Game 4 of the Judan was today. Here is a link to the complete video coverage of the game by the Nihon ki-in: https://www.youtube.com/watch?v=VDWwgXcSQLo.

And here is the game (transcription mine, hopefully without mistakes):



The game was quite nerve wracking and I enjoyed it a lot. It seems to me that Murakawa was often under a lot of pressure but that Iyama never really managed to make him pay for that, having problems of his own (most notably, managing the top right after losing the ko). Some sharp moves by Murakawa at 86, 96 and 116.
I also like the fact that the opening featured two 3-3 points and no hoshi. It's quite refreshing to see something else than the bots openings these days!

And if I am not mistaken, that makes the score 3-1 for Murakawa, which means he is the new Judan!

1st game of the 74th Honinbo, Kono Rin challenging Iyama is underway:

https://home.yikeweiqi.com/#/live/room/18001/1/15184532

Giving the ponnuki to capture the top stones in gote didn't look great to me, and the Yike bot thinks it's awful. At least play the n10 atari, it's not like Iyama to be wimpy. Also looks like he made a small endgame mistake in not playing t10 atari before r7 (makes t7 yose a little bigger). Is that something Cho Chikun would have found? Reminds me of a nice move Cho played (for good endgame anticipating stones which would be captured) that I think explorebaduk posted about but I can't find it now.

Kono started his Honinbo title challenge well, won the 1st game by resignation:



I added just a few variations at the end, from the ones found on the WBaduk server from the pro commentary, showing why Iyama had to resign.

_________________
Sorin - 361points.com

Game 2 is tomorrow, yike broadcast at https://home.yikeweiqi.com/#/live/room/18328/1/15630642.

Update: Kono won again! Is this the end of the Iyama era?

Indeed, it seems surprising for Iyama to have a 0-2 start in a major title!

I added the game record with some variations found on WBaduk.
In particular, I liked a lot the variation for black 147, which results in a spectacular exchange in black's favor.

LeelaZero seems to think that Iyama's losing move is 161 (5 moves before he resigned) - while if I understand correctly, the Japanese commentary seems to say that white was ahead anyway (I wonder if this is because of the 1 point difference in komi between LZ and the game rules).

_________________
Sorin - 361points.com

Hi sorin, at 120, one bot (maybe FineArt) gave 87% for ;
DeepLeela (much lower-powered) gave 60% for .
Very difficult for humans to follow-through like the bots.

LZ157 seems to disagree with the fact that white was ahead anyway, even if komi is changed to 6.5.
If Black plays at K5, LZ gives White 28% winrate.



If I add a white stone at K12, i.e. I make White play at K12 and Black passes, White's winrate becomes 38%.



I don't know how much worth is a white stone at K12, but certainly at least one point, so apparently LZ157 thinks that, at move 160, Black was leading by more than one point (but the lead wasn't huge, just a few points).

P.S. If, instead of K12, the extra white stone is placed at K13, then White's winrate becomes 49%.

Ed: move 120 is so far away from the end of the game, I had no illusion that one's advantage would hold for 40 moves. I am very surprised by the disagreement about who is winning after move 160 though, between the two camps - humans vs LZ

jlt: I realize my logical fallacy when I said that the disagreement in evaluating the board at move 160 can be due to 6.5 komi (Japanese rule) vs 7.5 komi (LZ), since if black had to give 7.5 komi instead, black would be less likely to win; but even so, LZ favors black. My LZ version is version 220 or so, and it seems in agreement with your 157.

So it looks like Iyama, in agreement with the pro doing the WBaduk commentary, thought that he is behind after white 160 and instead of the safe connection he tried to get extra points by playing risky, and his group died.

_________________
Sorin - 361points.com

We have to remember that winrate estimates are based upon players making mistakes. Otherwise they would be 0% or 100% and go might be less interesting. Furthermore, because of self play training, LZ's winrate estimates are based upon LZ's errors, which may be of a different kind than human errors. (In fact, we know that bots and humans make different kinds of errors. ) So top humans could be right that White is more likely to win the game, based upon human errors, while LZ is also right that Black is more likely to win the game, based upon LZ's errors.

