[computer-go] thoughts on 100,000 cgos games

[Don Dailey]

The "correct" komi probably can't be determined from these games 
because many of the good programs play to win the game, not territory.

However, we can probably determine if the komi is too high or
too low.

I have all the data in a sql database, when I get time I'll
do a win/loss for all the programs that have played a significant
number of games.

- Don



On Mon, 2006-05-01 at 17:33 -0700, David G Doshay wrote:
> Reaching 100,000 games seems like a good milestone for reflection, so  
> I hope that others will share their thoughts.
> 
> Of course, it is a great thing that Don has done to provide this  
> service, and I thank him for his initiative and efforts.
> 
> I have more of a list of random observations than coherent opinions:
> 
> 1) It looks like most any of the programs can have their rating swing  
> by 100 or more points, and it looks like it is the distribution of  
> other programs on-line at the time that is causing at lease half of  
> these swings. It looks like it is programs that are involved in  
> clearly non-transitive relationships with other closely rated  
> programs that this effects the most.
> 
> 2) With AnchorMan fixed at 1500, perhaps we should similarly fix  
> Random at zero. It has shown that it tends to less than zero. I think  
> this will really only change the scores of the programs that are not  
> too much stronger than Random. While this really does not matter, it  
> looks funny to have a program with a negative rating.
> 
> 3) While I had hoped that we might be able to do some machine  
> learning on SlugGo's parameters by having quick access to so many  
> opponents on this server, I have decided that without knowing who the  
> opponent is, any learning done in this context would not be  
> meaningful ... winning a game against any of the lowest ranked  
> programs does not mean that SlugGo's moves were any good and thus  
> gathering statistics about the success of moves in such a widely  
> varied field of opponents might be a big mistake. This is leading us  
> to some interesting discussions about machine learning, but we have  
> no deep conclusions yet.
> 
> 4) It seems to me that there should now be enough data to determine  
> if the komi being used is
> the best one. My expectation is that the "best" komi will end up  
> being different for different pairings, but hopefully something clear  
> will come out of the analysis.
> 
> I would hope that this discussion does not just turn into a cgos wish- 
> list, but if it does then please change the subject line.
> 
> Cheers,
> David
> 
> 
> 
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/