[Top][All Lists]

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

[Bug-gnubg] Re: gnubg bug?

From: Christopher Yep
Subject: [Bug-gnubg] Re: gnubg bug?
Date: Tue, 04 Aug 2009 16:56:40 -0400

It looks like you are correct; the problem is due to pruning.

Using a recent build (0.90-mingw 20090721):

With pruning turned off, gnubg gets the correct backgammon values for both moves (21/17 19/18: 8.5%, 20/16 19/18: 9.3%) at all plies (0-ply, 1-ply, ..., 7-ply).

With pruning turned on, gnubg gets the correct backgammon values at 0-ply, 1-ply, and 5-ply, but gets incorrect backgammon values (21/17 19/18: 16.6% BGs, 20/16 19/18: 12.4% BGs) at 2-ply, 3-ply, 4-ply, 6-ply, and 7-ply. It's strange that gnubg gets it correct at 5-ply, but not 6-ply or 7-ply, but note that the game always ends by the end of the 5th ply (the top player is guaranteed to bear off in 3 rolls).

How much slower is gnubg when pruning is turned off in racing positions?


At 03:44 PM 8/4/2009, Michael Petch wrote:
Howdy Team,

I find that a lot of the fringe cases (with such incorrect results) like
this usually occur on the race evaluator near end games. If this is a result
of pruning, maybe we might consider not pruning when the race evaluator is
being used?

On 04/08/09 1:23 PM, "Michael Petch" <address@hidden> wrote:

> This is cross-posted to bgonline:
> We call it "Gnu Math"?
> On a serious note this seems like a prime example of neurel net pruning at
> its worst. I don't have GnuBG handy but what do you get by turning "Pruning" > off? I recently had a discussion with Neil Kaz (I think it was Neil - but if
> it wasn't, it was another world class player) - about the settings I use in
> Gnubg. Besides using a Huge Filter I spend the extra time and turn the
> pruning off. I've seen this type of thign before.
> Has anyone tried the position without pruning? Sorry I can't test it from
> here.
> On 04/08/09 12:45 PM, "address@hidden" <address@hidden> wrote:
>> http://www.bgonline.org/forums/webbbs_config.pl?noframes;read=45761

reply via email to

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