Ideal win rates, Kill-all go

I saw the concept of Kill-all go and found it very interesting.

For those not familiar, it starts with some variation of this board position:

And the goal is for white to try to make one living group somehow. If white lives anywhere, black loses. Black wins only if it kills everything.

Seems like a great way to practice attacking / making eyes.

This got me wondering, assuming both sides play “perfectly,” will White be able to live somehow?

For example, if we put two strongest AI in this situation, what’d be the predicted win rate for Black and White?

(I’m down to play kill-all go if anyone wants. Send me a message / challenge on OGS! Shen0927)

3 Likes

Seems fun. I’d like to try it. How do we set up the board like that though? :thinking:
I’ll send you a message and you can send the challenge.

One way to start a game from a particular position is to the create a demo board to set up a position and then click on fork the game to create a challenge that can be sent to another player.

Here is a demo board:

If you set the forked game to use Chinese rules and with a komi of 360.5, then the outcome from scoring aligns with the desired objective.

8 Likes

By the way, another variant is called the Shape Game.

Here are some relevant forums threads:

1 Like

Challenge sent!

1 Like

@yebellz I’ve accepted the challenge and started our match as well! Let me know if you want to find a time to play live :slight_smile:

1 Like

This mode is actually quite a good training:

5 Likes

Here’s my game

3 Likes

hahaha, what an amusing coincidence: We both got an eye around L15 :sweat_smile:

2 Likes

haha interesting observation. Maybe I have a tendency to play the wrong moves around L15? :thinking:

2 Likes

From limited experience so far, I feel like Black is easier to play than White.

By coincidence maybe, I’ve played as Black in all my 3-4 kill all games, and managed to kill everything in each game, even against stronger players like @JethOrensin

I tried to put myself in the shoes of White during the game, and couldn’t really think of any good ideas myself to live. Black just has too strong of influence all over the board.

Either this board setup favors Black, or maybe I’m just better at attacking than living.

4 Likes

From an Ai perspective white can live at least two times in this setup.
Of course its Ai, but if it can live at least with two groups, humans should be able to live once I think, but it depens on the level of the players who is in favor I guess. Thats why this format is a great thing.

(I tried to set up a board position I could play kill all Go against Ai.
I didn’t find a good way yet because Ai often plays tenuki at crucial points where humans may still mess up, or it lets you live easy but ‘small’)

Just curious; How do you know?

Personally I’d be careful with jumping to conclusions with AI review, because the AI probably wasn’t trained on this starting position with 360 points of komi.

2 Likes

Well, I just played it out with some Ai :eyes:
You think that doesnt neccessarily mean much?

Well, it’s worth clarifying the experimental set up. What bot are you using? How did you set up the game? Are you using a large komi to force the appropriate winning condition?

Overall, your previous post seems a bit self-contradictory, since you also say:

2 Likes

I believe that positions as lopsided as this probably don’t come up in any AI training. Also an AI playing black may play sloppy moves, since the AI was trained with a normal Komi value, and optimises for win probability.

4 Likes

I remember seeing in the Computer Go Discord that someone had trained a KataGo network to play kill-all go in some different configurations, I can’t search it up right now but I recommend looking there for anyone curious! The network itself might not be publicly available but there were at least some screenshots of its analysis and winrates in initial positions.

4 Likes

I will hope someone will train an AI for playing this go. It must be interesting.

2 Likes