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

Sure, if we are talking fuseki I wouldn't take the relative AI winrate estimates of some moves too literally, but here are are at move 160 and not a lot of open areas remaining, so it's the kind of position where top humans are normally confidently estimating the result within 1-2 points I would say. From what I understand they are confident that white is winning.

LZ is also pretty confidently estimating that black is ahead (70+%).

One of these two groups is wrong. Which one?

It is possible that LZ running on my computer is missing something about the upper-left seki, that is a likely possibility, so I would be curious if others can analyze the position at move 160. Or if some pro is willing to play against LZ from that position...

_________________
Sorin - 361points.com

Based on my experience, I think they are a bit overconfident.



Humans do fine with fuzzy evaluations.



One of them would be wrong if they were both talking about perfect play, which neither of them can do. To be fair, I don't think that either is claiming that White or Black will win with perfect play.



Well there are a couple of ways to test LZ's estimate. One is to have LZ play against itself a lot of times. If LZ estimates a Black winrate of 70%, that means that White should win around 30% of the time, and one of those games may be the best approximation of perfect play. Another is to explore the game tree using LZ to find possible errors that self play might not reveal. For instance, in Beat Elf! ( viewtopic.php?f=15&t=16641 ) Elf gave Black an 85% winrate 4 moves before Black resigned. OTOH, such an exploration may reveal assumptions that humans share that LZ refutes.

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

LZ see the upper-left seki, and actually, it's willing to let the black stones die.

After playing a couple LZ vs LZ starting from 161 K5, I guess LZ see four interesting area for black :

- it feels that the E5 group is harassable
- there's also a weakness around O5
- the Q15 stones are a liability for white.
- B11 is also big.

On the contrary, black groups are very stable

This is a very interesting question. I’m under the impression that zero style bots still do not play the endgame as well as top human players. If that’s still true, I wonder if this game is far enough along that this effect is relevant.

_________________
Occupy Babel!

This is counter-intuitive to me, my understanding is that the closer the game is to the end, the more accurate AG-style bot estimations should be, whether the bot was trained zero-style or initial-AG-style, since the outcome from which they do the actual learning (the result of the game) is fewer steps away.

_________________
Sorin - 361points.com

But -- and this may be a big but --, by the time that the average gain per play is less than 3 pts., top humans can play (almost) perfectly to the end, while bots still make probabilistic estimates.

One thing that bots seem not to do that humans do is to guard against the possibility of their own errors. For instance, if one bot leaves a group killable, even by a 10 kyu human, while it tries to gain a point or two somewhere else, its bot opponent does not usually seem to kill the group. If it has overlooked a good play that the opponent finds, that could result in a lost game. A human would just nail the win down.

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

Kono (black) doing well in Honinbo game 3, o8 seems to be the big direction mistake, https://home.yikeweiqi.com/#/live/room/18718/1/16155667. Here is first day's play. I wonder if Kono will sacrifice his stones for now or attempt to pull them out (Yike bot says sac is fine).

A convincing win for Kono, who kept his lead from the first day and cryised to victory. Iyama now faces kadoban.

Too soon?

Iyama won by half a point.

Cheers,
Vesa

Oops, I got confused by the Yike Chinese UI and a long pause! Here's the game:



So I think c12 could be the start of the giving Iyama chance to comeback, if instead he cuts at k13 then by attacking the top right white group (using peep at p14, though m18 means that's not actually sente) black should be able to turn the top left centre area to dame, whereas in the game after f15 Iyama made a nice box with his potentially weak group in yose which is always nice. The bot says with c12 it's still very confident of a black win, but bots can play out a close endgame accurately, whereas going for a more decisive pressing of advantage to reduce chance of a slack yose loss could be wiser for a human. Hard to balance the risks involved.

Sorry, at move 141, when Black plays c12, I cannot find a peep at p14. There is a peep at m18, but I can't follow what you are saying.



Well, they can play out a close endgame accurately, and, OC, they ordinarily do. But that depends upon accurate reading. At move 141 I kind of doubt that is the case, without very powerful hardware. Consider the last gainful play in this game, Black 289. The game record gives a variation where Black plays a 0 pt. sente first, and the bot gives Black a winrate estimate of 90%. OC, there is nothing wrong with the play, but even at this stage of the game the bot's evaluation sucks big time. That does not inspire confidence. (Not a komi mismatch problem, either, as White wins with a 7,5 komi, too.)

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