Hi Gilles, That sounds really interesting, very nice!
I would never go for 1): This would require you to change things inside the Gecode kernel. This will most likely be a very frustrating experience. Even I would not dare doing that unless I had half year and under the assumption that I know Gecode as well as I do already. 2) sounds more promising but still very ambitious. If I were you, I would have a look at how activity information is recorded (the files are in gecode/kernel). The idea is as follows: a propagator is created which does not really do any propagation but it just records information about how variables change. To make it efficient this propagator is given a very high cost so that it will run only after all "real" propagators have reached their fixpoint. And as you suggest it uses advisors to do that. My gut feeling is that you can start very quickly from what is already there for activity recording. It is not that obvious how to share datastructures then. It is possible in principle, though. But I'd really leave that to later. Cheers Christian -- Christian Schulte, www.gecode.org/~schulte Professor of Computer Science, KTH, cschu...@kth.se Expert Researcher, SICS, cschu...@sics.se -----Original Message----- From: users-boun...@gecode.org [mailto:users-boun...@gecode.org] On Behalf Of Gilles Pesant Sent: Saturday, February 13, 2016 02:23 AM To: users@gecode.org Subject: [gecode-users] counting-based search in Gecode I would like to integrate counting-based branching heuristics in Gecode. Like AFC we need to collect information from each constraint in which a given variable appears. But our information is more involved than a simple updated failure count --- we have dedicated algorithms computing "solution densities" for variable-value pairs. We saw that we can define a Brancher to make an integrated choice of both variable and value. But to host our solution counting code we see two options: 1) In objects of a class Counter which, like Propagator, would inherit from Actor and subscribe to variables. They would mainly implement a getSolutionDensity(var,val) method and possibly react to domain events by updating data structures (via Advisors?). 2) By inheriting directly from Propagator, say in a class CountablePropagator, which may allow us to share data structures used by the propagator: for example for the regular constraint the layered graph is augmented with counters of incoming and outgoing paths at nodes. What's your opinion? Thanks, Gilles _______________________________________________ Gecode users mailing list users@gecode.org https://www.gecode.org/mailman/listinfo/gecode-users _______________________________________________ Gecode users mailing list users@gecode.org https://www.gecode.org/mailman/listinfo/gecode-users
