Collective development of a server for variants

Go Variants
82

A popular variant of go is “kill-all go”, where Black has a very large number of handicap stones, and White wins if they can make a living group.

If the handicap stones are already placed, then this variant can easily be played on any go server, for instance by using Chinese rules and setting a komi close to 361, so that White wins if they have at least one living group.

However, there is another variant with a kind of “auction” for the number of handicap stones at the beginning of the game.

In this variant, at the beginning of the game, both players play with black stones. This goes on until one player, instead of playing a black stone on their turn, says “I can kill you now”. Then, this player keeps the black stones; the other players becomes White. From this starting position, White plays first, and play alternates normally. White wins if they can make at least one living group.

This variant doesn’t seem particularly convenient to play on a standard go server. It might be well-suited for the server for variants.

8 Likes
83

I forgot fox had this go variant - 9x9 but once per game, randomly you get to make a double move

84

Oh, so it’s random. I kinda thought it’s an ability.

85

Yeah it’s completely random, so you have to play somewhat cautious, in case your opponent getting a double move would just completely end the game.

Sometimes you get a double exactly when you need it, other times you get it when it’s of completely no use.

I haven’t played it myself, but I watched thelovingsun play it a good few times on stream.

I can’t seem to find any vods to point to though, or people really discussing the variant anywhere. I thought maybe it showed up here on the forums at one point or something previously.

2 Likes
86

Actually @Allerleirauh - I think maybe I heard about it originally from @antonTobi on the variant go server

where they linked

https://www.reddit.com/r/baduk/comments/rtvhtl/anyone_know_what_the_new_buttons_are_on_the_fox/

“(in short: play two moves in a row at some random times)
Not as good as some of the variants on VGS, but the larger playerbase makes it easy to get a game :slightly_smiling_face:”

^ from antonTobi

and

At the start of the game two of each players’ stones are placed on board in certain positions (?). Then the game starts. Each player has 2 “flying daggers”. From moves 11-40, those “flying daggers” will be activated randomly, at which point the player can place two stones on the board continuously.

from the Reddit post.

87

How to make popular AIs, such as KataGo and LeelaZero, play some Go mods here, so that those mods could be much more challenging? At Github, I find Ya Tienna has make KataGo can play MagnetGo, but it’s too weak, I win every MagnetGo games against KataGo! Unbelievable :roll_eyes: And Cao has trained a new KataGo model to play TaoGo, and it has reached pro level. But I can’t find a link to download the model.

1 Like
88

I think that depends on the variant (or mod as you call it). Training new nets is always possible, but requires a lot of training games. In some variants KataGo could probably do fairly well with the existing nets, e.g. for Capture Go it would just have to evaluate the positions according to the new victory condition.

5 Likes
89

I don’t understand. In Capture Go, the first to capture stones win the game. But KataGo moves to win, not capture stones, there are two different goal.

Just like let KataGo to solve tsumego problems, it is some kind hard to implement. Kaorahi implement via “local search” feature of KataGo. It’s a “trick”, not a “proper method” from my view.

90

You mean edit the KataGo source code or use some feature readymade?

  • Editing KataGo source code is bit hard for me.
  • Local search solve tsumego, which feature solve CaptureGo?
91

Are there TaoGo professionals?

1 Like
92

Using Kata’s existing nets (trained for normal Go) to play Capture Go could be considered a “trick” too, but I’d guess it could work well.

But yes, it would require a changed source code to tell it about the new victory condition and to stop it from sacrificing stones.

Basically KataGo uses its nets (i.e. artificial neural networks) to decide which moves are “interesting” (my wording) and then plays the game to the end with lots of different interesting moves a lot of times.

Moves that are interesting in Go are likely also interesting in Capture Go, that’s why I think KataGo could be good at Capture Go without going through the ressources-intense retraining of the network.

1 Like
93

I think you can also retrain existing nets for a much lower compute cost than the original training. In fact, this is sort of how Katago training happened if I understand - it trains a net on small boards, then keeps that net and retrain on larger sizes.

But I have no idea how one would actually go about injecting new Go rules :sweat_smile:

3 Likes
94

I think Capture-N Go1 converges to Go with group tax when N tends to infinity, so, it makes sense that they are similar.

Hmmm, actually...

