[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Bug-gnubg] GnuBG faring poorly in backgames

From: Joseph Heled
Subject: Re: [Bug-gnubg] GnuBG faring poorly in backgames
Date: Fri, 24 Jun 2005 09:00:14 +1200

Yes, gnubg can't learn from it's own rollouts in this case. It can't
learn from humans since you need tens of thousands of examples to
train a reasonable "bootstrap" NN. So the challenge is - find a simple
enough strategy (simple enough == can be coded without an NN) that
will play better than gnubg in those situations. If you think it is
easy go ahead - I tried and found it is not so easy. when we have that
we can make some progress.


On 6/24/05, Rob Adams <address@hidden> wrote:
> Hi, I'm new to this list... I've learned a lot from
> gnubg over the last few years and am very grateful to
> all who have been a part of making it what it is.  I
> wasn't sure how to reply to this thread, sorry if it's
> starting a new one.
> I disagreed with gnubg's evaluation and rollout result
> of a backgame position posted on GammOnLine {from
> Chicago Point's newsletter actually}, so tried to play
> it out and quickly realized gnubg was just lost and I
> wasn't learning anything.  So I tried just to beat
> gnubg.  Since then, I've played several hundred
> matches against Xbot and Tbot on gamesgrid {starting
> from the nack opening} which I suppose are based on
> gnubg 0ply and 2ply.  And I'm winning over 60% of 1pt
> matches by just throwing blots at it...  I'm not a
> great player either.
> So, sure, it is going to be a lot of work to improve
> gnubg in these types of games.  You can't just let
> gnubg play itself because it is terrible.  It will
> take some kind of supervised training.  But it needs
> to be done.  Right now you just cannot say that gnubg
> is a world-class level player.  Neil Kazaross's quote
> on the opening page of the website an easily defeated
> challenge; with anti-bot strategy, it is beatable.
> Note that it isn't really backgames that are the
> trouble, in "normal" backgames gnubg is fine (I
> think); it is many checkers "back" games.
> Specifically if you start all 15 checkers in their own
> home boards like an acey-deucy game opening ... from
> there gnubg just doesn't do well.
> Perhaps a new Neural Net for situations where there
> are many checkers "back"?  I really don't know.  Maybe
> use Snowie 4 to help train it (is that even legal?).
> Or maybe just use human v. human training... let a
> strong human player or two play both sides.  I realize
> this is asking a lot.  But failure to improve gnubg
> will eventually render it into obscurity.  Newer bots
> will come along that play better.  Maybe Snowie 5,
> maybe Zbot3, maybe something else.  But if gnubg
> doesn't keep improving, it will eventually lose out.
> And this is one area where gnubg has obvious,
> horrifyingly bad shortcomings.  After these are dealt
> with, more routine , small errors can be tackled, but
> the need for improvement will never end.
> I don't know how much this will change gnubg's
> evaluations in more "normal" positions like the
> opening rolls and replies: probably not much.  I
> understand if people don't think it is worth the
> effort.  I know it would be a lot of work for perhaps
> little gain.  But until it is done, gnubg is
> beatable...by Amateurs... are y'all happy with that?
> Rob Adams
> ____________________________________________________
> Yahoo! Sports
> Rekindle the Rivalries. Sign up for Fantasy Football
> http://football.fantasysports.yahoo.com
> _______________________________________________
> Bug-gnubg mailing list
> address@hidden
> http://lists.gnu.org/mailman/listinfo/bug-gnubg

reply via email to

[Prev in Thread] Current Thread [Next in Thread]