Talk:Go ranks and ratings

From Wikipedia, the free encyclopedia

You have new messages (last change).
Image:Go-60sq.jpg

This article is within the scope of the WikiProject Go, a collaborative effort to improve Wikipedia's coverage of the game of Go. If you would like to participate, you can visit the project page, where you can join the project and see a list of open tasks.
Use the + tab at top to add new comments at the end. Sign comments with four tildes: ~~~~.

Contents

[edit] Separate page created

I've created this separate page, on a topic of basic interest, since it was crowding go (game). The existing go rank page has been merged here. Rather clearly, there is a long way ahead of copy editing to bring this all up to standard. Charles Matthews 09:49, 7 Sep 2004 (UTC)

11:09, 9 Sep 2004 (UTC) Tommie 11:09, 9 Sep 2004 (UTC) Tommie: I would like to widen the rank table at the bottom, to show 9p at top left and 30 k bottom right. How should I amend the cell spacings? I do not want to mess up.

Maybe the table isn't a great format, in fact, if we want to annotate. I just moved it in from the old go rank page. Some sort of graphic, plus a bulleted list, might be better? Charles Matthews 11:54, 9 Sep 2004 (UTC)

Tommie 18:30, 9 Sep 2004 (UTC): Actually I put already a figure at Go & Ranks, but I could not get it displayed, made smaller in size and moved to the right:

In Germany and The Netherlands a "classes"-system (German: "Klassen") was established. It comprised a further subdivision into Kyu/Dan halfgrades with classes 18 and 17 = amateur 1 dan with the 17 being on the stronger side. It is still in use for club ladders etc. where you pro-/demote after a won/lost game. [Image:http://www.gobond.nl/images/dan-kyu2.gif|right| Kyu-Dan-classes]

I have made extensive edits on the page today. It needs further work because some of the material is duplicated two or three times. The tables need work. The list of ranks at the end is useless (of course 5k is between 4k and 6k). Hu 22:19, 2004 Dec 24 (UTC). I think the better exposition is early in the article and later sections that have duplicate material should be merged forward and the duplication eliminated. Hu 22:41, 2004 Dec 24 (UTC)

[edit] Greater than 9 Dan? [Done]

Does anyone know any reason why there isn't any higher ranking than 9-Dan?70.111.251.203 15:01, 7 February 2006 (UTC)

My understanding is that ranks higher than 9-dan were traditionally reserved for Generals and members of the emperial family. Stuartyeates 15:18, 7 February 2006 (UTC)

Pro 10-dan is not reserved for Generals and members of the emperial family. Pro 10-dan is a special title which is given to only 1 person in Japan, which is the champion in the 10-dan match. It is similar to the case where 8d is a special rank among amateur Go world. For example, Kato Masao once got pro 10-dan title four straight years (five titles total). --Wai Wai 14:03, 10 July 2006 (UTC)

I am a go player and having looked at this site I feel like saying that the suggested correlation of elo ratings to go rating is not very accurate at all. If the ratings are reasonable, a difference in 4 ranks in go (assuming the player rankings are accurate) would correspond to a very low chance of the stronger person losing. By the time there is a difference in 6 stones in rating I would call it a zero probability of losing. Also a 1pro would probably beat a 6d amateur with a much higher probability than just 80%.

Ranking system is not a system which we should place absolute faith on. It's true even for the strictest ones. Although it's not common, a lower pro dan player can still beat higher pro dan player. At the extreme, it is possible for a 1p to beat a 9p. For example, Abe Yumiko 1-dan beats 9-dan --Wai Wai 14:03, 10 July 2006 (UTC)

[edit] Should we reverse order of table?

The description in text is from kyu to pro dan. However this table is from pro dan to kyu. It seems to be a bit inconsistent. What do others think? Can anyone fix the table?--Wai Wai (talk) 03:33, 1 August 2006 (UTC)

[edit] Statistical validity of "Winning Percentages" table

Like the commentor above ("...the suggested correlation of elo ratings to go rating is not very accurate...") I'm concerned with the "Winning Percentages" table in the main article, on two grounds:

1) Math

The kyu-dan ranking system in Go is premised on handicap stones; almost every amateur game is handicapped, so a 4 dan gives 3 stones to a 1 dan. Thus, I am concerned that the author does not have game statistics to justify the "Winning Percentages" table, but inferred the result by analogy to chess. That is to say, I doubt that a difference of one standard deviation in the distribution of players ordered by strength from handicap games, gives the same winning percentage in even games that one SD in chess does, and there is no mathematical reason it should.

