Another resign condition? about reversi-alpha-zero HOT 4 CLOSED

No I am not saying "resign by white" by black, but something like both players agree to stop the game under that condition. Maybe "resign" is not a correct word here, but it does the similar thing with alphago zero: stop the game earlier and keep training data simpler. And we can still keep same "false positive adjustment" mechanism.

About the MCTS: Resigning is self-play thing. When in evaluate/play_with_human, AI don't have resign mechanism, AI have to play till the end. But as AI resigned to avoid NN training on that situation, so MCTS is the only thing that can lead to final win.

Never mind, I can test locally and update the result later.

from reversi-alpha-zero.

Update: I realize that it is already implicitly done in current resign design: when selfplay, since both players are using same model, when one side's model predicts the state value is lower than threshold, than the other side will predict the state value larger than -threshold, and vice versa.

from reversi-alpha-zero.

@gooooloo

I can not clearly point out the problem on play data generation, but I feel a bit strange.

1. even if black is almost winning, it is strange to decide "resign by white" by black.
2. If we want to implement "think very deep(ex. search num to 1600) before resigning", that logic becomes an obstacle.

from reversi-alpha-zero.

Close it for now. Once I have result later I will update.

from reversi-alpha-zero.