Let me use the first moves of this game and the Elf commentary to explore some of the questions raised about go knowledge, bot choices, and making improvement.
One advantage of the Elf commentaries is that they are based on a large number of rollouts, many more than we ordinarily might have the time to wait for.
- Click Here To Show Diagram Code
[go]$$Bc Direction of play?
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . c . . . . . |
$$ | . . . 2 . . . . . , . . . . 4 , 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 5 . . . . . , . . . . . a . . . |
$$ | . . . b . . . . . . . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
First, let me start with
. It is certainly playable, but it is nearly off of Elf's radar, with only 8 rollouts. It loses 1% in winrate to Elf's first choice,
a, which gets 40k rollouts. Elf does not even show its second choice, which means that it got fewer than 1500 rollouts. My guess is that it is
. Looking at this and other games, it appears that Elf has a small but decided preference for 4-4 points, but that does not mean that other corner plays are bad. The margin of error matters, whatever it is.
is no surprise to those who study pro games or have read Kajiwara's book,
The Direction of Play.
bears a good relationship, we are told, to
. How bad can it be? Well, says Elf, it gets 0 rollouts and loses 6% to the nirensei at
5 (25.8k rollouts), which makes it a minor error in my book. It is hardly fatal, OC. (Elf and other bots do not seem to care about the human concept of direction of play. Until, perhaps, a little later in the game.) After the White nirensei, Black encloses the top right corner at
c. That enclosure is AI knowledge. It was not previously unknown, OC, but Black usually played more solid enclosures; it was White, usually playing catch-up, who played the thin enclosures. Elf's second choice for
is
b.
Otake occupied the last open corner with
. This is in line with my proposed bit of modern go knowledge, based upon Elf's choices, which is that playing in the last open corner is a kind of
last play. It is not always correct, OC, but it is a good idea.
In this case Elf prefers the play in the next diagram by 2½%.
- Click Here To Show Diagram Code
[go]$$Bc Zwischenzug
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . 6 5 7 . . . |
$$ | . . . 2 . . . . . , . . . 8 4 , 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 9 . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
Before occupying the open corner, Elf prefers to take profit in the top right corner with sente. The usual joseki is for Black to extend to
a, but that would allow White to take the last open corner.
- Click Here To Show Diagram Code
[go]$$Wcm10 AI joseki
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . O X X . . . |
$$ | . . . O . . . . . , . . . O O 6 X . . |
$$ | . . . . . . . . . . . . . . . 4 1 2 . |
$$ | . . . . . . . . . . . . . . . . 5 3 . |
$$ | . . . . . . . . . . . . . . 7 . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
When I was learning go, leaving the attachment,
, was generally frowned upon. After
White usually extended to
and Black crawled on the second line. But the bots often like the two step hane at
. This shape occurs 24 times in Waltheri (
http://ps.waltheri.net ), 9 times in the AI era.
BTW, some of you may be wondering how you can get a reliable winrate estimate from 0 rollouts
, or even 8. Well, you can't. What Elf does is to inherit the estimate from Elf's top choice for the next play. Thus,
got its winrate estimate from the White nirensei in the bottom left corner, and
got its from Elf's underneath attachment in the top right corner. Elf does this for any play with fewer than 500 rollouts. By contrast, how many winrates have you seen based upon 100 rollouts or less?
- Click Here To Show Diagram Code
[go]$$Bc Moves 1 to 10
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . 8 7 9 . . . |
$$ | . . . 2 . . . . . , . . 0 . 4 , 1 . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 6 b . . |
$$ | . . . 5 . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
This is as far as I am going to go, now. If
was a minor error, how bad is
? You might well ask. It loses 7½% to the outside attachment at
a, says Elf, another minor error. See, we figured out that
was problematic, based upon the fact that
was.
What about
? Here is where the human concept of the direction of play becomes relevant, IMO. If
bears a good relation to
, how much better will the relationship be between
and the stronger White position after White replies? It seems to me that direction of play points to playing in the bottom right corner. Elf prefers the bottom attachment there by 4%, which puts
in the top right into questionable territory. It may not be a mistake, but it might be.
- Click Here To Show Diagram Code
[go]$$Bcm7 Variation for
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . O , X . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . b . . . |
$$ | . . . . . . . . . . . . . . . . 6 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 4 2 . . |
$$ | . . . . . . . . . . . . . . . O 1 . . |
$$ | . . . X . . . . . , . . . 5 . , 3 . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
I don't know why
is high, but look at
, an unusual joseki for the top right corner. Again, I think that direction of play is a factor. The bots have taught us that the "standard" extension from the bottom right stones can wait, as a rule, but surely
works with
-
. It seems that Elf thinks that the attachment at
a or the keima at
b are each a little too far away.
- Click Here To Show Diagram Code
[go]$$Bcm13 Variation for , continued
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . 5 . . . . . . . 2 1 3 . . . |
$$ | . . . O . . . . . , . . . 4 O , X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 6 . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . 7 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . 8 . . O O . . |
$$ | . . . . . . . . . . . . . . . O X . . |
$$ | . . . X . . . . . , . . . X . , X . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
and
take profit in the corner, no surprise there, and then
approaches the top left corner from the top side, something that may go against the idea of playing from the wide side, but plainly obstructing White's development on the top side.
refrains from a pincer, something that we learned from AlphaGo. Next,
enters the right side with a pincer and
jumps into the center, something we also have learned to expect from AI.
- Click Here To Show Diagram Code
[go]$$Bc Moves 1 to 10, repeated for convenience
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . 8 7 9 . . . |
$$ | . . . 2 . . . . . , . . 0 a 4 , 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 6 . . . |
$$ | . . . 5 . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
In case we didn't know,
is another minor error, losing 7% to the solid connection at
a.
gets 0 rollouts, while
a gets 91.4k rollouts. Elf really likes
a. Even though White has made only minor errors, he has lost 15% in only 5 moves. It all adds up, as Kirby has pointed out. Keep in mind that all of White's plays so far have conformed to the collective go knowledge of the time, of which White was a master.