Re: [Bug-gnubg] New build of GNU Backgammon for Windows available!

[nedc]

I post here a reproducible case where 0.15 rollouts 
perform significantly poorer than 0.14.  I have 2 .sgf 
files with the rollout results if you would like me to 
send them.  I say poorer because it is essentially a 
one-sided checkerplay position.

If I get time I could also try this with a larger search 
space for the pruning net but I wanted to have identical 
settings for both nets initially.

I rolled this out with 0.14 2-ply and the wins were 75.1% 
with gammon losses at least 4% lower, and the standard 
error was 0.001 wins, and 0.006 cubeful (one-third the 
error of the 0.15 version) after only 600+ trials.  I was 
unable to save that 2-ply rollout result, so I am 
repeating the 0.14 rollout today and could forward the 
.sgf if you like.

Here is a discussion thread at GOL about this position 
with several additional rollout results including 0-ply 
GNU and Snowie 4.1.

http://www.gammonline.com/members/board/config.cgi?read=85765


My general observation after several rollouts with 0.15 is 
that the error rates for cubeful equities are taking 
several times more trials to converge than the 0.14 net.

This is a general observation as an experienced user, not 
a claim as a result of any scientific study.  

While we can all enjoy the immediate benefits of match 
analysis that is twice as fast, it could just be that it 
will take us some time to rediscover the ideal rollout 
settings with the 0.15 version.

--Ned


On Sat, 16 Oct 2004 12:42:24 +1300
 Joseph Heled <address@hidden> wrote:
> ... Please post a reproducible 
> case, with a smaller n as possible, full setting, and 
> seed ...
>
> -Joseph

5 prime redouble rollout_prune_2ply.txt
Description: Text document