Sedol's wedge in game 4 against AlphaGo

[Solomon]

Actually, I think I'm being too disrespectful to Lee Sedol here to suggest that the move is a bluff or a trick of some sort. I think everyone agrees that it is, at least, the best move. When one plays hamete or trick plays in the early game, they are not seeking to play the best move, but simply trying to trick the opponent. This 78 does not give off such an intention. It just happened to be that this move is the one that is most complicated and most likely to cause AlphaGo to make a mistake. At the same time, it is really the best path for Lee Sedol anyways, so I think that explains why the move can be considered "divine". But I still stand by the fact that L10 is stronger than K10, and that if AlphaGo played L10 instead, the result would have been worse for Lee Sedol.

[Solomon]

I've heard that Gu Li and Ke Jie both found variations that seemed favorable for Lee Sedol. Is this true and if so does anyone know what these are? If not, do they agree with Kim Myungwan's analysis?

I don't think this is true; here are video snippets / marks on both of their streams where they go over the move but don't go over L10:

Gu Li: https://www.youtube.com/watch?v=xgQddtHhrJk
Ke Jie (2h 42m 3s mark): https://youtu.be/3InZ3j0MgPo?t=2h42m3s

[Kirby]

But I still stand by the fact that L10 is stronger than K10, and that if AlphaGo played L10 instead, the result would have been worse for Lee Sedol.

I agree. But I think that this is how Go works, sometimes - even with an AI as strong as AlphaGo: You win the game by your opponent's mistakes.

I don't think Lee Sedol was trying to play a trick move as much as that he was trying to play the best he could come up with in the position that he was in. And in this case, it was good enough to lead to a blunder by AlphaGo.

If AlphaGo had played at L10 in response, it would have probably been a much harder game for Lee Sedol. But you can say the same for any mistakes that happened earlier in the game by Lee Sedol - had Lee Sedol played differently earlier, perhaps he would have been in a better situation before this point in the game. It would be somewhat more satisfying if there weren't a stronger refutation to Lee Sedol's move, but if that were the case, there would be an earlier point in time we could identify where the computer had made a mistake.

The satisfaction I get from this situation is that it shows that Go is a complicated game. And even if you have a distributed network of computers that have been trained against 30 million moves, having trained itself by playing against itself thousands of times... it's still not immune to mistakes that humans can identify.

Somehow, that's a bit comforting to me - at least until they come out with the next version of the software

[hyperpape]

Worth posting on here that An stated explicitly that he didn't see L10 while observing. So it seems that it's a hard position for professionals to read.

[Cassandra]

The satisfaction I get from this situation is that it shows that Go is a complicated game. And even if you have a distributed network of computers that have been trained against 30 million moves, having trained itself by playing against itself thousands of times... it's still not immune to mistakes that humans can identify.

Both (!) sides did what had been usual in the Four Go Houses during the Edo period in Japan:

Study the game in secret, in order to find a "new" move, which should work according to your own judgment, and which your opponent has never seen before.

[Sennahoj]

Both (!) sides did what had been usual in the Four Go Houses during the Edo period in Japan:

Study the game in secret, in order to find a "new" move, which should work according to your own judgment, and which your opponent has never seen before.

what??

[Uberdude]

In an interview Demis said that AlphaGo thought (so many anthropomorphisms!) the chance of white playing the L11 wedge was less than 1 in 10,000 (probability from policy network?), so it won't have been in many of its play-outs and evaluations before Lee played it (I believe MCTS can re-use information from analysis on previous moves in the game, so there won't have been as much already existing analysis for this unexpected move). Something I noticed when watching the stream is AlphaGo/Aja seemed to play k10 even more quickly than normal, does anyone have precise move timings? When a human faces an unexpected move they usually spend extra time to see if there was something they missed, but I wonder if here by being an unexpected move there was very little existing analysis of continuations from L11 so when it then spent 30 seconds or whatever on it that was essentially the only reading of it, whereas for other moves it has expected you get the 30 seconds of clock time plus this huge backlog of existing analysis from previous moves, and consequently was much weaker reading than normal.

P.S. I do find it odd Younggil didn't see L10, to me atari on top was my first instinct, then from below.

[Cassandra]

Quite apparently, AlphaGo does not have a face-washing mode ...

