I used a 40b256c network from KataGo that is in the middle of training right now, for around 50k playouts on each move.
- Click Here To Show Diagram Code
[go]$$Wc Base position
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . X . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . f . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . i j . . |
$$ | . . . X O . . . . . . . . . . . . . . |
$$ | . O X O O X X X . , . . . . . X . . . |
$$ | . O O X O X O X . X . . . . . . h . . |
$$ | . O X X X O O O X . . . . . . . . . . |
$$ | . . a c e g d b . . . . . . . . . . . |
$$ ---------------------------------------[/go]
- Base position: B+3.2. White wants to play h (but f is also just as good).
- After white a: B+5.0, so white lost about 2 points by playing in the lower left. Black now wants to play j (but i is also about as good).
- After white a, black b: B+2.0, so black lost about 3 points by playing in the lower left. White now wants to play h (f is now judged between 0.5 to 1.0 points worse than h, or about 4% winrate).
- After white a, black b, white c: B+3.0, so white lost about 1 point by playing in the lower left. Black wants to play i (but j is also about as good).
- After white a, black b, white c, black d: B-3.0, so black lost about 6 points by playing in the lower left. White should just respond and capture at e. (white playing h would be about 3 points worse, playing f would be about 5.5 points worse).
So:
It is repeatedly the case that playing in the lower left is too small. The per-move value of the raw territory in the lower left is something like 9 points (8 points due to the capture swing, 1 from territory on the first line). Which is quite a bit too small for the opening. The reason why playing elsewhere is only often 1 or 2 points worse is that there is extra value beyond the 9 points in outside forcing moves and group safety from resolving the capture. So this all makes sense.
The a-b exchange costs black 1 point. The c-d exchange costs black another 5 points, at which point capturing is actually the largest move for white, white no longer wants to tenuki. Why?
- Click Here To Show Diagram Code
[go]$$Bc Future forcing move
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . X . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . 1 . f . . . . . . . . . . . . |
$$ | . . . X O e . . . . . . . . . . . . . |
$$ | . O X O O X X X . , . . . . . X . . . |
$$ | . O O X O X O X . X . . . . . . . . . |
$$ | . O X X X O O O X . . . . . . . . . . |
$$ | . . a c . . d b . . . . . . . . . . . |
$$ ---------------------------------------[/go]
I'm pretty sure the answer is that KataGo senses that exchanging a-b in the lower left might in some variations reduce the effectiveness value of this forcing move when white resists by coming out at 'e', because black's stones down there have one fewer liberty. I'd be far from sure "1 point" is the right value for this, because it seems to me black can still follow with 'f' with about the same effect. So I don't entirely trust KataGo here, but it also seems plausible to me that it's right, there could well be variations where it matters.
And exchanging c-d further is very bad for black, now white wants to respond locally to finish the capture, and then black entirely loses out on this forcing move, which would be helpful in building the center and limiting white's left side.