Rank - Absolute Levels
Absolute ranks
The strength scale has two absolute points: random play at the bottom, and the hand of God at the top. However, both are very difficult to estimate, let alone to measure.
So, in practice, there is no such thing like absolute ranks, but see Rank - Worldwide Comparison and the European Go Database (statistics section) for two ongoing attempts. The only way to establish such a thing would be to have a massive tournament, with every player in the world playing many games against other players from around the world, with nobody's skill improving. (The tournament idea sounds like fun, but by the time it was over, i would hope to have improved a stone or three!)
There is not a single rank-setting authority. Rather, there are several rank repositories (each national or continental go association is one such, as is each Go server that tracks ranks), and they do not always quite agree. So one person may be regarded as a 1-kyu in Osaka and as a 1-dan in Tokyo, for example.
Within rank repositories, there may be variation such that a person with a weaker rank can easily beat a person with a stronger rank. This is because ranks do not always keep pace with changes in players' abilities. If a small group of players within a repository mostly play amongst themselves, their ranks may seem stagnant even though the group as a whole is getting stronger. When such a player contests someone from outside the group, s/he will be surprisingly strong for her/is declared rank (which the outsider might regard as sandbagging). The larger the repository, the higher is the chance of isolated pockets of players whose rank isn't accurately compared to the general level of strength.
Discussion
Jasonred Sometimes, I wonder... are go players in general getting stronger, raising the bar, so to speak? So, a 20 kyu now was equivalent to a 17 kyu 80 years ago? (or whenever the system came into existence).
Someone mentioned to me that once, 25 kyu was the rank given to an absolute beginner who understood the basic rules, with no knowledge of techniques. Apparently, nowadays 25 kyu means you have a certain degree of skill as well. And 30 kyu is the lowest possible rank. Weird.
Possibly, once upon a time, a 1 dan really would have had to give a 25 stone handicap to a 25 kyu...
Bill: I have given a 40-stone handicap based upon ratings. To my surprise it was a close game. ;-)
When I learned go, people did not take double digit kyu ranks seriously. After all, such players can have an insight that advances them several ranks at once.
These days, with internet go servers, double digit kyu players have become more visible. I also gather that they tend to play more with other double digit kyus. (Certainly if you were the only one at the local club you could not do that. ;-)) As a result, I think that two things happen. First, they reinforce each other's bad habits. Second, they get stronger as a group without the rating system reflecting that. So 25 kyus of today are probably stronger than 25 kyus of five years ago.
Mef: I think another problem that has arisen from go servers is that having enough players of similar rank means there are far fewer handicap games, since many players prefer to play even if the option is available. The result is that ranks may no longer be a full stone apart, since one really needs alot of 3 and 4 stone handicap games to make sure the spacing is right. In my experience this problem is not as prevalent at the upper ranks since they play more of such handicap games.
Tas: 40 stones? So he was a 35k? Based upon what?
exswoo I don't think Go Theory could really progress if such a thing didn't happen...although I'm not sure how much difference this makes (I would venture to say that the differences in skill would be most evident for the top players but barely at all for the amateur levels). Of course, the relative nature of the rankings makes this point moot. ;)
Ideally, as a kyu ranking is defined in handicaps, but dans are not, the natural point for shodan would be the point at which the handicap=rank difference system broke down?
Andrew Grant: This isn't true. Professional dan grades have no relation to the handicap system, but amateur dan grades certainly do. In any case, to say the handicap=rank difference system breaks down at any point is a circular argument, since all grades are man-made. You could invent a ranking system with a discontinuity at any arbitrary skill level; would that make it a "natural" point to call it shodan?
zinger: One of the dan players at my local club says that the original definition of (amateur) shodan was the weakest player who could beat a professional at nine stones. The other ranks were extrapolated from that, using the one stone per rank rule. If we apply this definition today, where would shodan fall on the various rank scales?
Bill: in one of his books Segoe says that an amateur shodan should take 4 stones from a pro. There has been amateur rank inflation since WWII.
- DrStraw: I was promoted to shodan in the UK in the '70s. I was recently looking over some games I recorded from that time and I doubt whether I would make it now. Dan ranks seems to be stronger now than 30s years ago - at least in the west.
Bob McGuigan: Some venues in which pros and amateurs play with the same ranking scale are various internet go servers (IGS, Tom, Cyberoro, Dashn). On IGS, for example, it seems that the 7d*, 8d*, and 9d* ranks are populated almost entirely by professional players. So on this basis some idea of the relative strengths of amateurs and pros can be determined, but of course there are issues such as the fact that server go does not seem to be directly comparable to "real life go".
Dieter: I'm a 2d. I think I stand a chance against Lee Chang-Ho with 9 stones but I wouldn't bid too high on it. No way I can win with 6 stones if a pro takes it seriously.
aokun: I just got back from a business trip to New York, where I go three or four times a year. I play in the evenings for at the go club on 52nd street and I love it, because, while I'm 7k on KGS and 4k in the AGA, I am at least 2d amateur in New York. There is always an amateur sensei from Japan in residence there and last night I was able to win a 3-stone game from the current guy, a 5d. My eyebrows went up when he said I'd be a 2d in Japan, but he just said, "inflation" and smiled.
