search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.56k stars 564 forks source link

how to use kataGo to analyse a kifu that if played by a ai or person? #290

Open minikiller opened 4 years ago

minikiller commented 4 years ago

Dear Sir: Now there are many persons to play Go game helped by kataGo or leela zero on the net, my problem is how to use Ai program to analyse a kifu to make a decision that kifu is played by AI? Thanks!

lightvector commented 4 years ago

No idea. This is not research that I've done myself, nor is it something that I have on my todo list in the near future. It might well be valuable for the Go community to have some careful and thorough research here, but someone has to do it. Are you interested in doing it? :)

Any such research probably should also consider the existence of many older bots with very different styles (like Zen 7) that are "AI" but whose styles would be considered very non-"AI-like" nowadays in the sense that they are very different than modern self-play-trained bots and might have lower move-for-move match rates, but that would be still more than strong enough to win games against amateur humans. It seems like one might also want to leverage a variety of bots - not just KataGo.

egri-nagy commented 4 years ago

Hi, We are working on it (others might as well), finishing a draft paper in a week or so. Tough problem, do not expect a clear-cut solution.

egri-nagy commented 4 years ago

Here is the preprint:

https://arxiv.org/abs/2009.01606

Short summary: you need an expert player paired with a strong AI.

See here as well: https://youtu.be/-cEL7I6BWTc?t=3909