Mixed up engine

@jomega
@dboing

I rather like the idea of objectively classified errors, with the threefold terminology. In education generally, what counts as a minor or a major error is not typically up to the student, but the teacher or professor. One would think that chess especially is akin to logic or mathematics and lends itself to objective evaluation. I suppose the difference is that chess is a game with opposing players, and engine evaluations assume a hypothetical opponent who is far stronger than any human, let alone the humans most of us are likely to play against. How often do we win by committing an inaccuracy (as in the game I posted) or a mistake, or occasionally even a blunder?

@Zubbubu said in #41:
> and engine evaluations assume a hypothetical opponent who is far stronger than any human, let alone the humans most of us are likely to play against.

I think the problem is that engine evaluation are scoring compared to perfect continuation from the position being evaluated. So that the number of errors, is not really progressively adapted to any of the player levels. There is no sense of student progression in there. I don't know how lichess winning odds correction might be related to that. But counting all errors without context, is likely to be information overload, and not conducive to wanting to work all of them , so which fewer.

Does the student want to be perfect right away or just improve? I may not have understood. In some way yes opponent far stronger but also player far stronger, both perfect (SF version of what that is, actually, within the chess tree it explores).

The lichess winning odds is also not adapted to player pair. it also is about best play on both sides. If both perfect such position would have such winning chances at legal termination of game (W,D,L the chess game 3 possible outcomes).

Maia has tried to model human pair behavior, and generated curves about winning odds in that context. One does not need the maia engines, but the preparatory work characterizing all levels of pairing (some rough binning though) I don't recall the details though. But there were well behaved curves based on lichess data, that span many levels of plays. My point being that those could serve as an adaptive goal of accuracy improvent basis. The expected level of error for that pair average could be the referential. how much a student would like to improve with respect to own level, per session (or game) or set of games, all paces of learning might be attempted...perhaps.

I had a situation where the evaluation goes from 4.5 to 7.2 and in the analysis it calls it an Inaccuracy. A 2.7 swing seems more than an inaccuracy, but this seems similar to your situation. The game is below.