Strange, I can read all the posts there

It's already been posted in the Computer Go sub-forum, but for the benefit of those who don't check there, DeepMind have today released their commentary (by Fan Hui supported by Gu Li and Zhou Ruiyang) on the 5 Lee Sedol games plus 3 AlphaGo self-play games.
https://deepmind.com/research/alphago/a ... s-english/

Have anyone took a look at those games where AlphaGo plays against itself? They are truly remarkable games. Looks like it has a totally different kind of consciousness or something. No human could play like that.

Yes, I have. What's so remarkable about them? I've seen plenty of 8ks play like that.

(Not saying AlphaGo is 8k, simply that I'm not anywhere near strong enough to really appreciate it)

Well I don't know any top level human pro players who play moves that look like they were played by 8kyu. And when AlphaGo plays them, they are not just random 8kyu moves but they really work as the commentaries show. That's quite remarkable as far as I can see.

We used to play a rank guessing game on KGS, where a player would find a random game, play 50-100 moves, and have the other people guess what rank the players were. Many times people guessed pro games were 5k or 8k or 1-2d. It's amazing how bad we are at guessing ranks, even relatively more experienced players who are stronger and played the guessing game quite often.

That reminds me, user 'maf' made a nice site where you can play this guess the players' rank game:
https://kyudan.net/

The problem with that site is you should always just guess 5d. They should give you a set of nine games each one representing one rank from 1 to 9 dan.

Eh? Guessing 5d as the mid-point is a good strategy if you have no skill in evaluating the strength of the players (and assuming the distribution of games in the database is uniform across ranks 1d - 9d), but if you do have some skill in that then guessing what you think will have a lower variance from the correct answer that always guessing 5d.

I always was terrible in math and school in general, but to me it seems like this: from a 5d point of view, 1d is 4 wrong and 2d is 3 wrong and so on. 20 in total. 20/9= 2,22. Uberdude's average offset is 3,86.

Check the user mr5d I just made using your strategy, ranked below Uberdude on the 95% confidence interval offset that sorts the leaderboard. Also Uberdude was quite a lot better on the first dozen or so guesses when I spent a long time trying to judge the level but then my score went down when I started doing it quickly. Maybe I'll make a new serious-guess-only one if I have time later and try to get a good score.

Edit: Sorry, after typing above did one wrong guess on mr5d in new tab without logging out (with 5 out and 0 out, got my colours reversed!) before that the 95% was 4.86 iirc

I also tried this and got to 3,57 average offset. I would like to know what was wrong with my math.

Looks like the "average offset" is the mean of the total offset in each pair of guesses. So if you guess 5d 5d and it is 2d 1d then total offset is 3 + 4 = 7.

the best pair is 4d.

I didn't check your math, so there could be multiple points of failure, but one thing you didn't take into account of is the number of games they have on record for each rank. They have a significantly fewer number of 9d games than 4d games, for example.

They said something about it being equal probability anyway.

Does anyone have either a SGF or a link to the recorded video commentary that Gu Li and Zhou Ruiyang did live during the Lee Sedol match?