Revised European go ratings

[Javaness2]

Yes to both

[Pio2001]

Thanks !

[Uberdude]

Is it true that the bottom rank is 20 kyu ? (the bottom rank in France is 30 kyu)

I've heard a reason for this is 20k+ players have a high variability and tend to improve fast, so if you put them in a rating system you get poor quality ratings (high uncertainty and quickly out of date). However, I've also heard a con in that it could alienate the weaker players, who feel like the ratings system doesn't think of them as real Go players and they don't get the motivation of improving their rating through tournament play. If a rating system could go down to 30k but have a confidence parameter (like Glicko rather than Elo?) and a liberal reset policy that'd seem a good solution to me.

[hyperpape]

AGA has ratings that go arbitrarily far down.

[Pio2001]

It is true that the ranks below 20 kyu seem quite useless.
But I have encountered once a situation similar to the one you describe : beginner players were coming in the annual tournament of Lyon (France, 70 players), and, as beginners, I registered them as 20 kyu, as I usually do (I consider people capable of playing an opening on a 19x19 board as 20 kyu).

But they were very frightened of being paired with too strong players, and not getting the right handicaps. In fact, as far as handicap is concerned, players below 20 kyu are considered as 20 kyu, so it would have changed nothing.

Another french tournament director (nickname Fenring on the http://go-on.forumactif.com/forum ) also advocates not to register beginners too high. He says that beginners have to bear with their lack of experience, the stress of their first tournament, and usually under-perform. No need to add an extra challenge on their shoulders on top of that.

Regarding the variability, the french system doesn't use any weighting for players below 20 kyu.
For example, if a 21 kuy plays a 15 kyu with 6 stones, the level variation is weighted for the 15 kyu player (because of the handicap), but not for the 21 kyu (full points are awarded).
The coefficients for fast games, 9x9 games or 13x13 games are also not applied. They get the full variation for each game.

[gennan]

In my kids club I use ranks down to 40k and Kalli Balduin (go teacher with currently 600 pupils in the Berlin area) even uses ranks down to 50k for his kids clubs. In my experience ranks below 20k are perfectly valid ranks when they are supported by handicap differences against stronger players.

So the revised system does not use a lower bound (for example, see http://goratings.eu/Home/History?PIN=16949867). With a liberal reset policy, not having a lower bound is perfectly fine as long as Go associations allow beginners to declare their rank freely in tournaments. When this is the case, the rating system doesn't need a special confidence factor for beginners (like a very large K factor).

[gennan]

Do handicap games count as much as games without handicap ? (in France, they are weighted with a coefficient equal to 1-H/10, and if White looses, her variation is again multiplied by 1-H/10).

The EGD uses a weight of 1 for handicaps up to 9 stones (for both players). For higher handicaps it uses a weight of 0 (for both players).

The revised system uses a weight of 1 for all handicaps. The reason is that I think handicap games help to maintain long range calibration. Reducing their weights would throw away valuable information.

I did include a mechanism in the revised system to reverse engineer ranks below 20k from apparently overhandicapped games against '20k' players. But still, most of the information about ranks below 20k is just lost, I'm afraid.

[gennan]

BTW, using the revised system's beta-function, I added a mapping to the Elo ratings used in DeepMind's papers (see today's Addendum on http://goratings.eu/Home/About)
According to this mapping, the different AlphaGo versions would have the following estimated ranks in the revised rating system:

AlphaGo Fan 7p
AlphaGo Lee 12p
AlphaGo Master 18p
AlphaGo Zero 20p

[Pio2001]

So, is your revised system complete ? Can we start to use it ?

[gennan]

So, is your revised system complete ? Can we start to use it ?

I think the EGD is basically doing a decent job, so it never was my intention that goratings.eu would replace europeangodatabase.eu (I have no desire to become a system administrator of anything). I only made the revised system to test the validity of EGD ratings and to find potential improvements. I do think the revised ratings are closer to "the truth".

The revised ratings are getting stale, because it hasn't been updated with tournament data since 2017-08-19. I suppose that keeping it updated would not significantly change my conclusions, but it would be nicer to keep the revised ratings current. So I am considering to add a mechanism to update it automatically by scraping newer tournament data from the EGD.

Then again, I talked to several people about my little project (not just on this forum) and I get the impression that very few people care about improving the rating system, so perhaps I should just not bother and abandon it.

[Pio2001]

That would be a shame.

Maybe what we need is to translate it into directly applicable guidelines :

New rules (auto-reevaluation, bottom rank etc)
A summary of the mechanism (to replace the page that describes the current system)
The list of changes from the previous system (the new values of "a", "con" and epsilon, etc)

And maybe the various national lists could then be joined into a unique european list. For example maintaining both the ratings of french players and european players means to have the work done twice.

I just discovered that I was currently 12 kyu in France, but with a 13 kyu rating and an 11 kyu ranking in Europe.

[gennan]

That would be a shame.
Maybe what we need is to translate it into directly applicable guidelines :

New rules (auto-reevaluation, bottom rank etc)
A summary of the mechanism (to replace the page that describes the current system)
The list of changes from the previous system (the new values of "a", "con" and epsilon, etc)

I kind of did that, but I guess the details are distributed over several pages and some details are only explained in this forum. Some editing could make it clearer, though. I could also publish the source code of the core of the system to remove any ambiguity about the details (if you can read C# source code). The core of the rating system is really quite small (I think less than 100 lines). I could even publish source code to reproduce the EGD ratings. The core of the EGD rating system may be a bit bigger, because it has more special rules, but not much.

