Life In 19x19

Click Here To Show Diagram Code

[go]$$Wc19
$$ +--------------
$$ | . . . . . . .
$$ | . . . a . . .
$$ | . . . b . O .
$$ | . . . X . . .
$$ | . . . . . . .
$$ | . . X . . . .
$$ | . . . . . . .[/go]

Yeonwoo says that 'a' is still OK for DDK level players, but nowadays the rest of us should know better. You should watch the video first, because it's very entertaining, but then here's a summary of the variations discussed.



For me the biggest surprises were not so much the actual moves, as the decisions about when it's OK to tenuki. I'll have more questions later, but right now lunch takes priority :-)

I think there are still positions where the slide is good.

I didn't check with AI, but this is a recent example where my feeling says it's OK: https://gokibitz.com/kifu/rJge9RJqB?path=28

There are also examples of AlphaGo games where it slides.

In this situation, Katago want to answer the slide by attaching at N3. The slide is a -1.5% wr/-0.5 pts "mistake" compared to the attach, but if black play the 3-3 after the slide, white gained a little in the exchange.

The slide is a -1.5% wr/-0.5 pts "mistake" compared to the attach, but if black play the 3-3 after the slide, white gained a little in the exchange.

Bill has given us a lot of insight into what win rates mean for Leela and Elf, and, as I understand it, even big numbers are not always as horrifying as they seem, while small ones can be treated as within a margin of error. But I think he has no experience with Katago.

To those who do have experience with Katago, what is your feel for reliability of the win rate/points equivalence? And is it the same at all stages of the game? The quotes around "mistake" here are suggesting similar caution.

In the comments at move 10 you say that Yeonwoo made a mistake by referring to B17 as the move that Choi Jeong played whereas it was actually played by her opponent. I didn't get that impression from watching the video. The impression I got was that this B17 move is normal, and the impression I got was that the move at E18 was the move that Yeonwoo was referring to as the Choi Jeong novelty. So, I didn't understand your annotation at this move where you stated that Yeonwoo got her players mixed up.

gennan wrote:I think there are still positions where the slide is good.

I didn't check with AI, but this is a recent example where my feeling says it's OK: https://gokibitz.com/kifu/rJge9RJqB?path=28

There are also examples of AlphaGo games where it slides.

Yeonwoo states in the video that the slide can still be good depending on the outside and gives examples where she would slide instead of playing attach.

John Fairbairn wrote:

The slide is a -1.5% wr/-0.5 pts "mistake" compared to the attach, but if black play the 3-3 after the slide, white gained a little in the exchange.

Bill has given us a lot of insight into what win rates mean for Leela and Elf, and, as I understand it, even big numbers are not always as horrifying as they seem, while small ones can be treated as within a margin of error. But I think he has no experience with Katago.

To those who do have experience with Katago, what is your feel for reliability of the win rate/points equivalence? And is it the same at all stages of the game? The quotes around "mistake" here are suggesting similar caution.

I don't think there is a simple winrate to points conversion. It depends on the specific AI, the stage of the game and the details of an actual position.
So I never found winrate a very clear measure for humans and I'm quite happy with this feature of KataGo.

KataGo may not be as strong as LeelaZero, but KataGo's score estimation is the best we have. No other freely available score estimation tool comes even close.

I'm quite convinced it's much more accurate than my own score estimation (and much quicker). I feel it's on par with professional level score estimation and probably even better. To test its accuracy, one could test it with reference endgame problems. But I don't think there exist reference opening positions with professional score estimations, except for komi and professional score estimation of handicaps (like here). So in most opening positions, trusting KataGo's score estimation is mostly a matter of faith, I would say.

One thing that can be tested is consistency by adjusting komi to 50% winrate and noting the score deviations, like xela did here. His results (albeit still anecdotal) seem to indicate an error margin of 2-5 points very early in the game for a wide range of winrates. When seen as the sum of 361 ownership fractions for area counting, the consistency error is only about 1%. That's already impressive to me and it confirms my feeling that KataGo's score estimation is at least on par with professional level score estimation in a wide range of positions.

I like that KataGo shows the score estimate as decimal number (not only an integer number) and I find this especially helpfull during endgame.

