Jon12345 wrote: ↑Mon Jul 12, 2021 5:25 pmBut we seem to be going around in circles here. If the objective result is draw, win or loss, why have engines evaluating in centipawns? Why even say, "White is better", or White is 1.2+?
Because engines do not give an 'objective measure' of the positions any more than GMs do. They are just strong players, that also make losing blunders occasionally. And their evaluation in centi-Pawn is usually wrong. Which they admit themselves: just let them search a few moves deeper, and they will give you a different score. Occasionally a very different score.
Centi-Pawn scores are not objective. They are the subjective opinion of an entity that is guessing after the statistics of the outcome against a group of equally fallible opponents.
I see what you are saying now. My idea has no merit because you cannot see the position calculated to the end. You are right, it is a waste of time to try to estimate the strength of a move because it is not possible using the methods that I and Ferdy have suggested. In fact, chess commentators should refrain from saying someone played a weak move, or a strong move, because they are not objective. Chess.com and lichess should remove their chart showing evaluation throughout the game. Chess books should remove !! ! !? ?! ??. Openings should not be ranked because no one knows.
Unless the evaluation is not 100% accurate, it is all a waste of time. I get it now. Thank you.
One thing you obviously overlook is that none of the parties you mention claims their judgement to be objective, and neither is their audience interested in an objective judgement. They are happy with the subjective judgement of a party whose judgement they trust. Neither do they attempt to assign Elo point to moves. You, however, want to classify GM quality moves that happen to be losing in a way that no human would ever discover as 'poor' on the Elo scale, because they are objectively losing. Apart from the fact that it is in practice not possible to make that judgement for most positions, I don't think there would be much interest in that from a user perspective.
hgm wrote: ↑Mon Jul 12, 2021 3:29 pm
If your goal is to assign the Elo of the player that would typically play the move to that move (as I recall you said it was)... Then yes, a move that typically would be played by a 2400-rated GM should get rated 2400, and a move like 2 Qh5 typically played by a 400-Elo patzer should get rated 400. Irrespective of how Stockfish scores it, or what it thinks would be the best move.
I think that the goal to assign the elo of the player that would typically play a move is something that you can never achieve.
The problem is that there are moves that only weak players or strong players are going to make.
For example there may be a capture that win material if you look one ply forward but lose material based on what a typical 2000 player see when only strong engines may calculate deeper and see that after many plies that the capture is winning
so basically players with rating 1000 or 3000 are expected to play the right move but not players with rating 2000.
If you don't get it already, you will never get it. You are too concerned with proving your original objection than seeing reality. I never let ego get in the way if I am trying to solve a problem, because it will impede progress.
You can look at it in two ways. "Here's how you do it" or "It can't be done." I belong to the former, you the latter. But the maths and equations are on my side.
Uri Blass wrote: ↑Mon Jul 12, 2021 6:26 pm
I think that the goal to assign the elo of the player that would typically play a move is something that you can never achieve.
The problem is that there are moves that only weak players or strong players are going to make.
For example there may be a capture that win material if you look one ply forward but lose material based on what a typical 2000 player see when only strong engines may calculate deeper and see that after many plies that the capture is winning
so basically players with rating 1000 or 3000 are expected to play the right move but not players with rating 2000.
My goal is to estimate the playing strength of a move. There is a strong correlation between that estimated playing strength and a players strength. On average, a stronger player plays stronger moves than a weaker player. They are more likely to play the engines move than the weaker player. This correlation means that the predictability of the estimated elo rating of a move is a reasonable guesstimate.
Do you have an example position to illustrate your point? Is your position that 2000 rated players make more mistakes than 1000 rated players, or that on occasion, the 2000 rated player will make a mistake on a move while the 1000 rated player won't? The objective value of the move is the same, whoever makes it. The predictive nature of the elo strength of the player making the move will have strong positive correlation, will it not?
An engines evaluation rating strongly predicts the likely outcome of a game. The equations are based on the engines evaluations. You can say sometimes a +10 evaluation leads to a lose. But most of the time it doesn't. We are back to the "its futile if its not 100% accurate" argument.
Last edited by Jon12345 on Mon Jul 12, 2021 6:52 pm, edited 1 time in total.
It is not possible for me to say whether 'it' can't be done, or how to do 'it', because so far you have been elusive and contradictory about what 'it' actually is.
But one thing is certain: you cannot measure the quality of a move (or in fact anything) by engine score, because there is no such thing as an engine score, unless it is a converged mate score. In all other situations the engine won't report a single score, but give you a (potentially) different score for each depth iteration. Which, in a won position, usually keeps rising with depth, and in a drawn position creeps towards zero with increasing depth. And in difficult tactical positions often suddenly switch from very high to very low from one depth iteration to the next. I have seen Stockfish evaluate dead lost positions as +8.5, and it takes an extremely long analysis before it turns negative.
Because it seems a good idea at the time. What they play will depend on how deep they search. Why do you think engines perform better when you give them more time to think (or, equivalently, run them on a faster computer)?
Why does it seem like a good idea (quality move) at the time if the engine score does not predict move quality?
Edit: I notice you want to use the term "good idea" instead of the interchangeable "quality move", because it then becomes apparent you are contradicting yourself.
