Hi! On Mon, 2009-05-25 at 16:30 +0200, Mikael Zayenz Lagerkvist wrote: > Hi, > > On Wed, May 20, 2009 at 8:36 PM, Denys Duchier > <[email protected]> wrote: > > I could be wrong about this, but I just had a quick glance at the > > implementation and its seems to allocate datastructures of a size > > proportial to the width of the domain. So if you have small values and > > very large values, the width is going to be very large. > > > > If I am right, then you might be better off encoding your N values into > > the interval [0,N-1] and then possibly using an element constraint to > > decode them where necessary. > > > > Then again, I could be completely off base :-/ > > Actually, you are completely right, and it also says so in the > documentation [1]: > Warning: > If the domains for the $x_i$ are not dense and have similar > bounds, lots of memory will be wasted (memory consumption > is in $ > O\left(|x|\cdot|\min_i(\underline{x_i})-\max_i(\overline{x_i})|\right)$ > for the basic algorithm (epk = EPK_MEMORY) and additionally > $ > O\left(|x|^2\cdot|\min_i(\underline{x_i})|\max_i(\overline{x_i})|\right)$ > for the incremental algorithm (epk = EPK_SPEED). > > The memory-consumption in the size of the domain is inevitable for > reasonably efficient extensional propagators. The translation into a > dense domain from a sparse domain needs to happen at one point or > another anyway for efficient data-structure management, and I choose > not to do it automagically.
Sorry, my bad, didn't noticed that on the documentation... Now I understand why it gets really slow, as the domain gets really big. So, there is no easy way of preventing the translation of this sparse domain to a dense domain, even if it is just to make some tests and see the behavior of Gecode under those circumstances? Thank you, Pedro Salgueiro > > Cheers, > Mikael > > [1] > https://www.gecode.org/doc-latest/reference/group__TaskModelIntExt.html#gf885611633b7b2bfb4d9512071dee9a3 > _______________________________________________ Gecode users mailing list [email protected] https://www.gecode.org/mailman/listinfo/gecode-users