John Fairbairn wrote:

The slide is a -1.5% wr/-0.5 pts "mistake" compared to the attach, but if black play the 3-3 after the slide, white gained a little in the exchange.

Bill has given us a lot of insight into what win rates mean for Leela and Elf, and, as I understand it, even big numbers are not always as horrifying as they seem, while small ones can be treated as within a margin of error. But I think he has no experience with Katago.

To those who do have experience with Katago, what is your feel for reliability of the win rate/points equivalence? And is it the same at all stages of the game? The quotes around "mistake" here are suggesting similar caution.

I mean, Katago think it's a 0.5 pts loss, I wouldn't call this a mistake. If Katago evaluation is correct, I would call this an inaccuracy/suboptimal choice. If you look at the moves played in this sequence :

: -4%
: -7%
: -1.5%
: -2%
: -7.5%

At this point of the game, a 1pt mistake is around 3% winrate

Jujube wrote:In the comments at move 10 you say that Yeonwoo made a mistake by referring to B17 as the move that Choi Jeong played whereas it was actually played by her opponent. I didn't get that impression from watching the video. The impression I got was that this B17 move is normal, and the impression I got was that the move at E18 was the move that Yeonwoo was referring to as the Choi Jeong novelty.

Ah, it makes more sense that way. Thanks for the correction!

Jujube wrote: Yeonwoo states in the video that the slide can still be good depending on the outside and gives examples where she would slide instead of playing attach.

Yes, this example is discussed at 9:15. I can't find the position in GoGoD 2018: I guess it's a more recent game. White to play.

xela wrote:

Jujube wrote: Yeonwoo states in the video that the slide can still be good depending on the outside and gives examples where she would slide instead of playing attach.

Yes, this example is discussed at 9:15. I can't find the position in GoGoD 2018: I guess it's a more recent game. White to play.

slide_example.jpg

What about playing in the unoccupied corner?

OK, on closer examination the position above has 14 white stones and 12 black stones! Adding a white stone at the top left 4-4, I find that both KataGo and LZ slightly prefer the attach to the slide, but it's one of those "within the margin of error" things. The follow-ups are interesting:

Click Here To Show Diagram Code

[go]$$B diagram 1
$$ ---------------+
$$ . . . . . . . |
$$ . . . . . . . |
$$ . 1 . . 3 4 . |
$$ . . . X 2 6 . |
$$ . . . b 5 a . |
$$ . . . . O . . |
$$ . . . . . . . |[/go]

Yeonwoo explains that black 'a' next would build thickness: I guess this is bad for white because it hurts the stone below. But the AIs say that black 'a' in this context is a small mistake (-4% winrate according to LZ) and black should play 'b' instead. What's wrong with 'a'?

Click Here To Show Diagram Code

[go]$$Wc diagram 2
$$ ---------------+
$$ . . . . . . . |
$$ . . . a . . . |
$$ . X . 5 X O . |
$$ . . 6 X O O . |
$$ . . . 3 X 2 . |
$$ . . . . O 4 . |
$$ . . . . b . . |
$$ . . . . . . . |
$$ . . . . . . . |
$$ . . . O . . . |
$$ . . X . . . . |
$$ . . . . X . . |[/go]

From here, the joseki move is 'a', followed by black 'b'. But LZ says that 'a' is a terrible mistake (winrate drop of 17%), and in this position white should play 'b' instead:

Click Here To Show Diagram Code

[go]$$Bc diagram 3
$$ ---------------+
$$ . . . . . . . |
$$ . . . 6 . . . |
$$ . X . O X O . |
$$ . . 1 X O O . |
$$ . . . O X X . |
$$ . . . . O X . |
$$ . . . . 2 3 . |
$$ . . . . 4 5 . |
$$ . . . 8 7 . . |
$$ . . . O 9 . . |
$$ . . X 0 . . . |
$$ . . . . X . . |
$$ . . . . . . . |[/go]

Of course this only works in this position because of the white stone at Q10. Without white friends below, is a mistake (-7% winrate according to LZ-242) and white isn't looking so healthy after .

Bill Spight wrote: What about playing in the unoccupied corner?

Yep, you posted as I was typing. Good question. I guess it's a transcription error and there should already be a white stone there. The image is a screen shot from the video at 9:15. Yeonwoo was talking about the top right with so much enthusiasm, I didn't even notice the empty corner!