[hyperpape]

P.S. I do find it odd Younggil didn't see L10, to me atari on top was my first instinct, then from below.

Certainly mine too (beginners play atari!). Maybe it means that he dismissed it?

[yoyoma]

In an interview Demis said that AlphaGo thought (so many anthropomorphisms!) the chance of white playing the L11 wedge was less than 1 in 10,000 (probability from policy network?), so it won't have been in many of its play-outs and evaluations before Lee played it (I believe MCTS can re-use information from analysis on previous moves in the game, so there won't have been as much already existing analysis for this unexpected move). Something I noticed when watching the stream is AlphaGo/Aja seemed to play k10 even more quickly than normal, does anyone have precise move timings? When a human faces an unexpected move they usually spend extra time to see if there was something they missed, but I wonder if here by being an unexpected move there was very little existing analysis of continuations from L11 so when it then spent 30 seconds or whatever on it that was essentially the only reading of it, whereas for other moves it has expected you get the 30 seconds of clock time plus this huge backlog of existing analysis from previous moves, and consequently was much weaker reading than normal.

P.S. I do find it odd Younggil didn't see L10, to me atari on top was my first instinct, then from below.

Chart of time taken: http://imgur.com/3HcJKbk AlphaGo took it's usual ~1 minute to answer the wedge. Also from Hassabis' tweet during the match:

Mistake was on move 79, but #AlphaGo only came to that realisation on around move 87

BTW after this tweet during the press conference he changed his wording to emphasize the brilliance of Lee's wedge 78, and saying that AlphaGo got confused. I haven't seen calling AlphaGo's 79 a mistake since then. I think this is (good) PR spin by Hassabis who has consistently praised Lee.

Hassabis also said AlphaGo's evaluation didn't drop until move 87. So it's very unlikely that simply thinking harder on move 79 would have helped, the position was too unclear too deep into the tree for AlphaGo to read. OTOH it does seem to show that AlphaGo doesn't increase it's thinking time just because the opponent makes an unexpected move. From the time chart you can see it increased it's thinking time for move 87. That would be consistent with it thinking longer on turns when the evaluation of the position changes suddenly.

[belikewater]

BTW after this tweet during the press conference he changed his wording to emphasize the brilliance of Lee's wedge 78, and saying that AlphaGo got confused. I haven't seen calling AlphaGo's 79 a mistake since then. I think this is (good) PR spin by Hassabis who has consistently praised Lee.

Hassabis also said AlphaGo's evaluation didn't drop until move 87. So it's very unlikely that simply thinking harder on move 79 would have helped, the position was too unclear too deep into the tree for AlphaGo to read. OTOH it does seem to show that AlphaGo doesn't increase it's thinking time just because the opponent makes an unexpected move. From the time chart you can see it increased it's thinking time for move 87. That would be consistent with it thinking longer on turns when the evaluation of the position changes suddenly.

If I remember correctly, Hassabis in the press conference said that AlphaGo was pressured into making a mistake. I found it odd that he would say it was pressured. One of the strengths of the program is that it is not pressured, but it is also its weakness, in that if Lee played the same moves as White again, the computer would likely repeat all the same moves as Black and make the same mistake.

[RobertJasiek]

it's very unlikely that simply thinking harder on move 79 would have helped, the position was too unclear too deep into the tree for AlphaGo to read.

AlphaGo does not read and this is the reason why it lost at this moment: it did not verify by reading. The AlphaGo team is too focused on neural nets.

[Uberdude]

AlphaGo does not read and this is the reason why it lost at this moment: it did not verify by reading. The AlphaGo team is too focused on neural nets.

Wrong. It uses MCTS for reading.

[RobertJasiek]

MCTS is NOT reading. It is incomplete sampling with "arbitrary" gaps and without complete decision-making down the tree.

[Tactical] reading means functionally complete tree exploration and decision-making down the tree. (Obvious failures and obviously inferior moves may be ignored but all interesting or unclear moves must be read unless pruned by the reading method / principles.)

EDITED

[Uberdude]

By your definition then I and other humans don't read either, as our reading is not perfect or "functionally complete tree exploration and decision-making down the tree". Perhaps you meant AlphaGo did not read correctly, that is different from AlphaGo does not read.