Find the best move (AI will be no help)

Gomoto · #1

Black to play in the lower left. LZ and Kata have got a blindspot in this case. Do you find it?

Click Here To Show Diagram Code: [go]$$B $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . , X . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O X . . . . , . . . . . , . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . X X O O . . . . . . . . . O . . . | $$ | . . . O X . . O . , . . . . . , . . . | $$ | . . . O . X . . . . . . . . . X O . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

TelegraphGo · #2

Though you claimed AI would be no help, I totally used AI assistance to find the best move. Else, how would I know which moves Kata and LZ weren't looking at? I wouldn't know whether my human analysis just found the computer move, or if there was still some mysterious slightly better one.

Before looking at AI, I gave myself as complete a list of candidates that might be the best move for the fight.

Click Here To Show Diagram Code: [go]$$B $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . , X . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O X . . . . , . . . . . , . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . X X O O B . . . . . . . . O . . . | $$ | . . B O X B B O . , . . . . . , . . . | $$ | . B B O B X B B . . . . . . . X O . . | $$ | . . . B B B B B . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Obviously, most of these moves look like bad shape, but there's usually something deep involved when AI's are wrong, so I kept my options very open. After checking to see which moves my AI was analyzing most deeply, and seemingly accurately, I had the remaining candidates:

Click Here To Show Diagram Code: [go]$$B $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . , X . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O X . . . . , . . . . . , . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . X X O O B . . . . . . . . O . . . | $$ | . . . O X B B O . , . . . . . , . . . | $$ | . B B O . X . . . . . . . . . X O . . | $$ | . . . . B . . B . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

I clicked into each of these moves with the AI, to see if it immediately caught itself. It didn't. So I looked at the suggested variations for each move, trying to see where there could be some deep reading that the AI missed. Most of these become simple broken/inefficient shapes for B when W plays as AI suggests, without much chance for a miracle. The most promising two moves were:

Click Here To Show Diagram Code: [go]$$B $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . , X . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O X . . . . , . . . . . , . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . X X O O B . . . . . . . . O . . . | $$ | . . . O X . . O . , . . . . . , . . . | $$ | . . . O . X . . . . . . . . . X O . . | $$ | . . . . B . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

My AI was changing its mind a lot about the second line move, and I was enjoying watching it discover some clever timings. It seemed to settle on an 'optimal' variation that was somewhat better for W than the AI's main line. I felt I understood the reasoning behind each of the moves, so I started analyzing the other promising candidate. If W tries to cut at the waist, then B will cut with atari.

Click Here To Show Diagram Code: [go]$$Bm27 $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . , X . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O X . . . . , . . . . . , . . . | $$ | . . . O X . 9 . . . . . . . . . . . . | $$ | . . . O X 8 . . . . . . . . . . . . . | $$ | . . O X . 6 7 . . . . . . . . . . . . | $$ | . . . . X 4 5 . . . . . . . . . . . . | $$ | . . X X O O 1 . . . . . . . . O . . . | $$ | . . . O X 3 2 O . , . . . . . , . . . | $$ | . . . O . X . . . . . . . . . X O . . | $$ | . . . . . 0 . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

W's five stones become trapped! They have only three liberties, so they can't escape. W's only chance is to reduce the liberties on the black three stones, which can only be done as above. But is W strong enough to hold B as B pushes to the side?

Click Here To Show Diagram Code: [go]$$Bm37 $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . O . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . , X . . | $$ | . . . X . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O X . . . . , . . . . . , . . . | $$ | . . . O X . X . . . . . . . . . . . . | $$ | . . . O X O . . . . . . . . . . . . . | $$ | . . O X . O X 9 . . . . . . . . . . . | $$ | . . . . X O X 8 . . . . . . . . . . . | $$ | . . X X O O X 6 7 . . . . . . O . . . | $$ | . . . O X X O O 5 , . . . . . , . . . | $$ | . . . O . X 1 3 4 . . . . . . X O . . | $$ | . . . . . O 2 . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

W dies in a ladder, and the netted stones die too, so this works for B. It's a variation that is very difficult to read out for the AI, since it depends on the grazing ladder collision with the upper right corner a full 50ish moves after the first move. If W can't cut the knights move, then W has to cower and accept a bad exchange or get squeezed, so obviously the move is better than any other. In retrospect, I could have found this without my AI friend, and probably even have guessed that it was the variation the AI missed. The nose tesuji triumphs again! My AI gives B a 57% winrate after being shown that it works, and a 50% winrate beforehand. Cool position, B's stones all work so perfectly.

Gomoto · #3

baduk · #4

I am very interested what the strong response might be,can you tell us?

baduk · #5

And if the first move is the attach,why is it a blind spot?leela suggest this move immediately,maybe I misunderstand something

jlt · #6

I only have LeelaZero 157 on my computer. After 10k playouts, the nose attach is only visited 60 times and gets 43% winrate for Black, but once played it gets 57% winrate. Here is LZ's main line, still with 57% winrate for Black at the end:

Gomoto · #7

@baduk: Does your LZ version find the move in this diagram reliable too? (I found the move on Josekipedia and was puzzled that LZ and KataGo did not find it reliable. When I am repeatedly analyzing this kind of positions, only from time to time the bots consider this move on my setup. I thought this might be interesting.)

