[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Bug-gnubg] Fwd: Neural Net program for For Traditional Iranian/Turk
Re: [Bug-gnubg] Fwd: Neural Net program for For Traditional Iranian/Turkish Backgammon - from Rick Janowski
Wed, 29 Jun 2011 08:17:06 +1200
A far more interesting and difficult task would be to go for one of
the other variants In my family they used to play machboosa
(surrounded), which is the same as the Greek Plakoto, which I find
strategically orders of magnitude harder. I can hardly tell if I make
sensible moves or not, as this is a game where you have to slow
yourself down usually not just a race to the line that is backgammon.
I even have doubts that a neural net can capture this complexity,
since most consequences are far ahead in the future - it is like
playing a backgame all the time, and we know how poor computer players
are at that, GNUBG especially.
2011/6/29 Øystein Schønning-Johansen <address@hidden>:
> This is indeed interesting, however I think it is a lot of work....
> The problem is that you have to recode a lot of things. First of all the the
> move generator. Then the bearoff evaluator must be reconstructed. I guess
> you can use the same system as in one side bearoff databases, (where you
> keep the a table of the #-rolls-to-get-all-off distribution for every
> position), but it has to take the bearoff rules into account. Makes the
> distributions different.
> This makes a difference also for race positions as well, so the race network
> must be retrained as well....
> I guess you can make a simple TD(lambda) training of a neural net when all
> that's in place.
> So, challenging that it would be cool to try, however I think I rather spend
> my time on other projects...
> Hmmmm... it isn't that much changes..... maybe I should at least take a
> deeper look.
> Bug-gnubg mailing list