Can you break the AI with 3-dimensional go

I want to spice go up so that a human can beat the maschine again,
Do you think 3d Go would be sufficient?
Or would the engine have no problems?
Or would this variant even to complex to human players?

3D go is not that ‘easy’ to implement as it seems. If you go for a regular grid for a board, suddenly every intersection has 6 liberties. However, having most intersections have exactly 4 liberties is quite fundamental to the complexity of go.
If you have one less, you get TriGo, which is still a proper game, but much simpler since most sente moves lead to long, forcing sequences which are comparatively easy to calculate.
When you have additional liberties, groups tend to be ‘trivially alive’, since chasing a stone, it can immediately extend and get 5 new liberties (regular 3D go being the example).
An interesting 3D version of go (imo) is Diamond go which you can look into if you’re interested. I’m not aware of any AI approaches there, though most intersections still have four liberties, just with different geometry, hence I would assume that the complexity scales in similar fashion to regular go.

1 Like

Current go bots can’t, AFAIK, play 3d go so you’d win by default. But once a version for 3d go is written and trained I’m pretty sure it will easily beat humans, and be further ahead than top humans than in regular 2d go. The level of top humans at 2d go is relatively close to perfection because we’ve had thousands of years of collective learning. Humans are still kyus at 3d go.


I think 3d go is a completely different game. Humans who are good at 2d go would need significant training to become good at 3d go. I think their skills at 2d go would not carry over well to 3d go.
Personally, I wouldn’t even be interested at putting much effort into it.

1 Like

Off the top of my head, I don’t see how you would get ladders in 3D so computers would be even stronger relative to us?

Fun fact, Arimaa is a game that was invented in 2003 and designed to be hard for computers to beat to humans. A computer program beat a top human a year before AlphaGo appeared. That program was written by David Wu, who also wrote KataGo.


The point of this modern wave of ai is to be versatile. I’d imagine very little coding (comparatively) will need done to switch something like leela zero over to 3d and then start training it. It will become “strong amateur” level decades faster than the humans would even take a shot at defining what a “strong amateur” would mean, let alone superhuman level.