Life In 19x19

Long time ago, Josef Moudrik and me worked on an automatic Go style/strength estimator that takes a look at your game records and tries to estimate your style and strength based on a large body of games it used for training.

Now, Josef Moudrik has pushed the concept much further and created a shiny web app that allows you to get your games analyzed with no hassle! You can either upload a .zip with your SGF games or it will download your games from KGS if you supply your username (entering password not required). Try it out at: http://gostyle.j2m.cz/webapp.html

Example pages with analysis of the latest TheCaptain's and pachi's KGS games:

http://gostyle.j2m.cz/WAPP_RES/edb042f9 ... 04c13f47b/

http://gostyle.j2m.cz/WAPP_RES/840fc3ad ... e7b0631d3/

(they last two months; just enter their nicks to the form at the main page if they already expired)

Let us know what you think!

I'll give it a shot. I supplied my KGS username and it's been loading for quite a while now. Probably offline or crashed.

I looked at TheCaptain's stats that you gave. The Distinctive Patterns chart is hard to read. So 5x4, 3x4, tenuki, 3x3 attach joseki pattern is something strong players play but TheCaptain doesn't play? This pattern happens about every 14th game in professional play based on the red mark? What is it about strong players and weak players?

Anyway, very nice idea. I'm excited to find out what it says about my style.

Edit:
Took a couple of tries, but now the results are in:
http://gostyle.j2m.cz/WAPP_RES/388e6238 ... 9231ae01c/
Rank estimate at 2d±2.63 is about 1 ¼ stones too high. I've been hovering at 1k, occasionally dipping to 2k and peaking at 1d. I agree with style evaluation. My games are often full of fighting. I'm pretty bad at shinogi too.

Awesome idea! This is pretty cool.

I tried uploading a SGF. It said 'pending' for about half an hour. How long should it take?

Seems a little generous in its rank estimation, calling me a 3d (when the account hovers between 1k and 1d). The results are a little vague and feel a bit meaningless. What does 5.5 mean on the scale of territory vs. moyo? It suggested one move I should consider playing, a one space jump, but it meant nothing to me, as it lacked context, i.e. it lacked examples from my games where I could/should have played this.

A 3k friend did the test, it ranked him a 1d, and told him to stop playing empty triangles. That was perhaps a little more concrete, but again, without context it's hard to know whether this advice is meaningful.

I understand this extremely localized pattern matching is fundamental to the technique, which I presume comes from go engine research, but it does feel like there is some way to go to make this meaningful or instructive to humans. Still, I look forward to further improvements.

Pasky,

This is very interesting read
"On Move Pattern Trends in a Large Go Games Corpus"

http://arxiv.org/pdf/1209.5251.pdf

This is a fun idea. I did the test by entering my kgs account name, and within a minute or so the results were posted. I later entered another one of my accounts to compare, and although it showed different patterns, the judgements about my preferences were very close. I really like the idea of showing typical patterns and correlating them to strong or weak - but it doesn't seem particularly informative at this stage. Although the app correctly picked my rank, it seemed to think that I had a tendency to make strong moves.

I downloaded my username from KGS and got 1 dan, which seems high.

One thing I noticed looking at the list of games is that it treats demonstration games, including reviews, as separate games to be analyzed. Perhaps this should only look at rated or free games?

Inkwolf,Toge: Yep, it was down for a half day, now it is up again. It should take (once the stautus changes to working) less than a minute..

quantumf: Of course, it might be off, especially when the user submits a small number of games (but even without it, the method has a certain error). The empty triangles and other advices have even bigger error and I am trying to tweak the process to be relevant. The problem here of course is, that the method correctly learns that weak players play the empty triangle a lot, while strong players generally do not. But, sometimes, they do and when they do play it in larger number in the sample submitted, the program warns them on thing they already surely know. The 5.5 on territory <-> moyo scale marks what compromise between territory-based game style and influence-based style the player makes. Similarly to empty triangle, the one-point jump basically says that "better players play this more" and yes, more context is needed - this statistics is basically a byproduct (which I considered interesting though) of the strength estimation.

Anyway, it is not just patterns, the program also analyzes the awareness of "sente/gote" (for a certain definition of these terms), high/low plays throughout the game, and some other "features".

skydyr: Hmm, yes it should disregard the demonstration and review games, as it does with handicap and other board sizes, I will look into this, thanks! I've looked on your output and I think the cause for the mis-guess in strength is the fact that the games (not much of them anyway) are relatively sparsely scattered over a long time during which your strength changed.

I got 1k, and I was about to say that it's too high, but then I realized my KGS ranking has actually been floating around there recently. It's even 1d? or 2d? now. Whether that's my real rank (I haven't been playing much) or not, that's where KGS has had me.

As for the style, I think it's about right--it pegged me as territorial, orthodox, just slightly above average in fightiness, and evenly split between safety and shinogi.

JoeM wrote:skydyr: Hmm, yes it should disregard the demonstration and review games, as it does with handicap and other board sizes, I will look into this, thanks! I've looked on your output and I think the cause for the mis-guess in strength is the fact that the games (not much of them anyway) are relatively sparsely scattered over a long time during which your strength changed.

Yeah, that doesn't surprise me too much. I don't actually play on KGS particularly frequently, to say the least.

I tested it 3 times with:
1. 40 Tygem games
2. 40 KGS games (by entering my userid)
3. 40 IGS games

Of course for IGS and Tygem I used my latest games.
With my tygem games it estimated me as KGS 2k (pretty accurate)
With KGS: 3k (again very accurate)
With my IGS games (at 5k range) it estimated me as 3d (needless to say that's very wrong).

The estimated playing style of all of my games was pretty much the same and very accurate!
The recommended professional players were always different except of two playerse which was shown in 2 of the 3 trys.

As problems it pointed out:
1. Tygem: Sente and Gote
2. KGS: Sente and Gote, Empty triangle, likes to capture stones (the last one is very true )
3. IGS: Nothing (that's not a surprise as 3d )

The diagrams at the end don't really tell me much.

Now my thoughts: I really like it! It's a great and interesting idea. I always knew I play territarial but seeing a number on how much (8 in my case) is really interesting.
The other areas of playing style are also very helpful.
The recommedations of professional players that have similar style and whose style is very different is very cool.

I would love to see an offline/downloadable version, of course new information must still be exchange online but that way you won't need to upload all games. And maybe can even use more than 40 games.
Even if that isn't possible, I would still like to see this being improved.

Now here is something I didn't like: It told me that I like to play empty triangles, I would have liked to get a list of the games where I did play it. Without context it's hard to take it seriously.

I'm not entirely clear on what is meant by the Orthodoxity axis. Could you explain it?

Cool, tried again with 8 recent SGF files. They estimated me at 10k KGS...which is, in fact, my current rank.

No surprises on style...low in fighting and innovation, high on territory, medium on thickness. I will check out some of those suggested pros to study! Thanks for the fun new tool.

BYW, this seems like an interesting way to scout out your opponents on KGS, too...

I'm 7k, I guess. And I like to capture stones.