search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.33k stars 546 forks source link

Issue in training: low visit counts and strange initial conditions(board dimensions, komi settings, piece position) #947

Open Centaurea-Platinum opened 2 weeks ago

Centaurea-Platinum commented 2 weeks ago

I recently began to contribute to Katago distributed training. I noticed that the network is trained on strange initial board/komi conditions and are running on low visit counts. Is the strange initial board/komi conditions to prevent overfitting? Will the low visit count negatively effect how the network run on high visit counts?

lightvector commented 2 weeks ago

The variety of initial boards is for variety. Many initial positions are from random human games, to ensure that KataGo is familiar with a variety of the positions that arise from human games to be good at analyzing them, not just the more limited set of openings and josekis that would occur in pure AI vs AI games. A lot of the variety of komi values is to balance some fraction of these games to be fair given that they don't start in equal positions, some other parts of komi variety is to ensure that KataGo has experience with a wide range of komi (similarly for board sizes and rules).

Low visit count is simply a necessity - if we increased the visit count a lot the training data might be better, but it would be impossible to generate enough training data.

lightvector commented 2 weeks ago

Anyways, welcome and thanks for the contribution! If you do notice any particular strange games or results, feel free to link more specific cases. It wouldn't be the first time that someone manually looking has found something wrong with the data, which has led to the discovery of a bug or some other improvement.