Life In 19x19

With the recent success in preliminaries, Sumire has broken into top 30 female players! She is now #28 globally and #4 in Japan (having passed Xie Yimin)!

She was actually top 30 at the beginning of the year, according to this.

In talking about ratings, I am way out of my field, but I do remember talking to Nathan Divinsky and Ray Keene about them. They produced a book (?Warriors of the Mind) which aimed to show who was the strongest chess player of all time. They backtracked Elo ratings by adding all known results.

As far as I can recall, they said there are, nevertheless, major problems inherent in the methodology. For example, you get different results if you take a player's whole history or his recent history. In addition, there was a special problem with young people who were improving fast.

I suppose there may also be problems with which games you include e.g. official only, or time-dependent. I infer this is especially crucial from the fact that chess produces separate ratings for classical, rapid, blitz and online games - and the results are indeed very different in each case.

The go rating list above is based on whole history, I believe. Does this disadvantage Sumire, a very fast improver? Does it favour, say, Rui Naiwei, who doesn't play much now and certainly doesn't win much now, but who has a huge and stupendous back catalogue she can call on?

In terms of very recent games, official or otherwise, Sumire seems to be ahead of both Xie Yimin and Suzuki Ayumi. She would, however, be behind Rina and Asama, but, strangely, also Mukai Chiaki.

The old question seems to apply: if it were not a rating list of go players but a list of surgeons and you had to undergo life-saving surgery, would you trust what the list tells you?

gazzawhite wrote:She was actually top 30 at the beginning of the year, according to this.

The historical ratings are not fixed. If she wins a few games now, all the historical ratings will get higher too. For example, in this post on 15th March she was only ranked #35 in the same beginning of the year ranking list. This makes interpretation of the historical snapshots a bit difficult.

It is a bit related to some of the problems John is highlighting in his post above. Players tend to become stronger or weaker over time. Goratings is trying to take that into account and provide an estimate of the ratings at different times. It's not easy because when a player defeats a stronger opponent, we cannot really tell if it was because of better luck or because they've actually gotten stronger. We will know better once we get more data (ie. when the player plays more games). Only then the estimates become more accurate and then goratings will also update its earlier estimates. That's why the historical ratings are fluctuating too.

Understanding the math behind goratings (or WHR) is beyond my paycheck, but more details here:
https://www.remi-coulom.fr/WHR/

According to the WHR description as used in goratings.org, all games played by all players are taken into account. So, if Rui Naiwei played and won a lot, and then nearly stopped playing, she doesn't get a lot of punishment. The usual is that an active player wins more games in its prime, then loses more often, and naturally, the number of loses taken into account increase compared to the wins. The paper explaining the algorithm doesn't say that old games are discarded. Actually, in KGS, if you stop playing, your ranking doesn't drop. It is the server that marks you as ?. Then, you are off the ratings. In goratings, I just checked, Cho Chikun has rating and games from 1971.

In the case of players who improve fast, I feel that the rating can't follow the player. Sumire is now over 3000, but not too long ago, she was under 3000, while other young players (boys) were over 3000, with not so good results. I think that, even though she plays and wins a lot, many of those wins are against weaker players, so her rating is dragged to that level. Eventually, her rating will improve.

pajaro wrote:According to the WHR description as used in goratings.org, all games played by all players are taken into account. So, if Rui Naiwei played and won a lot, and then nearly stopped playing, she doesn't get a lot of punishment. The usual is that an active player wins more games in its prime, then loses more often, and naturally, the number of loses taken into account increase compared to the wins. The paper explaining the algorithm doesn't say that old games are discarded. Actually, in KGS, if you stop playing, your ranking doesn't drop. It is the server that marks you as ?. Then, you are off the ratings. In goratings, I just checked, Cho Chikun has rating and games from 1971.

What you are saying is not the most accurate:
-Rui Naiwei didn't nearly stop playing. Infact she played 45 games in 2021. (Could be more. Goratings doesnt have all games). For comparison she has 9 games for 1999 (her peak year).
-Rui Naiwei's rating has gone down 183 points from its peak (which was in the end of 1999). Her global (all gender) ranking dropped from #21 --> #366.
-Her rating went down 32 points in 2021.



Here is a summary of her games in 2021: Her rating at the end of the year 2021 was 3119. Is it inflated? Maybe a bit? Hard to say how much because she won a bit more than she lost. But at least it is not as dramatically wrong as you could imagine, is it?

According to the WHR description as used in goratings.org, all games played by all players are taken into account.

This doesn't mean that all games are given an equal weight for all years. I mean, how else can you create a rating graph that reflects a player's strengh development in different years? Most likely games from 1990 have the biggest effect on the rating in 1990 and most likely they have a very small effect on the rating in 2020. And the other way around. Right?

