search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.49k stars 564 forks source link

Known phenomena? #715

Open michito744 opened 1 year ago

michito744 commented 1 year ago

The following game record is the actual game of the "Two-Headed Dragon". KataGo_sample_20221217_0001.txt

White is sure to live because wins against the ko of A. 2022-12-17

In order to know what KataGo does not understand, I let the game proceed under the assumption that White moved to a. The result was unbelievable: KataGo could not recognize that the white stone in the center was dead until it was down to two liberties. It did not understand life and death at all.

lightvector commented 1 year ago

Yes, this is known. All known AlphaZero-trained resnets have issues evaluating two-headed dragons, similar to ladders, it appears to be something that AlphaZero bots consistently fail to learn without intervention. I'm currently already experimenting with special training to see if it can be improved.

lightvector commented 1 year ago

By the way, thanks for the game record! I can also use it as a training sample too for such positions! :)

michito744 commented 1 year ago

@lightvector Thanks for the reply.

I understand the above. I see that KataGo does not understand the meta-structure of "the opponent's living stones can exist in an area that is completely enclosed by myself - by doing so, own moving prohibition point is taken away".

By the way, it would be great if you could also check the seki collapse in #705. That example also misinterprets life and death until the number of liberties is decisively reduced.