Since the creation of internet Go servers, this argument hold much less weight. There are at least 10 servers, on which hundreds of even amateur games are played daily, and the results are typically meticiously recorded, since recording game stats is something computers do very well. Translating these results into results recognised by national bodies is another thing, but it should at least be possble to analyses these multitude of results. Stuartyeates 13:57, 13 September 2006 (UTC)
2) Experience

Some of us suspect (merely from personal experience) that winning percentages are higher, in Go, for fixed skill discrepencies; i.e. a two stone handicap might give reliable 50% results for two players, but the stronger one (taking white, giving the two stones) would virtually always win an even game, with a higher percentage than inferred from the table for a 0.7 SD(approximately) Elo rating difference.

Anecdotal Example: I myself have never beaten someone more than two stones above me, in an even tournament game (or even in serious club games), but I have beaten people more than 0.7 SD above me in tournament chess (over 1 SD) (but I have played a great many more tournament chess games. Ironically, I'm about the same strength in both, 2050 USCF and 1 kyu AGA)

N.B. the "Greater than 9 dan?" subtopic is not the best place for this discussion, but I wanted to follow the comment cited, and this is my first attempted contribution to the Wiki. --Peter a.k.a. yoof

P.S.: I was about to link http://en.wikipedia.org/wiki/Arpad_Elo but have just read it for the first time. The statement "Elo was giving US players lower ratings than they deserved" is imo contraversial, but not disputed by any Chess organization to my knowledge. After the USCF modified the rating system, following Elo's retirement, around 1980, USCF ratings became consitently higher by about 100 points than FIDE ratings for the same players. The first master I beat (about 1983; I had been inactive from 77 to 83 for University, and my own rating jumped 100 points very abruptly when I returned) was 2200 FIDE (international) but 2300 USCF (american), and I began to track the differences (among the minority of players with both ratings) on the tournament crosstables. Some of us concluded that the USCF had introduced inflation for the apparent purpose of encouraging new players (the Fischer Boom had ebbed). I note that the chessmaster and professional statistician David Burris disagreed with me on this point (which is why I gave up trying to debate this). Peter H. St.John, M.S. 19:51, 4 August 2006 (UTC)

[edit] A Query

The depth aassertions are useless without references to other games that use an ELO system. Scrabble, Chess, Backgamon, Shogi, Amazons, Twixt - can we include comparisons with any of these?

Also it should be noted in the table borrowed from Ales Cieplys website that the Grades are often awarded by different sources. Since Claimed Grades are often not equivalent to Actual Strength, the statistical validty of that entire table has to be questioned. -Zinc Belief

[edit] Grammar

  1. I inferred that the clause "K is 30 and 20 for players below 2400 ELO" should read, "K is between 30 and 20 for players below 2400 ELO". Please correct this if I have made a mistake.
  2. Both the terms "go" and "Go" are used extensively throughout this page (presumably by different authors). Is there a consensus on which is correct for Wikipedia, and if not can we reach one? Note that the same question has been asked on the page Talk:Go (board game), and any comments you may have should really be made there (and the consensus applied to this and other go/Go pages if it is reached).

Thanks, Stelio 21:24, 14 September 2006 (UTC)

Go is the current choice to distinguish the game from the verb.--ZincBelief 16:19, 20 September 2006 (UTC)

[edit] Rate of improvement

I've seen similar numbers elsewhere regarding the rate of advancement (especially at the low kyu levels), and I must say that I'm highly dubious. At least for adults, these numbers seem to be inflated to the point of discouraging people from continuing to learn the game (although maybe it's more attractive to first-time players, I can't say), once they begin to imagine that they must not be cut out for playing it if they're advancing so slowly.

Are these numbers for people spending 8 hours a day with an instructor? Jdmarshall 03:23, 10 December 2006 (UTC)

I don't know about Go, but I can comment from my experience with Arimaa. The Arimaa server has a range of computer opponents rated from about 1100 to 1900. I have seen complete newcomers enter the system and master the top computer opponent within weeks, i.e. in less than 100 games. By the same token I have seen people play 1000 games over the course of months without mastering the top computers. I am not sure whether to attribute the difference to innate talent or method of study, but it is in any case it is misleading to say something like, "Newcomers can beat all computer opponents in under a month of study," because it isn't true of all newcomers.
Someone with more Go experience than myself should make the edit, but I believe the minimization of the difference between lower ranks is overblown. I suspect it is true that, "A jump from 30k to 10k could happen within weeks or even days for quick learners," but it is equally true that it could take some newcomers a year or more of play to get to single-digit kyu. It is simply rude for "quick learners" who breezed through introductory material to dismiss low-rank distinctions as insignificant. --Fritzlein 19:32, 10 December 2006 (UTC)

[edit] Game Depth

The section "Game Depth" should be deleted; it is (mistaken) Original Research, and it does not cite it's sources; but the interest level is high and hopefully we can replace it with someting statistically professional, if not insightful. I'll quote the whole thing here and critique it line by line.