Source for data: https://www.goratings.org/en/players/115.html

Edit. I re-read your post and realized that you probably used her as a hypothetical example. Sorry for misunderstanding. To find out how the rating system works in the scenario that you described, I tried to find a player who has nearly stopped playing and also became weaker, but couldn't find a good example. Do you have anyone in mind?

If there is an algorithm that actually tries to model what is happening with regards to »strength«, I think WHR is one that is by far closer than anything ELO-based.

However, I deeply suspect that there simply is not enough data (i. e. game results) by at least two orders of magnitude to make any such algorithm actually work to the displayed precision.

There are people here who seem to have a fascination with whether a player has gone up from 54th place to 51st in the ratings list. I suspect this is a tiny minority (maybe not of people here but of those in the wider world). There was a good example from that wider world yesterday. I am not a golf fan, but I read a piece about how excited golf fans are that Tiger Woods may play in the Augusta Masters. It made no bones about his slim chances of winning, but said the news had given golf a great fillip. In the course of the article, the world's current No. 1 was mentioned. That made no impression on me whatsoever, and I now can't remember his name at all or anything else about him. But I do know Tiger won lots of majors.

In similar vein we could have the fan test, of a different kind, in go. If you were offered a free fan signed by Iyama Yuta, holder of seven major titles and dominant force n Japanese go for over a decade or another by Pyeon Sang-il, who is far above him in the ratings list, which fan would you choose? I, and I believe the vast majority who have no desire to demonstrate sheer perversity, would choose Iyama. Hiw many go players have even heard of Pyeon?

Similarly, I would choose Sumire over Rui, even though Rui's calligraphy is much, much, much better.

Ratings fans talk about strength. They just mean a changeable number. It is safer to talk about significance. It is certainly more attuned to the wider public. Years ago, in Japan everyone except babies in cots, knew the name of shogi champion Oyama Meijin even if they didn't know how a knight moves in shogi. They didn't even known his given name - he was just Oyama Meijin. A bit later everyone knew the Honinbo, even if they didn't know he was actually called Takagawa, because he won it 9 years on the trot. Then everyone knew the full name of Sakata. He won so many different titles they didn't tag a title onto his name. He got his full name. Except he dodn't. Most people, reading it only in the newspapers, though he was (initially, at least) Sakata Hideo, a commoner reading for Eio.

None of this vast public knew, or cared, what rating numbers were attached to their names. It was their general significance that mattered.

Now that ratings proliferate, it's hard to escape them, but still there are various ways of looking at them. My feeling, bolstered by what I see in the chess world, is that what matters most, to most people, is the peak rating. For some, but only within the chess world itself, there are landmarks such as a 2800-player (but even that jus means an "elite player."

My question earlier, about whole-history ratings, is whether that way of looking at ratings obscures the significance of players (forget daily strength). It seems to me, prima facie, that it brings in troughs as well as peaks, and so flattens out a player's profile. Very wrongly, I feel. If you look at Yi Ch'ang-ho that way, he "only" won over 140 titles, including many world titles, and once had an annual score well over 80%. Greatest player ever? Where is he in the ratings list now? His form has been abysmal in recent years, and ratings lists "accurately" reflect that, but what if we measure his "significance" instead.

A quite different way of looking at ratings, and more "honest" n the baseball sense, is to say they measure "current strength". I don't know if they do, hence my earlier question. But, again at a prima facie level, I suspect current can be distorted by having too much weight put on historical results (e.g. Rui Naiwei, but even also Yi Ch'ang-ho for that matter).

I'm no statistician, but I do recall Disraeli's aphorism: there are lies, damned lies, and statistics. And in go, I tend to think there are lies, damned lies and rating lists.

The very title of this thread, and its length, points to the truth that many of us are interested in a mere 2-dan, Nakamura Sumire. Where is the thread on Pyeon Sang-il, NINE dan? Or even, Sin Chin-seo or Ke Jie, when we do have a thread on SEVEN CROWNS Iyama Yuta.

> How many go players have even heard of Pyeon?

Few people have heard of Pyeon Sang-il. Quite a lot know Byun Sangil.

Seven of Byun's games have been reviewed on Youtube (see https://senseis.xmp.net/?ProGameReviewsOnYoutube). That's vs thirteen of Iyama's and three of Sumire's. (Of course, the list is not complete.)

I agree with the point you're making, but Byun is not popularly obscure.

John Fairbairn wrote:My question earlier, about whole-history ratings, is whether that way of looking at ratings obscures the significance of players (forget daily strength). It seems to me, prima facie, that it brings in troughs as well as peaks, and so flattens out a player's profile. Very wrongly, I feel. If you look at Yi Ch'ang-ho that way, he "only" won over 140 titles, including many world titles, and once had an annual score well over 80%. Greatest player ever? Where is he in the ratings list now? His form has been abysmal in recent years, and ratings lists "accurately" reflect that, but what if we measure his "significance" instead.

