Resignation: add resignation mechanism to reduce compute time

[barakugav]

No description provided.

[barakugav]

from alpha zero:

In order to save computation, clearly lost games are resigned. The resignation threshold v_resign is
selected automatically to keep the fraction of false positives (games that could have been won if AlphaGo had not resigned) below 5%. To measure false positives, we disable resignation in 10% of self-play games and play until termination.

[barakugav]

Even if a game is clearly lost, the search still yield good training data for the moves probabilities head.
Not sure we should invest in it, as it saves computation, but i think the training data produced is still reasonable.

[poja]

Interesting.... I would guess that it could be a significant improvement in later stages of training but not in early and early-medium stages - because the endgame is critical for "understanding the game". How much percentage of the game do you think this can save on average? 30-70%?

…

On Mon, Nov 7, 2022 at 10:34 AM Barak Ugav ***@***.***> wrote: Even if a game is clearly lost, the search still yield good training data for the moves probabilities head. Not sure we should invest in it, as it saves computation, but i think the training data produced is still reasonable. — Reply to this email directly, view it on GitHub <#50 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABDHPIQGI5DSJ6UJOHSLPWDWHC5JTANCNFSM6AAAAAAQLCHUHM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[barakugav]

Its very hard to say.
If the players are good, big advantage should lead to quick victory in most cases, so i dont think 70%.
And in early stages we need the end game.
I think we should abandon this feature... also it is not trivial to implement