[Top][All Lists]

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

Re: [Fwd: Re: [Bug-gnubg] Re: Battle of the Bots]

From: Joseph Heled
Subject: Re: [Fwd: Re: [Bug-gnubg] Re: Battle of the Bots]
Date: Mon, 06 Jan 2003 12:55:03 +1300
User-agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.2) Gecko/20021202

Here are 48 more games.

So far I have
W wins 46 + 0 + 0    = 46 points
B wins 18 + 21 + 11  = 93 points

(46 - 93) / 96 = -0.489

If you got different results, can you send me a record of the games please.


Joseph Heled wrote:

In my small sample (48 games), W won 23 time, which is 48%

I guess it is too small, but this is in the Snowie 4 range ...


Jim Segrave wrote:

On Sun 05 Jan 2003 (08:41 -0600), Neil Kaz wrote:

Yes..Snowie's cubeless rollouts assume no Jacoby rule.

I will try to import these games into Snowie 4 and have it analysis them and
see if it thinks there are gross blunders.

There must be some enormous blunders after hitting a checker from the
backgame side since my 1296 game 2-ply rollout of the basic 31 backgame
resulted in the backgame player only winning 32% while losing a gammon over
35% for a cubeless equity of -.806.

This is to be compared with Snowie 4 where the Gammons lost are about the
same, but the backgame player won 44.7 % games and the cueless equity
is -.554 !!

I'm curious about this, so I decided to start a 1296 game rollout
cubeless (and cube disabled). untruncated, no Jacoby, no bearoff
DBs, with some almost promiscuous filter settings:

at 0 ply, accept 1 add 20 within .250
   1 ply, accept 1 add 10 within .200

I wanted to see if there were some moves which were likely to be
getting discarded too early for being non-promising.

Just started, I'd guess it will be about 6 to 10 hours to roll
out. This is with the 0.12b nets, not the .13 if this matters. I'd
guess if it gives similar results, it indicates that, for this
position at least, theres a systematic error in gnubg's approach
causing this or (just a possibility), there's something wrong with
Snowie's play.

For what it's worth, after the first 60 games, gnubg is showing
.3451 wins (std .0732), gammon losses of .3717 (std .0073). This
suggests it's going to settle on the figures Niel got and that it's
probably not simply a failure to take a move forward a couple of

Is there a way to get a few samples of what Snowie plays from this
position to run through gnubg to analyse? A 12% difference in wins
seems huge enough that one or the other is making a fairly big
mistake and, I would think that there should be a point where one
bot's analysis of the other bot's play should be screaming blunder and
it might be possible (given the large difference in outcomes) to
identify who is "doing the wrong thing".

(all the above to be taken with a certain large grain of salt, as I
don't rate myself as a backgammon player at all.

Bug-gnubg mailing list

Attachment: n5.sgf
Description: x-sgf

reply via email to

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