Great job!
Interesting analysis as far as it goes, but I was a bit shocked to click on a headline "opening stats for the 2024 Candidates" and not see any stats at all about which openings were played.
And that probably applies to 99% of the people who clicked the link.
Perhaps "Opening Novelty Rates for the 2024 Candidates" would be a better title.
And that probably applies to 99% of the people who clicked the link.
Perhaps "Opening Novelty Rates for the 2024 Candidates" would be a better title.
@Aighearach said in #3:
> Interesting analysis as far as it goes, but I was a bit shocked to click on a headline "opening stats for the 2024 Candidates" and not see any stats at all about which openings were played.
>
> And that probably applies to 99% of the people who clicked the link.
>
> Perhaps "Opening Novelty Rates for the 2024 Candidates" would be a better title.
Good call, thanks for the suggestion :)
> Interesting analysis as far as it goes, but I was a bit shocked to click on a headline "opening stats for the 2024 Candidates" and not see any stats at all about which openings were played.
>
> And that probably applies to 99% of the people who clicked the link.
>
> Perhaps "Opening Novelty Rates for the 2024 Candidates" would be a better title.
Good call, thanks for the suggestion :)
is there a way to define novelty based on the position information, not just the move sequence?
And not use some surrogate non terminal position evaluation, but game result. That might take more games to fill in the statiscial questions with enough position information visiting over the chess positoin wilderness.
And then one might use that slight tasking pressure (W D L), to test various metrics in the positon world, of the notion of novelty.
I would not go occam razor in that directoin, maybe yes, but not in chess complexity on the board, more about the formalism of the metrics and probablity law being set up. One might go straight to A0 pre-implementation position world (also the input planes might do, and then use any metric necessary (means hypothesis research, I known I known, not tangible technology yet, but technological instruments maybe to be found for later technology, ... wait, am I talking about chess?).
And not use some surrogate non terminal position evaluation, but game result. That might take more games to fill in the statiscial questions with enough position information visiting over the chess positoin wilderness.
And then one might use that slight tasking pressure (W D L), to test various metrics in the positon world, of the notion of novelty.
I would not go occam razor in that directoin, maybe yes, but not in chess complexity on the board, more about the formalism of the metrics and probablity law being set up. One might go straight to A0 pre-implementation position world (also the input planes might do, and then use any metric necessary (means hypothesis research, I known I known, not tangible technology yet, but technological instruments maybe to be found for later technology, ... wait, am I talking about chess?).
Doing a little statistical analysis would be good.
To distinguish a "novelty" from an "inaccuracy", you might also consider the evaluation after the opponent responds. For example, the criteria might be that a move is considered a "novelty" if:
1. The move is new*
2. The change in the engine's evaluation of the game position is no worse than a loss of 70 centipawns**.
3. The engine evaluation goes above X within the next Y plys. Here, X and Y would be two more arbitrary parameters that you need to set, but it would give an indication of whether the move was a good practical move. I.e., maybe the engine evaluation was based on a line that isn't playable to a human.
*Even "new" is hard to define. It could be that the resulting position does not occur in the historical database, or that the move has never been played in the prior position. Those are not necessarily the same because of transposition.
** At least I think this is what you meant in your article. You could also base it on the absolute evaluation of the position. E.g., the person who played the move is at no more than a 70 centipawn disadvantage after playing the move.
1. The move is new*
2. The change in the engine's evaluation of the game position is no worse than a loss of 70 centipawns**.
3. The engine evaluation goes above X within the next Y plys. Here, X and Y would be two more arbitrary parameters that you need to set, but it would give an indication of whether the move was a good practical move. I.e., maybe the engine evaluation was based on a line that isn't playable to a human.
*Even "new" is hard to define. It could be that the resulting position does not occur in the historical database, or that the move has never been played in the prior position. Those are not necessarily the same because of transposition.
** At least I think this is what you meant in your article. You could also base it on the absolute evaluation of the position. E.g., the person who played the move is at no more than a 70 centipawn disadvantage after playing the move.