Thx for the pointers! I'm afraid my bg simulator won't interest you that much. I'm trying to learn a bit about neural networks and implementing/training the standard TD gammon neural net seems like an interesting way to do that. And maybe it'll help improve my unimpressive bg game at the same time. :)
On Jan 1, 2010, at 7:09 AM, Øystein Johansen wrote: > On Thu, Dec 31, 2009 at 10:33 PM, Mark Higgins <[email protected]> wrote: > I'm coding up my own little backgammon simulator in python and want to > benchmark it against gnubg. I was hoping there's a gnubg python module I can > install and import that would expose the underlying engine so I can pass my > generate boards and dice rolls into it, and have it return the set of > possible moves ordered by what it thinks expected point scores are - that > sort of thing. > > Does such a module exist? I googled around a bit but didn't find anything. > > Hi! > > I would strongly suggest that you use the benchmark databases from Josephs > work. The benchmark databases can be found here: > ftp://ftp.demon.nl/pub/games/gnubg/nn-training/benchmarks > > (Plain ascii files) > > The classify positions like gnubg main code, and we therefore also use the > same benchmark (We mainly focus on the contact class). I believe this is the > ideal way to check the strength of a bot, since it is quite fast and accurate > enough. If you like this idea, look at Josephs code. I can also provide you > come C code (depending on glib) which runs through a benchmark. > > The other way of finding strength of a bot is to play long sessions against > each other. Was that more of what you where thinking? > > -Øystein > > PS: Can you tell us more about your project? What's the AI technology? Neural > net? What's the state? Can it do deeper searches? Cube evaluations etc? We're > curious here.
_______________________________________________ Bug-gnubg mailing list [email protected] http://lists.gnu.org/mailman/listinfo/bug-gnubg
