I ran an Elo ranking over all the games in the go4go database. This graph is still a work in progress (anyone good with R?) but here is the top players by Elo for the last 25 years:



The first thing I notice is how truly dominant Lee Changho was. Not just the best, but the clear best for 10 solid years. Another interesting point is that by Elo Iyama Yuta is the strongest player in the world right now. Here's the current top 10:

Iyama Yuta: 2788.9701869012265
Ke Jie: 2766.0026491949425
Park Jungwhan: 2755.3027742005374
Shi Yue: 2734.911306108628
Lee Sedol: 2721.9747143862674
Tuo Jiaxi: 2719.2818172904167
Mi Yuting: 2715.9730263812394
Jiang Weijie: 2713.4313622455284
Lian Xiao: 2702.344198033096
Kono Rin: 2701.814513146868

I decided to set every players' starting rank at 2400. This isn't ideal, especially for the earlier games where 9d players have their ranks set the same as 1d. I'm hoping by the time of these charts those problems have mostly gone away. The go4go database emphasizes newer games at the expense of not having as much of the older stuff.

There's still a lot more I want to experiment with, like highest rated games in the db, biggest upsets, predicting future tournaments, etc. I also want to keep tinkering with the graph to get it to look better. It's less visually confusing over a smaller time scale, but then you don't get the full progression.

Interesting, please keep it up. But I have a pet peeve -- do you really think the ratings are accurate to better than one part per trillion? On a more serious note, does your system assign a probable error to the ratings?

Definitely not, I was too lazy to format the output so you are seeing the raw doubles.



Nope, there are other ratings systems that do, like glicko, but I chose Elo because it's very easy to code, and very simple so everyone understands it.

Lee Changho's dominance is indeed striking, but something else I noticed was the generally upward trend in the ratings. Are pros on average getting stronger, is rank inflation inherent in ELO, or perhaps is your initial seeding at 2400 too low? What happens if you started everyone off at 2600?

I'm not sure starting everyone at 2400 is going to be accurate across regions - given the fairly limited number of international games I think this will skew the Elo in favour of top Japanese players since the average level of pro is somewhat weaker than Korean and Chinese (at least for newer pros where competition is much more fierce in China and Korea).

Yes, I noticed the ratings increasing too. Elo is a zero-sum rating system, so if the winner gains 12 the loser loses 12. The only way for points to enter/leave the system overall is when new players enter or old ones stop playing. My guess is that there are just way more players in the database now than in 1990, so there are more points to spread around.

I'm not sure if there's a way to adjust or account for that, I may tinker with it some. It might be worth it to try other rating systems too and see how they do.

Aren't them some versions of Elo with a little epsilon factor that allows a small amount of points to enter the system to reflect that people can get stronger without others getting weaker. The EGF ratings have this: http://www.europeangodatabase.eu/EGD/EG ... system.php.

I am still messing with the rating system (reading about Whole History Ratings which seems very interesting) but in the meantime I wanted to share the other reason I did this: using elo to decide which games to play through. Every week go4go emails me new games. These are the top three games last week by Elo, and the winning odds it predicts:

#1) 17th Chinese City League A, round 7
Park Jungwhan 56% -- 44% Jiang Weijie

#2) Japanese Honinbo, Title match #4
Iyama Yuta 65% -- 35% Yamashita Keigo

#3) 17th Chinese City League A, round 7
Mi Yuting 47% -- 53% Shi Yue

So I'm off to play through these games, the other nice thing is I can spit the output out spoiler-free, so I can replay the game and be surprised by the outcome. I've used sites like Mr Kin's Go News in the past, but every game is listed with its result, so you always know who won before you start.

I had a similar idea, and produced WHR ratings from go4go.net data: http://www.goratings.org/

I am getting a different ranking from yours, though.

Before trying WHR, I tested Glicko. With an incremental rating system, Iyama Yuta had a better rank, but was still a bit far from #1.

I believe incremental systems don't work well for such data, because Japanese players play mostly Japanese opponents. WHR is not incremental, and can deal very well with situations where two strongly connected groups of players interact rarely.

Rémi

It may be true that Japanese pros mostly play only other Japanese pros, but that must be true for pros from any country, they mostly play other pros from their own country. The number of pros who frequently play in international tournaments is very small.

I think Japan is certainly more isolated than China and Korea. Quite a few of the top Korean players play in the Chinese league which is 11 rounds a year. And picking two recent international tournaments there was the China-based MLily cup, which had 253 Chinese, 79 Korean, 23 Taiwanese, and 16 Japanese players in the prelimnary tournament (http://igokisen.web.fc2.com/wr/ml.html). And the Korea-based Samsung cup preliminary (http://igokisen.web.fc2.com/wr/sc.html) had 220 Korean, 87 Chinese, 25 Japanese, and 16 Taiwanese.

Yes, the point is if you start all pros with same elo, then Taiwan and Japan will get an equal ratings start which is obviously not true from results.

Remi your website is awesome, much prettier than I can do. Any chance you could make some comparative graphs like I tried to do for Elo?

Compare the ratings for Cho Chikun and Lee Changho.

Thanks for your suggestion. I probably won't have time to work on the web site very soon, but I add this feature to the todo list.

Remi, well done.

A small problem - your code does not handle games with 'Void' result correctly. For example, game ID 47707 is treated as B+R in your system. Should be easy to fix.

Thanks for noting this problem. It is now fixed.

I also optimized WHR parameters. The first version was using a parameter that I optimized for KGS data. But the volatility of pro. ratings should be much lower, and the prior stronger. So the new graphs are much smoother than before. This changes the ratings significantly:
- An Jungki was 16th, and is now 91, which looks more reasonable
- Kang Dongyun was 3rd and is now 7th. His 8-win streak did not make his rating grow as much as in the previous version, thanks to the lower volatility.