The ELO rating depth also states something about the depth of the game. In the abstract, the total depth of a game is defined by the number of ranks between a random player and the theoretical best play by an infallible creature.

Actually, no. Elo ratings say nothing whatever about the "depth of a game". They are a percentage measurement based on Gaussian Distribution; about 68% of Tiddly-winks players would have an Elo rating of between 1200 and 1600, same as in Chess and the same in Go (if the mean is taken as 1200 and the standard deviation as 200, which is what Elo uses in Chess) and any other competition of skill.

In practice, the depth of the game is the number of ranks between a beginner and the best player in the world. [Who's definition is this?] The number of ranks in this latter, practical definition is increased by the age and popularity of the game, as a richer literature and greater playing pool both tend to move apart the end points of the scale. Even so, the practical definition allows a rough comparison of Go to chess, as both have an extensive literature and a huge playing population.

A rank in this definition does not necessarily correspond to a traditional Go rank. Using the EGF scale of standard deviation there are about 25 standard deviations between 20 kyu and 9 dan.

If this were true (that go ranks were Normally distributed, which is true, and that there are 25 standard deviations from 20k to 9d, which is not accurate) then the Mean (average) would be about 6k, and almost sixty-eight percent of all players would be between 7k and 5k. (68% of everything in a Normal Distribution lies within one standard deviation of it's mean. 97% lie within 3 SD's, so 97% of us would be ranked between 9k and 1k. The standard deviation for go is closer to 3 stones than to 1.)

Chess ratings run a range from about 1000 to 2800... No they don't. The minimum rating is about 200; the standard deviation is about 200 and the mean is about 1500, so the range would go about equally with half the players from 1500 up almost 7 SD's to 2800, and half down almost 7 sd's to about 200. (However, much more attention is paid to ratings of professionals close to 2800 than to beginnners close to 200). You can get statistics of chess ratings from the USCF's website at uschess.org.

... but they are differently scaled. Converting the FIDE formula...to be on the same scale as the EGF formula yields...and (2800-1000)/174 = 10.4. Therefore, when converted to be on a similar standard deviation, we can say that chess has a depth of about 10.4 compared to a depth of about 25 for Go.

This is comparing the precision of one measurement (incorrectly) to the precision of another; it's like saying, my one pound apple is better than your 0.493001 kilogram orange. Both games have averages, so does tiddlywinks if you record results of tournaments; and someone will be 99th percentile, in any game with more than 100 participants. Also, the number 1000 in his formula means nothing at all. I think he's mistakenly using the mean from an old AGA rating alternative.

It is a bit artificial to cut off chess ratings at 1000, but this is no more artificial than cutting off Go ranks at 20kyu.... There is no such cutoff. For a published example, the rating formula described at http://beta.uschess.org/frontend/section_171.php uses 750 as a starting point for rating beginners who do not have prior ratings.

...The USCF measures chess ratings down to zero... I thought he said they were cut off at 1000?

...usually among kindergarten players, so one could argue that there are as many as 16 ranks of depth in chess. However, the AGA also has no cutoff, and measures ratings below 40 kyu as well as above 9 dan, so if chess can be said to have 16 ranks, then Go can be said to have over 50 ranks. Well, go ranks are a 2 digit number, and Chess ratings are 4 digits, so Go has only 40+9 = 49 divisions and chess has 2800-200=2600 divisions, so what, is he saying chess must be 50 times more deep than go?

There's alot to be said comparing go and chess, and for some reason we all want to. Here are some observations of mine own.

