[computer-go] KGS Computer Go Tournaments

[David Doshay]

On 5, Dec 2006, at 12:48 PM, Nick Wedd wrote:
> But I understand that SlugGo is off sick at present, and as it is  
> the program that would most enjoy such time settings, maybe I  
> should wait until I hear that it has recovered?



Hi all,

One of the most persistent misunderstandings with respect to SlugGo  
is that it is constrained by thinking time. Our earliest data did  
indicate this, but that data was from games against GNU Go, SlugGo's  
underlying engine. In those games we found that by deeper search,  
taking up to 6 hours per game, we could beat GNU Go by ever larger  
margins, with the top not yet in sight ... in other words, we would  
have an even higher winning percentage with even longer lookahead.  
Unfortunately, we discovered that this was really what we now call  
the "evil twin" effect, and it is not true against less well  
understood opponents, either human or other computers.

Our published data indicates that against non-GNU Go programs, much  
shorter lookaheads are optimal, with the longer search times leading  
to worse play than the shorter. As discussed in the last week or two,  
this is because SlugGo's algorithms base move decisions upon a  
combination of initial board and leaf node evaluations. The deeper  
leaf nodes become increasingly unlikely in the real game, (the  
opponent branches from the lines of play we have checked) and thus  
their value in the final decision process is lower. I think that we  
do have more work to do investigating branching factors in the  
lookahead paths, but preliminary data also suggests that too much  
branching does not help and in fact also hurts. This makes little  
sense to me, but it is what our data shows. I eventually will get  
back to investigating this oddity, but not for the next few months.

The cooling is back on in SlugGo's machine room, so SlugGo could play  
in an tournament if, according to the rules, the GNU folks give their  
approval. The last few times I asked I did not get any reply, but  
then something broke so it did not matter in the end anyway.

The software bugs that I knew of are fixed, so SlugGo can play. I  
would be happy to enter SlugGo in another KGS tournament, but if you  
want it to be the "long time" tournament it would be my preference  
that we wait until Feb or March of next year so I can further  
investigate wider branching factors. But if you don't want to wait  
and the GNU Go folks are willing to let me play, I will enter  
whenever it works for everybody else. It does seem to me that the MC  
folks could also use a few months to get some memory management  
issues worked out if they are going to take more than something like  
10 - 20 minutes per move.

Cheers,
David