Definitions for pedants:

Position-sequence: Given a position P (stones on the board + captures on each side + ko restrictions), you generate a sequence {Pn}n≥1 where Pk is the position P in the game Capture-k go. This game state is (game theoretically) said to be either winning (W), losing (L) or undefined (?) and also to have an optimal score S(Pk) which is a real number or undefined.

Positional convergence: Given position P, it is said to game-result converge to R in {W,L,?} if the game states Pk converge to R (with discrete topology) when k tends to infinity. Likewise, it is said to score-converge to S if S(Pk) converges to S (in the real numbers plus the isolated undefined symbol) when k tends to infinity.

Game convergence: Given a sequence of games {Capture-n Go}n≥1, they are said to converge to game G if for all positions P, the position-sequence {Pn}n≥1 converges to the position P in G.

Well, but...

Rules for the nit-picky:

Capture-N Go defined under the following ruleset:

  • First person to obtain N or more prisoners wins.
  • You must pay one stone to pass.
  • Game ends by the first rule, resignation or agreement (but not necessarily by consecutive passes).
  • Some kind of ko rule I guess? (As some infamous youtuber would say “It’s complicated, mmmkay”).
Still not satisfied... Well, I don't know what else to say. You've been summoned though, if you know you know. The meta game works. Here is a video for your entertainment for having made it down here

Notes:
1: Atari-Go = Capture-1 Go

1 Like
95

Not needed, because each capture (in the ko or not) counts down toward victory by N captures.

Does that count towards the N captures?

96

I mean this is more important than the stuff that came before it.

In particular the choice to allow or disallow passes makes a huge difference.

Simairly whether passes lift ko bans could have a big chance in results.

I can imagine for instance that if you have a double ko seki on the board with. A capture limit, there could be parity effects, where all other things being equal, one where the number of points needed to win depends on whether N is even or odd etc.

With passing lifting ko bans, you can decide whether to give one point to the opponent, or to give both players a point with consecutive captures, forcing the other player to pass next.

So scores could be like B-W: 0-0, (1-0, 1-1), 1-2, 2-2, (3-2, 3-3), 3-4…

If Black captured whenever it became possible it passes lifted a ko ban, bracketing the double ko captures. So you could see that Black might be the first to 1 point, White to 2, Black to 3, etc with this sequence.

There’s an alternate sequence where Black passes first and white captures the double ko first

0-1, (0-2, 1-2), 2-2, 2-3, (2-4, 3,4), …

But that doesn’t seem beneficial.

I think there’s so many choices of rule combinations I wouldn’t be confident in saying “Capture N go” tends to go.

97

It’s probably needed if you want the game to resemble go in the long term? Like that the results of the games match maybe.

I was kind of assuming it did.

98

Recently, I’ve used the modified KataGo to train a Daoqi model. After >2 weeks of continuous training and approximately 100k games, I got a pretty strong model.

How strong? I’m an AGA 6D, and the model can beat me with 4 handicaps. That’s pro level. (Even though Daoqi is different from Go, I think the ranking can be translated from Go to Daoqi.)

gcao at lifein19x19

99

That is true, except it would make for a very disbalanced game.

If one of the players is ahead in the prisoner count by even one prisoner, they’d be incentivised to simply retake every ko, making them effectively the master of all kos. Essentially, that player wins all kos and the other player loses all kos, regardless of ko threats or tactical considerations.

I find that to be pretty boring.

Of course.

100

Indeed, without the ko rule, the game would play out very differently, and it would not be the same as Go.

One strategy that I can forsee would be rushing to be ahead on the prisoner count in order to become the ko master and use that as an overwhelming advantage.

It would be as if (normal) Go had the following rule: “When you repeat a position, instead of a loss or no result, the winner is whoever has more prisoners”.

101

I somewhat doubt that, because a particular game of go can easily become finite.

For instance, imagine I am very much ahead in territory. Once territories have been decided, and neutral points have been played, we will start filling our own territories, until my opponent has exactly two eyes par group. Then my opponent puts themselves in self-atari; I capture their group; now I’m even more ahead in territory. And so on; eventually all my opponent’s groups are captured. Which means every empty space on the board is my territory. We keep playing for a while, until eventually my opponent has no legal moves left because every space on the board is a 1-space eye of my group which encompasses the whole board.

At that point, I have captured a very large number of stones, but assuming there was no unreasonable ko fight, that number should be less than (19x19)^2 which is less than 160000.

So, if this was a game of standard go, I would win, but if this was a game of capture-N go with N > 160000, I would not win.