Life In 19x19

Getting KataGo to solve tsumego is an interesting game! Often, if you put a life and death problem in a corner and leave the rest of the board empty, KataGo will just tell you that the first move in an empty corner is bigger. The challenge is to construct a whole-board position where the tsumego is hotter than anything else. But it's worse than that. Even if you can persuade it to pick the "right" corner, KataGo will suggest different moves depending on the whole-board context.

Does anyone have tips for this kind of analysis?

Here's an example from the Hitachi tsumego.

Click Here To Show Diagram Code

[go]$$B Black to play and kill
$$ --------------------
$$ | . . . . . . X . . .
$$ | O O . . . O O X . .
$$ | X X O . . O X O X .
$$ | . X O O O X X . . ,
$$ | . . X . X . . . . .
$$ | . X . . X . . . . .
$$ | . . . . . . . . . .[/go]

Solution Alternative variation I've constructed three full-board positions:

• Balanced: black can win if and only if black kills the top left group.
• Ahead: black can win without killing (but killing is still presumably the best move)
• Behind: killing makes the game close, and is still the best option, but white is ahead

In the first situation, KataGo solves it quickly. But in the others, it has a surprising amount of trouble finding the key move, and prefers the alternative variation.







Note: I'm still on KataGo 1.3.2 with a 20-block network, too lazy to upgrade at the moment! Please post here if you get different results with newer versions.

You can restrict KataGo's analysis to a range of coordinates (for a prescribed number of moves).

If you do it via GTP command kata-analyze, the syntax is described here: https://github.com/lightvector/KataGo/b ... ensions.md

If you use a GUI, most of them support this in some way.

If you analyze a board with your corner stones only and restrict the analysis for 20 moves to the smallest rectangle that includes all stones, KataGo will quickly find the right move and continuation (with a 40b network, 20b doesn't get it right).

I just uploaded a short article about a tsumego solving AI on the Sensei's Library: https://senseis.xmp.net/?PandaSensei