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

Re: [computer-go] future KGS Computer Go Tournaments - two sections?



On 9, May 2005, at 2:24 PM, Nick Wedd wrote:

I would like to see SlugGo playing in KGS tournaments, but I'm sorry, I won't be able to let it play in a "formal" division. Or indeed, if we decide not to split the tournaments into divisions, to let it play at all in regular tournaments. The same goes, in future, for the weaker GoFigure, and the much weaker DumbBot, two other "metaprograms" that use GNUGo, in various ways, for move generation.
Interesting. May I ask why you feel this way? See below for my specific concerns.

But I would like to see SlugGo (and GoFigure, and DumbBot) playing on KGS. I am sure people would like to watch it playing. One possibility is to have them play in a "casual" division. I don't know how you would feel about that, SlugGo won't get much of a challenge from the other players in the lower division.
What I am trying to discern is what purpose the two divisions are serving. If it is the separation of strong from weak then I do not understand why SlugGo would be put with the weak programs. If it is the separation of mature from new programs then I agree that at this time SlugGo is quite immature, and it shows in the lack of features like "real" time controls. In that case it seems that the criterion should be a specific list that programs must implement.

If the problem is that it is a derivative program ... will you also eliminate programs that are derived from Ander's early work? I understand your reluctance to get involved in deciding how different is "different enough," and I had thought that your earlier intent was to leave the decision to the developers of the base program.

Another possibility is to have a "slow" tournament. The opponents that don't like playing against GNUGo derivatives would realise that a slow tournament was likely to attract SlugGo, and know to avoid it. Can you remind me, what are the fastest time limits that SlugGo would be happy with?
OK, if speed is the reason then I don't understand at all, because I thought that I was supposed to self-select SlugGo out of any tournament where I thought the time controls were prohibitive.

Sorry, I am still not clear about what you are trying to accomplish and what you are trying to avoid.

The weekend of April 30 SlugGo and SmartGo were both competing in the Cotsen Open in LA against people. SlugGo entered as a 9 kyu, SmartGo as a 10kyu. SlugGo won one game (against an 8 kyu) and SmartGo did not win any. SlugGo was overwhelmingly ahead on points in another game (against another 8 kyu) when the other player started playing tactically complicated loosing moves in order to run down SlugGo's clock, and SlugGo did loose on time. The time controls were 40 minutes of basic time and five 45 second byo yomi periods. SlugGo took an average of 48 seconds per move at that stage, so the periods cranked right down. The only control I have on SlugGo's time is the specific branch factors and depth of global lookahead, which are set at the start of the game by reading a parameter file. For the second day I set the lookahead one step less deep, and the one time we went into byo yomi SlugGo took 20 to 22 seconds per move and never did use a whole period. So, SlugGo can play within "normal" tournament time limits ... it just cannot have the lookahead parameters set very deep. And this was with the traveling cluster, which runs at about 1/2 the speed of the G5 cluster.

Due to a mistake on my part, the wrong sgf files were saved (too tired I guess) and I do not know the percentage of moves that were identical to GNU Go's evaluation and how many were from "further down the list" and promoted by SlugGo's evaluation. We may try to regenerate that information.

Anders and I were able to meet on the Friday before the Cotsen and play 2 games between SmartGo and SlugGo. Anders chose to resign the first (even) game when SmartGo indicated that it was something like 100 points behind for a while. In the second game SlugGo gave SmartGo 4 handicap stones and won by 60, although if SmartGo had selected its second choice on just one move it should have won. I found it interesting that the basic architecture of SmartGo and SlugGo are similar: a wide set of candidate moves are selected and then global lookahead and evaluation determines which move gets selected for play. SmartGo chooses 24 wide at that top level. For the Cotsen, and these games against SmartGo, SlugGo would select up to 26 wide at the top.

Of course, if you wish you can let SlugGo play as a bot on KGS, outside a tournament, accepting challenges from allcomers.
Our main problem has been the willingness to let complete strangers tie up that many computers. I built up a new traveling cluster out of 24 Mac minis and one dual G5 tower for the Cotsen, and perhaps when it is online at the University we will consider making the mini cluster available. But I am more interested in having it play against other programs than against random people out there on the internet.

I am deciding now about taking the traveling cluster to the US Go Congress.

Cheers,
David


_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/