Click Here To Show Diagram Code: [go]$$B $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . O X . . . . , . . . . . , . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . . O X . . . . . . . . . . . . . . | $$ | . . O X . . . . . . . . . . . . . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . X X O O . . . . . . . . . . . . . | $$ | . . . O X . . O . , . . . . . X . . . | $$ | . . . O . X . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Gomoto · #8

Attachment:

blind2.JPG [ 161.41 KiB | Viewed 10217 times ]

Gomoto · #9

Attachment:

blind3.JPG [ 162.13 KiB | Viewed 10217 times ]

Uberdude · **#10**

@jlt, if you are still using 157 because your computer is too slow for the bigger networks I suggest you get Lizzie 0.7 as it comes with a much newer and stronger 15-block network that has been trained on 40-block games.

jlt · **#11**

Thanks for the suggestion. I see that 15-block networks keep being trained here:

https://github.com/leela-zero/leela-zero/issues/2192

John Fairbairn · **#12**

Quote:

@jlt, if you are still using 157 because your computer is too slow for the bigger networks I suggest you get Lizzie 0.7 as it comes with a much newer and stronger 15-block network that has been trained on 40-block games.

I note that Lizzie 0.5 gave a win rate of ~54% for White on start-up but 0.7 now gives ~57%.

What does this tell the experts among us?

dfan · **#13**

I assume that the Leela Zero network that comes with Lizzie 0.7 is stronger than the Leela Zero network that came with Lizzie 0.5, and stronger networks tend to be more opinionated (the usual gedankenexperiment is to imagine an infinitely strong network, which would give White a winrate of either 0% or 100% at the start of the game).

jann · **#14**

John Fairbairn wrote:

I note that Lizzie 0.5 gave a win rate of ~54% for White on start-up but 0.7 now gives ~57%.

What does this tell the experts among us?

That in recent selfplay games W wins more often than used to in the past.

Bill Spight · **#15**

John Fairbairn wrote:

I note that Lizzie 0.5 gave a win rate of ~54% for White on start-up but 0.7 now gives ~57%.

What does this tell the experts among us?

I am no expert, but the difference is interesting.

As dfan indicates, as bots get stronger, their winrate estimates tend to the extremes. Remember that these estimates are of slightly different things, being estimates of how often each player wins with different colors in self play. Since the bots are different, their results from self play may well be different.

But the shift in White's favor is interesting, because it suggests a lower effective temperature of the empty board. Normally we expect that between evenly matched opponents the first player's advantage will increase as their level of play increases. The opposite is the case here. OC, it's only one datum, so we can't draw any conclusions, but it may be that the stronger White player is better able to take advantage of the symmetry of the empty board to lower the effective temperature. If so, that should show up as a change in White's overall strategy.

Note that the shift in winrate estimates is the opposite of Elf. Assuming that, when it first arrived, it was stronger than most other bots, its winrate estimate for White on start-up of 50½% suggests that as Black it was better able to make use of the first play. The initial estimates by LZ and Elf now differ by around 6-7%, which is large enough to make me think that one of them is wrong. I suspect that it is Elf, because in the Elf variations in its commentaries White follows Black around noticeably in early play, which does not allow White to take full advantage of symmetries. OC, on my own I could not say that White Elf is wrong to do so, but maybe LZ can show us Elf's mistakes.

Uberdude · **#16**

Bill Spight wrote:

Normally we expect that between evenly matched opponents the first player's advantage will increase as their level of play increases. The opposite is the case here.

Bill, you've said this before, but I still don't see why you say that. If the 7.5 Komi is slightly too big, as seems to be the case, then as a bot gets stronger I would expect its winrate for White on the empty board to increase as the stronger player is better able to carry that advantage through to the endgame.

Gomoto · **#17**

When the bot was at a local maximum before, it would also be possible that the stronger bot shows a lower winrate.

The bot realizes that go is more difficult, than it anticipated ;-)

Gomoto · **#18**

The blindspot in the analyzed joseki is about 4 points better than the normal AI move according to KataGo when you continue the variations.
That is quite a gap.

jlt · **#19**

dfan wrote:

I assume that the Leela Zero network that comes with Lizzie 0.7 is stronger than the Leela Zero network that came with Lizzie 0.5, and stronger networks tend to be more opinionated (the usual gedankenexperiment is to imagine an infinitely strong network, which would give White a winrate of either 0% or 100% at the start of the game).

On the other hand, Alphago says that on an empty board, White's winrate is 52.9%. More likely, Leela Zero is going through a path in which the playing style favors White.

Gomoto · **#20**

KataGo attaches about +1 point to white with Komi 7.5 and -1 with Komi 6.5

Find the best move (AI will be no help)

Who is online