Hi Hideki, I think what I will do is use ELO and a simple formula for determining handicap. The formula will impose a slight curve on the value of a handicap stone, it will slightly increase with each ELO point. In other words a stronger player will benefit more from having an extra stone and the handicap will be chosen appropriately.
Of course the formula will be an assumption. I will either build something in to the server to make gentle modifications over time, or I will manually adjust the parameters from time to time based on the data collected from the server. It will be easy to tell if the handicaps are too aggressive or too conservative after a lot of data is collected. The initial formula will assume the stronger players on 19x19 CGOS need 1 handicap stone to overcome 100 ELO points. This seems to be fairly standard in servers and I think it's probably a good starting point. Since computer programs do not represent very strong players at 19x19 I don't think there is a great deal of "rank compression" at these levels, so I can imagine that this will be a reasonable starting point for the stronger players. Of course 100 ELO per stone will probably not work well with the really weak players (and may even be wrong for the strong computer players) so one way or another we will have to converge on a formula that tries to be as fair as possible. There seems to be a large range of computer playing skill on CGOS, from zero ELO to 2200 without any large gaps. This may look different at 19x19, we shall see. I think our formula will require 2 constants that can be adjusted to control the shape of the curve. I will come up with something but I will be happy to take suggestions too. Like everything else that has to do with ratings, rankings, handicaps, etc this is all an estimate and will never be 100% perfect. But I think we can make the attempt to fit the data according to the results to be as fair as possible. - Don On Mon, 2006-12-25 at 14:35 +0000, Jacques Basaldúa wrote: > Hideki Kato wrote: > > > In Nihon Kiin's ELO system(1), 1000 ELO is 1 rank, > > The Elo rating is based on two assumptions: > > a. The performance of each player in each game is a > normally distributed random variable. > b. All players performance have the same standard > deviation. (This is controversial, and other > rating systems modify this.) > > Anything else is arbitrary! Including how many Elo > points produce a given probability. > > Elo *tradition* (from chess) has always used the > convention: > > 100 points = 1/(1+10^.25) = 0.3599 approx = 1/3 > 200 points = 1/(1+10^.50) = 0.3204 approx = 1/4 > > 1000 Elo points give a probability of 1/317 > > It is obvious that Nihon Kiin's use a different > scale (may be 1000 Nihon Kiin's = 100 traditional). > > On the handicap subject: > > I am very happy to have a 19x19 server either with > or without handicap, so I welcome it as it is. > > Nevertheless, I have certain experience (not with > MC) of computer go with handicap and I can tell: > Waiting for the opponent to blunder is only a good > strategy if the handicap is lower than it should. > E.g. 7 kyu difference & Handi 3. If the handicap > approaches its real value, that does not work. > I have seen (many times) GnuGo not being able to > win a H7 game to an opponent more than 10 kyu > weaker. That happens because it had to invade > unclear positions. The more the invasion is > postponed, the worse. The weaker player simply > does defensive uninteresting play and so does the > stronger player (with better yose, but that's not > enough). If I (manually) use two or three turns > just to invade, GnuGo tries to save the invading > stones and that's more than enough to win the game. > > As I said before, its a different game and the > more accurate you determine the handicap, the > worse. If at all, handicap should always be > underestimated by a factor of 1/2. > > Handicap is used to make the game interesting > enough to white (but usually white still wins) > to honor a lower player with a learning game. > I hope weaker bots will learn at lot from the > games played against the stronger. ;-) > > Jacques. > > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
