I have played a lot of large board games recently, mostly 23x23 but also a few 21x21 and 25x25. Here are some observations and questions that I have been unable to answer yet.
Everything is about influence. The game play changes very much from 19x19. It is like the difference between 13x13 and 19x19.
Even if you lose a medium sized group, you can usually make up for it later in the game.
There are a lot of things “going on” on the board which interact with each other in surprising ways.
19x19 feels very small afterwards. It becomes easy to understand what is going on in a 19x19 game, even if you still lose.
Even 25x25 is quite playable, and hardly the largest tolerable size.
Correspondence games are really good here. People tend to time out with byoyomi though, Fischer seems more lenient.
Handicap needs to be adjusted.
Are computer players good at this or will it be a throwback to the GNU Go stone age?
I think computer players would at least have to be adjusted and recompiled, even if the required changes in current programs are small, which might discourage cheating. I never wanted to play 19x19 correspondence because of the temptation for the opponent to cheat.
I made a few of these boards (painted and lacquered MDF) about 10 years ago. I made one board for every uneven size from 5x5 to 27x27. I originally intended it for multi player go with up to 6 players (you only have to trivially change the ko rule), but I got enough “stones” (actually colored bead things that resemble go stones) to be able to play a 2 player game on 27x27. It takes a really long time OTB.
What do you think about playing on these crazy boards?
Strong AIs are usually trained to a specific board size. The training time for bigger nets is much higher, but the same techniques would work there as well.
That’s really neat that you made physical large boards.
I played a few games on 37x37 on KGS a long time ago. I thought it was pretty fun. We played with about ~10 second moves, to keep the game length down.
It’s really a different game. I think I was experimentally opening on the 5-5 points. And the most interesting thing was how small the 19x19 felt after even one game on 37x37. Would play again.
49x49 is an epic 2401 intersections. If the stones occupied only half to finish the game, and 10 moves were played each day, you could finish in four months. If it took more moves to finish and you played slower it could last over a year.
If you played live and averaged 10 seconds per move, it’d take 3 hours and 20 minutes to fill half the board. A marathon!
Funnily enough, AIs can be strong on large boards without ever training on them. Recently released a version of KataGo’s code that works up to 29x29, and the same neural nets in the current run that were only trained on 7x7 through 19x19 seem to work okay on the larger boards. https://github.com/lightvector/KataGo/releases/tag/v1.3.5%2Bbs29
One key to making this work is actually that convolutional nets don’t physically care what the board size is. You feed in the board as essentially a “image” (e.g. 19x19 pixels, except instead of RGB you have channels that specify the stone on each spot and whatever other properties you want), and then almost all the layers in the net merely specify a sequence of local transformations by which pixels should change based on the pattern of their neighbors. And magically those transformations result in a map that shows what all the good moves are, who owns what territory, who’s winning, etc.
Since these transformations are simply uniform rules about how pixels should adjust values based on their neighbors, they can always be applied no matter how big the board is - every pixel still just follows the same rule and updates based on its neighbors. The only question is whether those transformations all together also happen to produce good moves and evaluations on a big board the same way they do on a smaller board, when the only thing they were trained to do was to produce those good moves and evaluations on the smaller board.
It seems in practice they do, to an extent. (if you compile KataGo to try to scale up the board even more, and more to size 40, 50, 60,… you’ll finally start to see the nets say really crazy things given that they were trained on size at most 19).
Haze and @Leidang, There are slightly over 20 Go Players in a 37x37 board size “infinite tournament” (what is called a “ladder” here on OGS) by correspondence on the “Little Golem” Game Server. One of the games which has yet to be finished was started over 3 years and 3 months ago! As @Aumpa mentioned, there were people playing 37x37 games on KGS long ago (and maybe still doing so, IDK, I stopped playing on KGS a few years ago) - the most famous was “Beer Slayer”. Anyway, 37x37 is playable if you are willing to put in the time, but IMO there is a “diminishing return” on going bigger. A 21x21 game is not much different than a 19x19 game, but a 25x25 game is easily noticeably different than 19x19 in how the strategic thinking/planning is more important relative to the tactics (local fighting / life-and-death battles). A 25x25 game, to me, feels like a war with 8-10 important battles - one in each corner, at least one on each side and often/usually one in the center (which is basically the same for 19x19 except the side battles are more strongly influenced by the corner battles)… However a 37x37 game can have 15-25 (or more) battles going on. It is a little crazy. If you want see the over 3-year long 37x37 game, it is taking place here: https://littlegolem.net/jsp/game/game.jsp?gid=1880016
We’ve seen how interesting a game can be for a 19x19 board, obviously the properties change a bit for larger boards (I think it was mentioned influence becomes a lot more important). I was wondering if at some point when you “zoom out”, so to say, if things start getting more simple again. There’ll be some interesting stuff at the borders of domains, I guess, but I wonder if there’s some interesting macroscopic behaviour that might teach us something.
I’m interested in getting AI to compete against one another on larger and larger boards. I’m interested in doing AI automated matches on boards of size 100x100 and up. It presents an interesting technical challenge (how do you optimize calculation time on boards of that size?) and it would also be an interesting tournament setup (give set computational time per turn for each computer, maybe something a little unrealistically small, and see which computer wins. You set matches to be an hour long, for example, and figure out how long each computer can afford to compute a turn). I feel like there’s a lot of things we could learn from seeing these larger board sizes, but maybe my interests are a bit more scientific than competitive.
Any ideas what programs or nets I could use to try and play these games? I’ve currently got a few nets on my machine but I don’t know where to look to start finding anything that could realistically play on these extra large boards…
I’m really curious how a trained bot would handle 37x37. Like, even just looking at the basic principle of corner-sides-center. The sides are so long and the center is so much bigger, it’d be really interesting to see how the stones spread out, not to mention the new corner josekis.
Hi, I’m wondering if we can get ranked games for board sizes higher than 19x19? Also, how hard would it be to add board sizes to OGS larger than 25x25? Thank you.
This is speculation, but I would guess that an issue with rankings for less common board sizes is that their limited activity might result in limited data to produce confident rankings. Also, since theory is not so well developed for such sizes and people are generally less familiar, I would guess that player performance would be more volatile on uncommon board sizes, leading to more noise in trying to estimate an overall rating.
You mean, “we cannot get ranked games with non-standard (9, 13, 19) board sizes here on OGS.”. . . But we can on the Little Golem Game Server where they play 27x27 (just added) and 37x37 WeiQi —> Little Golem
