This is interesting, but I’d be cautious about reading too much into the winning percentages without a baseline. What’s the percentage for black if black passes on the first move? Or if white passes the first move? If an early pass drops the win rate to 1%, then most of these “mistakes” are close to even. If, on the other hand, a pass drops the win rate to, say, 30%, then some of the mistakes Viktor cites are close to losing a move, which even as amateurs we can consider a problem.

I remember a lecture Jim Kerwin 1p gave some years ago about how to study pro games. He put up a position from a pro opening and asked who is ahead. People watching offered their opinions. Then he said that it’s about even and that, in fact, unless you have a really good reason to believe otherwise, you should assume pro moves are even. Even if you play like the losing player in a pro game, it’s still much better than you are probably playing now. (Of course, most of the audience was kyus and low dans.)

I still haven’t ruled out the possibility that AlphaGo’s opening is not actually better than that of top pros. With over 200 moves in many games, there is still a good chance its strengths lie elsewhere.

Still, I am happy to have this, because it inspires me to keep seeking the truth.

Thanks a lot for sharing this Marcel, this is very interesting! Indeed it may not matter a lot for us amateurs which move is one percent better than the other, but the fact that pros overlooked so many moves as "bad" is what is really amazing.

_________________
Sorin - 361points.com

I doubt pros are going to play any single moves that are as bad as passing (exclude blunders like Nakano's self atari!), though after 50 moves of a bad opening against AlphaGo I could well believe they are about a stone down. It's an interesting question what "win%" would it give for passing, but probably we can't know (unless I can persuade Aja or Lucas to ask!), though maybe we could ask another strong bot? I do wonder though how much confidence you can give to its answer though, as in its training it will have seem very few positions with one player passing! But pros are sensitive to tiny losses, for example in this very common opening when I was at BIBA On Sojin 7p said he considered the pincer of 6 a slight mistake (379 games in waltheri) and better to back off at a (1174 games) and black just taking the corner is fine for him (he assumed the normal solid connect joseki which AG disapproves of). AG agrees pincer is worse, by 2%. He thought pincer was ok if black's bottom right was at b because then if white gets sente after black takes the corner he has a good approach there, and you are trying to prevent black making nice mini-Chinese style formations whereas the 4-4 isn't threatening to make anything particularly good. With 3-4 AG thinks pincer is the same 2% worse, though now back off is not best move but wants to approach white immediately (as an aside this does seem to be a valuable general lesson from this tool, it often prefers to approach an open 3-4 over answering an approach to a 4-4, as was perhaps the thinking in older opening theory).

Click Here To Show Diagram Code

[go]$$B
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . a . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X b . . |
$$ | . . . . . 5 . 6 . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Browsing the WeiqiTV videos I do find the biggest total win% drop I've seen; this position black is at 23%. I've always thought that joseki is pretty ugly for black, but do recall being taught (maybe on China trip from Peter Liu 2p) that it's definitely bad if black's lower stone is at k4 (because h3 gap), but with L4 instead it's better (though maybe still bad overall). (AG doesn't give this line with k4). AG does also say o4 extend is better than o5 atari, which I seem to recall Go Seigen suggesting in his 21st century opening book.

Click Here To Show Diagram Code

[go]$$Bc
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . O X . . . |
$$ | . . O . . . . . . . . . . X X O O . . |
$$ | . . . . . . . . . . . . . X X X O . . |
$$ | . . . O . . . . X , . . X . X X X O . |
$$ | . . . . . X . . . . . . . X O X O . . |
$$ | . . . . . . . . . . . . . . O O . O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

P.S. I think a change in the underlying "win%" metric of half a percent is significant if a pro wants to play the best opening, but also the numbers we get in the tool have some noise of about half a percent.

We know that the cost of passing on the first move, in points, is double the fair komi, by symmetry considerations, so it's about 15 points. So if we take the distribution of normal game results (by points), and look at the area of a single tail that shows a player winning by 15+ points, that would tell us about what the win probability is after passing at the beginning of the game. (Many approximations were taken in the writing of this paragraph, I know.)

I'd be surprised if AlphaGo gave even as high as a 1% chance of winning after a first-move pass, given how sensitive its evaluations seem to be to slightly (by pro standards) suboptimal moves.

So I can't help but ask if Viktor himself stopped playing 5-5, since it's less favored by this tool.

Somewhat ambiguous. I would have thought that approaching a 3-4 was preferable to replying to a 6-4 approach to a 4-4, in older theory, as it prevents an enclosure. That is, the AG tool is a bit retro in this regard.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.