say I cannot come up with any suggestion but
> > to use a debugger...
> 
> >From what I see here, you may have hit one of the main design decisions in 
> Gecode: When a space fails, it can be in an arbitrarily screwed up state. If 
> I understand you correctly, you perform some propagation, and if that fails, 
> you go on with some other propagation in the same space. If failing the first 
> propagation means that one variable is actually failed, you may not use the 
> space any longer. I think this can easily lead to segmentation faults, can't 
> it, Christian?
>

I will write a part of the corresponding propagator as a
pseudo-algorithm:

Suppose Bc is a propagator already implemented.
I am interested on writing K3b which has the VarArray x.
K3b<Var,Bc>::propagate(Space* home) {
	VarArray<Var> clone(x);
	for (int i = n; i--; ){
		//new guess upper bound
		double new_upper = clone[i].lower() + 2.0;
		clone[i].upper(home,new_upper);
		Bc narr(home,clone);
		if(narr::propagate(home) == ES_FAILED)
			x[i].lower(home,new_upper);
	}
}

>From Guido's answer i infer the problem occurs because both propagators
share the same space. Another way to accomplish the same behavior is by creating a new Space and to put there the propagator. This will prevent the parent
space to become failed or "...in an arbitrarily screwed up state".

I will try to write the function guessing some space primitives because
i have not seen much about spaces.

K3b<Var,Bc>::propagate(Space* home) {
        VarArray<Var> clone(x);
	for (int i = n; i--; ){
		//compute a guess upper bound
		double new_upper = clone[i].lower() + 2.0;
		//Create a new space
		Space test();
		Bc narr(test,clone);
		clone[i].upper(test,new_upper);
		if(narr::propagate(test)==ES_FAILED)
			x[i].lower(home,new_upper);
	}
}

In this new version, the space 'test' is created to test the propagator
(narr) in another independent space.  Note that the processes performed on test influence subsequent propagation on home. I don't know if this can work
but i don't feel so good creating additional spaces in propagation steps.
										All the best,
										Gustavo Gutierrez