My conlusion from limited data is that there is some rank deflation going on because of go servers, with huge numbers of rated games being played online by a motivated subset of on average younger players. Think about this: take a group of people all of one rank and isolate them, causing them to play thousands of games only with each other over the space of a couple of years. Their ranks will spread out as smaller differences in skill become the prime determinants of wins and losses, as some improve in skill and others get stale, and from simple statistical dispersion, particular if they are not all playing the same number of games.
Online players are not all of one rank, but they do tend either to have played a lot in the past or to be playing a lot now and learning quickly, and, as someone pointed out, with so many players available, they're not doing a lot of 4-stone type games. Hence, lots of, say, 10k players who are only a couple of stones worse than some 3k players and who can beat 3k club players.
Malweth: When I played recently in Japan at the ki-in I beat a 2d easily. I am ~3k AGA. My teacher explained the reason for this: the ki-in sells rank certificates so it's to their advantage to increase your rank quicker.
The oddest part was that they kept asking if I was 3k inteneto and I had to keep saying no -- I'm about 7k online. It is kind of nice when they're impressed at your skills when you're really only 3k :D.
I also imagine that the stronger players don't usually play at the ki-in because it's more expensive than local clubs. I wasn't in Tokyo for long enough to check out any other go clubs. 2 days is not enough!!
aceofspades: Most of the discussion here is about rank movement in groups, but there's nothing here, except for the first line, about the title - absolute levels. So: I've been thinking of two theoretical absolute ranks. One is simply how you do against the Hand of God. I don't believe anyone has calculated the maximum score difference in territory scoring, but for these purposes (not to mention others) area scoring is convenient. Then every computer program and every player who plays against kami no itte has a result ranging from loss by 361 + komi (if they claim nothing on the board and take black) to draw (perfect, integer komi assumed). Naturally the result takes averaging over games to be fine-tuned, or games could be played on a larger board. Theoretically the same system could be used against random play, but almost everyone would win the whole board. Perhaps large handicaps and a large board would give rise to a vairety of player performances though. Finally, a more interesting practical idea occured to me while writing. Using either "x stones stronger /weaker than" or "loses by 26 points on average" systems, one could rate everyone in relation to a computer program, using large handicaps and komis and large boards for an even greater variety of ranks. Although, large boards would cause problems because the significance of most go knowledge is dependent on board size. Anyway, it's an interesting idea because it would provide stability to ranks - 9 stones stronger than a particular version of GnuGo is a level that never changes and can be compared to anyone else who has played that GnuGo.
aokun It is a very interesting idea, worth thinking about more. (a) because winning is not dependent on the score difference, I think the "26 points on average" approach is not good, as it gives a different incentive to the player. If you are ranked on wins and losses and you are 15 points behind, you might make bolder, riskier moves, more questionable invasions, trickier end-game tesujis and the like in an effort to turn it around. If you are ranked by point difference, then you might not take the +/- 25 point risk, but instead play cautiously, since a -26 player getting a -15 score is an improvement. This would produce different play even than if all handicapping was done with variable komi. (b) Ranking against fixed software, like GnuGo might be of questionable value because beating the software can involve different moves than you would play against a person. ("Get Strong Against GnuGo"?) In getting better against software opposition, I have found a few patterns, not joseki both because they aren't even result and no human opponent would follow them, that invariably produce a good result for me. The software doesn't learn, so they always work. There are probably a zillion of such patterns to be found and if ranking against software were to count, there'd be a parallel "GnuGo Theory" developing. Against a purely random player is a tempting idea. Would that require different play?
blubb: This idea is what has induced my last year's evaluation attempts on the rank of random. I agree with aokun about the two problems involved. The first doesn't seem to be such a hard one, though. Imposing a winning threshold different from 0 points equals altering komi. It does probably affect the playing style, but since this applies to all players, it does not spoil the results. Yet, to make these "komi ranks" useful for usual games between human players, the margins have to be translated to handicap stones. Anyway, the second problem is trickier. Knowing the weaknesses of the particular engine in use is a huge advantage (just ask any Go programmer about this, or see
this article). We could smoothen the consequences by having a couple of different engines play with a large number of players, e. g. at online Go servers, and use them as anchors. However, as long as the first-sight strength of engines differs from their asymptotic ones by many stones, the basic flaw would remain.
SloSilver? Actually, I can't think of a way even the 'Hand of God' could defeat a 180 stone handicap by a properly instructed newcomer, so using 361 (the number of points) as the bottom end would probably be inefficient. I find it interesting that ranks are tied to 'stones' in some systems and that some of those systems allow for a rank of lower than 180. So the logic of the rank coming from a stone difference doesn't seem to flow all that well for me. In concept, if 180 stones is the level on which it would be completely improbable to defeat, and likely the defeat possibility would increase on a curve from that point (using the 'perfect player' concept espoused), it should be possible to generate a rank system that kept the win/loss odds in the 33-66% chance range by using a 0-180 absolute stone ranking system (more likely a 0-99 stone handicap rank system to keep it in a better statistical scope). It would not ever be a perfect 50/50 chance but you could pretty easily keep it between 1/3 - 2/3 win possibility. Interestingly enough, if you adopt the 0-99 rank system, 0 being best, 99 being total newbie, you might give a even the most disparate players an even game (newbie vs. pro), if that is indeed the purpose of a rank handicap system.
kokiri - I'm game. I say there's no such thing as 'absolute rank'. There seems to be this unwritten principle that how people play on the board is an approximation of some universal constant, their rank but really it's surely the other way round - the rank is a way of trying to approximate all of a player's game - all the blunders, the oversights, everything, into a single number. For most amateurs, consistancy is such a problem that this is at best a hazy approximation.