  • The complexity of a game can be measured with Game, Ergodic, Complexity, and Coding Theories. One could bound the computability of determining a winning strategy or the existence of a winning strategy, calculate the entropy of a board postion, etc.
  • I would estimate that the statistical reliability of the outcome of a game of go would be comparable to the reliability of the outcome of a two to four game match of chess. I don't know of anyone measuring this; it might be a good project for statistics students in college. There is something to be said for the aphorism chess is a battle, go is a war; not that war is more complex than battle, but there are differences in pace and scale and statistical reliablity (the outcome of a large-scale war is more deterministic than the outcome of a small-scale battle).
  • Both Chess and Go are Too Hard. Go may be in some sense more complex than chess, but both games defy human comprehension and nobody would say that Garry Kasparov must be dumber than Honinbo becaue Honinbo plays a harder game.
  • It is often said that go must be more complex than chess because computers master chess but not go. There is indeed something to that, programming go is very hard, but not as much as you think. Research in artificial intelligence began with chess; von Neumann himself built a chess computer (with a small board and only a few pieces) in the process of more or less inventing computer science as the first electronic computers were built. The World Chess Champion Emanuel Lasker was a mathematcian who tried to invent Game Theory before von Neumann, and another world champion Mikhail Botvinnik was an electrical engineer who built a chess computer shortly afterwards. Chess has been integral to the study of artificial intelligence from before the beginning. Contrast this to the situation in Asia; China and Japan were both very busy socioeconomically after World War II. While the west could invent chess computers the East was rebuilding it's commerce and indusdtry. Now, sure, NEC has as much money to throw at Go as IBM has to throw at Chess, but it's decades later. Meanwhile, I could give about 15 stones (an absurd handicap, rarely even used to teach children) to Many Faces of Go in the early 90's; now, I can give maybe 6 stones to the best "robots" at KGS. That's an improvement of about 4 standard deviations, similar to chess computers going from about 1300 before Ken Thompson to about 2100 with his computer Belle in the 80's. Go is just behind in this chronologically. We can enjoy our anthropocentric superiority in go for a another decade maybe, the way chess players did in the 70's.
  • NB: I may have found a source for some of this confusion in the discussion of Go Rankings. The Standard Deviation I'm talking about is "sigma" of the Normal distribution of ranked players; that is, 16% of all players are higher ranked than one standard deviation above average, by definition. But for example this PDF from the AGA uses "sigma" to refer to the variability in the probability of one player beating another, based on rating; which is a different thing.

Pete St.John 18:24, 23 March 2007 (UTC)

[edit] I pity the fool who reads this article

There is no universally applied system. The means of awarding each of those ranks and the corresponding levels of strength vary from country to country and among online go servers.

Okay. But are they really so divergent that so little can be said about them in general? Are there no major systems that Wikipedia could describe? At least for the higher levels? I mean, there are regional divergences in the Go rules as well. Wikipedia still did a half-decent job of explaining them.

Ranks between about 10k and 30k have very limited usefulness and meaning since there are few discernible differences in each level. It is not surprising to see a 19k defeat a 15k player, so these ranks are of little significance. They are mainly used in teaching, to mark learning progression.
The requirement of a "rank up" at this stage is very loose.

Sure, but then, they are stricter later? In what way?

The progression from DDK to SDK is a significant turning-point for learners, as in one saying "you are not a Go player if you cannot attain a SDK".

Okay, and why is it significant? What is a bloody 9k anyway?

This page hints at the existence of rank systems, has a really unhelpful rank table down the right side and lots of bizarre statistical tables, and runs a severe risk exploding the heads of our dear readers. Regards, --anon 192.75.48.150 16:15, 28 March 2007 (UTC)

  • I pity the fools trying to write this article :-) Go is like Polo, which gives handicaps by goals, and ranks the players with those goals; if a professional is a "3 Goal" it means his team must give 3 goals to the opposing side in a handicap match. Adding up the goals of all the players determines the net effect. In Go, we give moves; a "three stone handicap" means essentially that the weaker player gets to make 3 moves in a row at the start of the game (usually these moves are placed at standard grid points). This is used to rank players, so a 5 Dan gives a 2 stone handicap to a 3 Dan. In Japan, tradition is to start ranks at 9kyu (in Karate this is "White belt"), with the numbers going down as experience increases, up to 1 kyu ("Brown Belt"). Then the first step of "experienced" practitioner, 1 dan ("First Degree Black Belt"), counting up to 9 dan ("9th Degree Black Belt", usually an elderly and much honored teacher in Karate). So a 3 dan in Go would give 4 stones to a 2 kyu, who in turn would give 5 stones to a 7 kyu (the arithmetic across the 1k/1d boundary can be confusing because there is no "zero" kyu). In Karate, ranks are not given to people below 9k, as it is assumed they can be caught up to that level before taking any rank tests; this leads to the idea that go players may as well get up to 9k before playing ranked games, but it's not so easy. Matching this traditional, and in Go very practical, ranking system to modern rating systems (particlulary Elo) is not so easy. Pete St.John 18:28, 28 March 2007 (UTC)
Retrieved from "http://en.wikipedia.org../../../g/o/_/Talk%7EGo_ranks_and_ratings_24e9.html"