A quite different way of looking at ratings, and more "honest" n the baseball sense, is to say they measure "current strength". I don't know if they do, hence my earlier question. But, again at a prima facie level, I suspect current can be distorted by having too much weight put on historical results (e.g. Rui Naiwei, but even also Yi Ch'ang-ho for that matter).

I think you have slightly misunderstood the WHR. It doesnt sum up the whole history into just one number. Not at all! It shows the history as history. Ie. it doesn’t just show the “current strength” but the strength at any single date in their whole career.

Take a look at these:
https://www.goratings.org/en/history/
https://www.goratings.org/en/ladies/
Lee Changho 16 years as #1.
Rui Naiwei 26 years as #1.

Can you see that the significance of their career there? Both of these are produced by the WHR. Their past accomplishments are not lost. They are there in WHR.

John Fairbairn wrote:But, again at a prima facie level, I suspect current can be distorted by having too much weight put on historical results (e.g. Rui Naiwei, but even also Yi Ch'ang-ho for that matter).

Do you have any evidence to support your claim? All the data from goratings seems to point to the opposite direction, ie that WHR doesn't seem to put too much weight on historical results when it calculates the current strength.

You can find a summary of Rui Naiwei's games in 2021 in my earlier post above. Is her current rating really that much off?

Here is a summary of Lee Changho in 2021: His rating at the end of the year was 3334. Is it accurate? Well, I cannot say it is completely accurate, but at least it doesnt seem to be very significantly off. At least not very badly distorted by historical games, is it?

I think you have slightly misunderstood the WHR.

I freely admit I may have probably massively understood WHR, but I also start from the premise that I don't like obsession with numbers. And I note with approval that the pro go world takes relatively little interest in ratings, especially in Japan.

I can see for myself that the likes of Yi Ch'ang-ho, Rui Naiwei, Cho Chikun are not achieving what they used to achieve, so to say to me that their current ranking is not much different from an older ranking goes in one ear and straight out the other. It has no meaning for me. I like to admire people for what they achieved and like to remember them as they were at their peak, not as Darby & Joans. The company of D & Js can be wonderful, but in a quite different way.

My intuition about WHR is the following:

Take a big physical table (you know, the wooden kind). Mark the years in one direction and »strength« in another (no need to put any numbers here, you can regard this as unitless).

Then for each player, take a somewhat flexible, springy rod. These tend to be straight, but you can bend them with force (and they will straighten when the force goes away). They are as long as each player's career. Put them on the table, roughly in the time direction.

Then for each game, take a slider: two rails (loser and winner) gliding against each other, but with a spring inside that applies a force when the loser rail is higher than the winner rail. Fix these to the rods at the time of the game, winner to winner rail, loser to loser rail.

When all these forces are balanced out, you get strength graphs for the players.

Of course, how it looks exactly depends on e. g. the springiness and the stiffness of the rods, and the force of the connector springs.

I do appreciate this attempt to get rid of numbers, but as far as I can visualise this contraption, the numbers still seem be lurking there - in their dozens, allegedly just like the lurkers on SL. So, in the end, it's mutton dressed as lamb. 'A' for effort and bonus points for creativity, nevertheless!

I prefer to rank players like one of those problems from grade school.

• Fan Tingyu played black and lost to Iyama Yuta.
• Shin Jinseo has black hair, wears black glasses, and beat Park Junghwan.
• Ke Jie has black hair, black glasses, played black, and beat Fan Tingyu with black hair and wearing black glasses.

What colors are Mi Yuting's hair and glasses?

Toukopouko wrote: Edit. I re-read your post and realized that you probably used her as a hypothetical example. Sorry for misunderstanding. To find out how the rating system works in the scenario that you described, I tried to find a player who has nearly stopped playing and also became weaker, but couldn't find a good example. Do you have anyone in mind?

I used Rui Naiwei as an example, because previously John had used her too. I don't follow her, and I just took a peek in her data to write the post.

About the use of old and new games in the algorithm, I'd like to explain something better:
from what I read in the paper, all games taken into account have the same weight in the result of the algorithm. goratings.org is based on that paper. But maybe the implementation is a bit different. I don't know this. I might even be wrong about what I understood in the paper. I used KGS to explain that they used the same algorithm, but also a way to remove games. In KGS, an inactive player is marked as ? after some time, and that player's games are out of the algorithm. This affects the result of the algorithm in other players (not only direct opponents). Maybe goratings.org keeps all games of every player (you can see it, I used Cho Chikun as an example of a long history), but perhaps only games newer than some time are taken into account when running the algorithm. Again, I don't know this kind of details, and again, sorry for any mistake in my English.