pytorch / ELF

ELF: a platform for game research with AlphaGoZero/AlphaZero reimplementation
Other
3.37k stars 567 forks source link

The weight seems not to learn Go rule. Why? #143

Closed y-ich closed 4 years ago

y-ich commented 5 years ago

Thank you for your great work.

I tried to let ELF OpenGo to read a famous ladder problem and found that ELF OpenGo weight v2 cannot suggest the move to pickup ladder stones.

52990846-744f7000-344d-11e9-9676-a817fb9a760c

(Actually I used the converted weight with Leela Zero, not the original ELF weight)

I understand that learning whether no liberties or not of large stone string is extremely harder than learning whether illegal or not in surrounded intersection for CNN. But I have no idea to improve this situation.

Isn't 20-block ResNet enough? Does 40-block ResNet solve it? Or do you need full-connected layer to learn liberties of large stone string?

I want to hear opinions of ELF team since your paper picked up the ladder issue.

y-ich commented 4 years ago

https://github.com/leela-zero/leela-zero/issues/2445