search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.57k stars 564 forks source link

I don't really understand the point of your new 9-way ai release. #853

Open awsjgy opened 1 year ago

awsjgy commented 1 year ago

It's all down the 19 now, so what's the point of you releasing the 9 engine, please?

awsjgy commented 1 year ago

I know a Chinese guy trained himself a custom rule before: white counts as a win if he makes a live piece. Are your weights trained by it? Are you planning to train dead-live games in the future? All black eats up all white wins, how many discs do you let black have the most?

OmnipotentEntity commented 1 year ago

The latest release is a network trained specifically on 9x9 boards, so it will do better on 9x9 than the normal 19x19 network. This isn't an engine, it's just a network for the engine. The normal KataGo engine will run both the 9x9 network and the standard 19x19 network. The point of releasing the 9x9 network was so that other people can use it. It was used to create the 9x9 book. https://katagobooks.org/

The 9x9 network was trained on 4 handicap using the rules you describe (about a 78 point komi, iirc).

awsjgy commented 1 year ago

I have a couple questions, can the weights of the board paths other than this 9-way weight be released for our research? What is the approximate difference between the strong 200 elo. Can this 9-way weighting be played as a handicap? Or is the split-first stronger and the handicap not as strong。This latest 9-way weight, which is just used to study the opening pool, still performs better than a regular katago on the 9-way

awsjgy commented 1 year ago

The latest release is a network trained specifically on 9x9 boards, so it will do better on 9x9 than the normal 19x19 network. This isn't an engine, it's just a network for the engine. The normal KataGo engine will run both the 9x9 network and the standard 19x19 network. The point of releasing the 9x9 network was so that other people can use it. It was used to create the 9x9 book. https://katagobooks.org/

The 9x9 network was trained on 4 handicap using the rules you describe (about a 78 point komi, iirc). I have a few questions, this 9 way weight specializes in 9 way openings, so why would you go and try to do dead-end problems and eat all the white. And is it possible to play let's play around the empty space? What is the right setting for this breadth and aggressiveness

ChinChangYang commented 1 year ago

https://github.com/lightvector/KataGo/releases/tag/v1.13.2-kata9x9

One fun feature is that this net also has a little bit of training for 9x9 handicap games, including the "game" where white has a 78.5 komi** while black has 4 or 5 handicap stones, such that white wins if they live basically anywhere. This training did not reach convergence, but enough that if you try searching with a few millions of playouts, the results are pretty suggestive that white can live if black starts with all four 3-3 points, but not if black gets a fifth stone anywhere reasonable.

awsjgy commented 1 year ago

I'm sorry to say that I've found that the official 18b weights in general are exactly the same as this particular 9-way weight, with exactly the same trick picks. There is no stronger