One of the best things about go is the ability to choose a handicap that gives a decent game between players of a wide difference in ability. That this gives rise to the ranking system is understandable, but it also seems unfortunate that we then get tangled up in debates about rank deflation, inflation, anchoring, where shodan should be placed etc. To me none of this really matters - it's meaningless really. Worse than that, the ranking system encourages bad behaviour - such as people on KGS who won't play '?'-ranked players, focusing more on rank than the go.
tb: I avoid playing '?'-ranked players on KGS, though I do play them now and then because somebody has to. But the reason is not some kind of rank protection. It's because I don't much enjoy playing a wildly mis-matched game, and nearly every such game I play is wildly mis-matched. When I have a choice of two games to play, one that is a close match, and one that is likely to be a bad matchup, why should I choose the latter?
Malweth: I avoid playing on KGS lately because of this. I've been finding that Yahoo is one of the BEST servers around simply because 1) you never know what kind of game you'll get and 2) your chances of playing a much stronger player are about even with your chances of playing a much weaker player (I'm 3k AGA). On KGS, if you have a solid rank your chances of playing a stronger player are slim. If you don't have a solid rank, your chances of playing anyone are severly limited. Not impossible, of course, but I don't want to wait 15 minutes or more to play a game.
If we didn't have ranking systems, would we have to invent them? Probably we would, but only in situations such as the tournament scene where we play people we don't know. In a club environment, where there are few people who haven't a kokiri number of more than two, it's unnecessary and I would rather throw them out, and play kadoban style systems with other players on an individual basis.
Velobici: Agree with kokiri. I have given up on using rank or rating to select the number of handicap stones to use a my local go club. Rather I have been tracking my performance against other players and recommending a change in the number handicap stones following a kadoban system. Interestingly, the president of the club, who is also the strongest player, does the same. He occasionally will give a player the opportunity to place one handicap stone less for each consecutive win till the first loss. In this way, one player moved from eight to four handicap stones in four games. The club president is very strong playing against a handicap; stronger than his even game play. A handicap from being the strongest player for so long and playing so many handicap games.
Malweth: I also agree... "Absolute" rank is impossible, simply because of play styles. It is easy for a certain 5k to be strong against a certain 3k, for example, even though both are still 5k and 3k respectively (on any given server).
aceofspades: Whew...I'm back to Go and SL after about a year just to comment on this page :-) Anyway, I certainly accept that in a certain sense humans never have an absolute rank, which is a concept based on seeing actual human play as a stochastic fluctuation about a Platonic "true rank". Although you have to admit that practically nobody would consider this justification for abandoning all notions of quantitative characterization of strength.
Regardless, what I really meant to ask was: Does a Turing machine - assuming of course that you have a proper Go-description language it works in - have absolute rank? And if not - consider Go for a second as if there were no such things as human players, as if it was merely a benchmark-type task you were giving to a few different programs on your computer. Specific to that group, does each have an absolute rank? It would seem obvious that this is true, since you can play out all possible games within that finite number of programs. Now consider taking the limit of a program's rank according to a suitable such group-definition, as the pool of programs increases to infinite size, populated by randomly chosen programs from all possible well-defined ones.
And yes, I'm aware I'm being highly theoretical.
Finally, one more small thing. The very top of this page gives random and perfect as the absolute ranks. But surely there is at least one more separate one - namely, always passing. Forgive me if this has already been mentioned and I missed it. Furthermore, one can derive further levels from combinations of these by playing according to each of the three strategies with a specific probability. For example consider the "half-vegetative, half-omniscient" player: they pass with probability 50% and play the optimal move (including possibly as pass) with probability 50%. In fact, one might consider this to be in some way a half-way point in ranks.
Actually, now that I think of it, maybe this whole "derived ranks" thing has a deep connection to the discussion on absolute ranks for humans or computers. Because consider the follwing: define the level of the fundamental player the Hand of God as 1, and always passing as a fundamental player with level defined as 0. Then a program that, as described above, passes with probability 1-p and plays optimally with prob. p is the fundamental player with level defined as p. If a human or program draws most often, among all fundamental players, with the fundamental player of level p, then they have level p.
Well? Please ready your catapults with volleys of scathing criticism...1...2...
Anonymous: ...3! It is far from obvious that even Turing machines have absolute rank. For instnace, in your system, we can imagine two AIs which behave identically except in reaction to a midgame pass. They would perform identically against all sensibly programmed AIs, but would have wildly different absolute ranks.