And maybe the various national lists could then be joined into a unique european list. For example maintaining both the ratings of french players and european players means to have the work done twice.

I just discovered that I was currently 12 kyu in France, but with a 13 kyu rating and an 11 kyu ranking in Europe.

I never saw the French rating list before, but I found it just now: http://ffg.jeudego.org/echelle/echelle_niveau.php
These are the French ratings according to the EGD: http://www.europeangodatabase.eu/EGD/cr ... dgob=false
These are the revised French ratings from 2017-08-19: http://goratings.eu/RatingList?country=FR
I suppose these different systems should agree with each other fairly well, if these lists if all are well calibrated and if all systems are fed with the same tournament data.
For your rank, the systems seem to be in agreement within a variation of about one rank, which is not bad. I suppose 13k is your EGF rank and 11k is your revised rank?

So what you suggest already exists, I think. I don't know which national go associations maintain their own rating system and why. Before I started my project I knew Belgium had their own rating system. Only on this forum I found that France also maintains their own rating system (which is almost the same as the EGD since 2013).
I think most countries don't have their own rating system, but some countries regulate ranks. For example, the Netherlands don't have a separate system, but we do have a classification committee (currently lead by HermanHiddema) that processes tournament results to regulate dan ranks. I know some countries regulate kyu ranks as well and other countries don't regulate ranks at all.

I don't know why the FFG maintains their own rating list. If they submit al results to the EGD, they could just use the EGD rating list. But I suppose the FFG wants to apply their own rules, which are somewhat different from the EGD system (like a lower bound of 30k, which modification has been requested to the EGF by the FFG, but it has not been granted yet).

Perhaps it's difficult to make one european rating system that complies to the different wishes of each national go association. Even if it were possible, I guess some national go associations will still want to use their own system because they just prefer to keep some sovereignty (even though it takes effort to maintain their own system).

[Javaness2]

I think that the rating database deserves some improvements, so I hope your work isn't discarded. Isn't there a committee that is supposed to be investigating such projects?

The faux argument is often made that allowing ranks below 20 kyu will introduce inaccuracies into the system. You just have to look at the rating distribution to scoff at that - players are competing at 30 kyu in tournaments, and they are recorded instead as having competed at 20 kyu. There is no genuine problem to extending the system.

[gennan]

I think that the rating database deserves some improvements, so I hope your work isn't discarded. Isn't there a committee that is supposed to be investigating such projects?

The faux argument is often made that allowing ranks below 20 kyu will introduce inaccuracies into the system. You just have to look at the rating distribution to scoff at that - players are competing at 30 kyu in tournaments, and they are recorded instead as having competed at 20 kyu. There is no genuine problem to extending the system.

From my emails with Aldo Podavini (the EGD manager since 2009) and reading the reports of the Annual General Meeting (AGM) of the EGF (see Minutes of Annual General Meetings at https://www.eurogofed.org/egf/) and a summary at https://www.eurogofed.org/egf/commissions.htm stating The AGM of 26th July 2012 agreed to reduce the number of EGF commissions [...] The Ratings Commission was suspended for review., it seems such a committee does not exist anymore. And from my emails with Aldo, I get the impression that the mathematical details of the rating calculations are not really his concern.

It seems this is the situation since about 2011: As the EGD manager, Aldo Podavini validates tournament result files, he reports annual participation trends to the EGF board and AGM meetings and he implements changes decided by the EGF board. But since its inception by Aleš Cieply in 1996, the calculations haven't changed much AFAICT. I can only find the increase of epsilon (from 0.014 to 0.016) in 2009 and some changes in the rules for tournament classification (which determines the weight factor of each tournament).

So perhaps I'd have to lobby the EGF board directly to have them take my suggestions in consideration? I kind of doubt that general executives would concern themselves with the mathematical details of rating calculations.

Perhaps I should just go ahead and scrape updates from the EGD and have goratings.eu coexist with the EGD.

[gennan]

Interestingly, I just found this document by Aleš Cieply from 2001: http://gobase.org/studying/articles/myo ... eply.ps.gz. The format may be somewhat inconvenient, so I converted it to pdf and uploaded it here: http://goratings.eu/content/ales-cieply.pdf. (I think it's from 2001, because the modification data of the ps file inside is 2001-05-21)

Apparently, he did a similar statistical analysis of winning probabilities (with data from 1996 to 2001) as I did and he found the same discrepancy: see figure 2 on page 7 and compare it with http://goratings.eu/Probabilities (the dashed line in his figure is the black line in my figure, the solid line in his figure is the black line in my figure). He notes the discrepancy in the lower half of page 8, but he offers a different explanation in the upper half of page 9 (he thinks it's caused by the McMahon pairing system that most tournaments use), but he does offer a fitting function 1/a = λ => 7 / (3300 - rating), which is similar to what I used in an earlier version of the revised system. The EGD uses 1/a = λ => 20 / (4100 - rating).