Use the v6 field played_q to do a more direct blunder rescoring

[Naphthalin]

Draft for how to use the new field played_q from v6 training data LeelaChessZero#1468 in rescoring:

• compare best_q and played_q
• if their difference is above a threshold, a proven win was missed, or a proven loss played without being forced, rescore all positions until that position
• rescore q, d and m (last one is still missing)

There is currently no code dealing with the special case of playing a longer/shorter proven line which gives inaccurate m but accurate q and d.

In general, the current code base is split between the final position where no consecutive position exists, and all other positions for which a consecutive position exists. I suppose this split isn't needed for the v6 rescoring following https://github.com/LeelaChessZero/lc0/issue/1308/ though I worked with it here.

[masterkni6]

Is the rescorer supposed to change result_q and result_d of the blunder move to best_q and best_d? I don't really understand the training all too well but doesn't that make it think the blunder is a good move?

[Naphthalin]

Is the rescorer supposed to change result_q and result_d of the blunder move to best_q and best_d? I don't really understand the training all too well but doesn't that make it think the blunder is a good move?

There might be a small misconception on your side about the training: The Lc0 training trains on positions, not on played moves; the played move is only relevant for the game outcome. Changing result_q/result_d when a blunder occurs is the equivalent of a human "experimenting" in a position, i.e. "the current position is slightly better for white. This move looks interesting, let's see what happens even though I think it's a bad move." It doesn't make the blunder itself look good, but the position where the blunder happened because of temp exploration.
Does that make sense?

[Tilps]

Is the rescorer supposed to change result_q and result_d of the blunder move to best_q and best_d? I don't really understand the training all too well but doesn't that make it think the blunder is a good move?

No - it avoids having the game outcome after the blunder, affect the training target before the blunder.
Nothing about that causes the policy of the bad move to increase, or for the training target after the blunder to be changed. (Unless there are multiple blunders in the game... but the logic still holds there.)
The main risk of this change is deblundering too often per game. Unless there are runs of multiple positions between each blunder - this change almost becomes like training on a high q ratio. Which is known to be a tiny bit weaker than what we already have. The hypothesis is that there exists some happy medium where runs of training target are left unchanged far enough that learnings propagate back through the game effectively, while not also learning to expect blunders so much.