- Blind mode tutorial
lichess.org
Donate

Taking a closer look at a single game with engines

#2

WDL graphs and heatmaps look visually impressive.

WDL graphs and heatmaps look visually impressive.
#3

How exactly did you calculate the piece activity? You mentioned, that if the piece could attack the opponents half it was giving an additional score. How exactly was that additional value calculated?

How exactly did you calculate the piece activity? You mentioned, that if the piece could attack the opponents half it was giving an additional score. How exactly was that additional value calculated?
#4

@Not_Super_Mario said in #3:

How exactly did you calculate the piece activity? You mentioned, that if the piece could attack the opponents half it was giving an additional score. How exactly was that additional value calculated?

For each piece, I counted the squares it attacks. If a square is in the opponent's half, it's counted twice. If the square is adjacent to the opponent's king, it's counted a third time. The total sum of the squares for each piece is the activity. More details and examples here: https://chessenginelab.substack.com/p/calculating-piece-activity

The formula can certainly be optimised, but it's always difficult to say which activity value is "objectively more correct"

@Not_Super_Mario said in #3: > How exactly did you calculate the piece activity? You mentioned, that if the piece could attack the opponents half it was giving an additional score. How exactly was that additional value calculated? For each piece, I counted the squares it attacks. If a square is in the opponent's half, it's counted twice. If the square is adjacent to the opponent's king, it's counted a third time. The total sum of the squares for each piece is the activity. More details and examples here: https://chessenginelab.substack.com/p/calculating-piece-activity The formula can certainly be optimised, but it's always difficult to say which activity value is "objectively more correct"
#5

The piece activity and sharpness scores are pretty interesting. Is there a good way to calculate how easy the position is to play with white or black, or would this act the same as the sharpness score. For example, the evaluation is 0.0, but it's much more difficult for black to maintain equality than for white.

Also, does the WDL calculation take into account rating levels or live accuracy scores to make its predictions? i.e. for a winning king and pawn endgame, is the WDL the same for a 2500 as it is for a 500 rated player?

The piece activity and sharpness scores are pretty interesting. Is there a good way to calculate how easy the position is to play with white or black, or would this act the same as the sharpness score. For example, the evaluation is 0.0, but it's much more difficult for black to maintain equality than for white. Also, does the WDL calculation take into account rating levels or live accuracy scores to make its predictions? i.e. for a winning king and pawn endgame, is the WDL the same for a 2500 as it is for a 500 rated player?
#6

@SneakyPil said in #5:

The piece activity and sharpness scores are pretty interesting. Is there a good way to calculate how easy the position is to play with white or black, or would this act the same as the sharpness score. For example, the evaluation is 0.0, but it's much more difficult for black to maintain equality than for white.

The sharpness is more a measure of the difficulty of the position for both sides. Splitting it between white and black is interesting, but currently I'm unsure how this can be done.

Also, does the WDL calculation take into account rating levels or live accuracy scores to make its predictions? i.e. for a winning king and pawn endgame, is the WDL the same for a 2500 as it is for a 500 rated player?

The WDL comes from LC0, so from a strong engine and doesn't change depending on the player ratings. I tried to use settings to make LC0 weaker to estimate the WDL with a more human view, but the play of an 2000 engine isn't the same as the play of a 2000 player.

@SneakyPil said in #5: > The piece activity and sharpness scores are pretty interesting. Is there a good way to calculate how easy the position is to play with white or black, or would this act the same as the sharpness score. For example, the evaluation is 0.0, but it's much more difficult for black to maintain equality than for white. The sharpness is more a measure of the difficulty of the position for both sides. Splitting it between white and black is interesting, but currently I'm unsure how this can be done. > Also, does the WDL calculation take into account rating levels or live accuracy scores to make its predictions? i.e. for a winning king and pawn endgame, is the WDL the same for a 2500 as it is for a 500 rated player? The WDL comes from LC0, so from a strong engine and doesn't change depending on the player ratings. I tried to use settings to make LC0 weaker to estimate the WDL with a more human view, but the play of an 2000 engine isn't the same as the play of a 2000 player.
#7

Nice blog post!

I wonder about the activity measure.

Does it make sense to increase activity for squares on the opponents side? Intuitively it sounds like it does make sense. But what about a bishop fiancheto'ed King side, while both Kings are castled there and the bishop ends up being "very active" on a long diagonal pointing towards an empty Queen side with no action taking place there?

About adding extra score for piece activity around the King - I do not think that's a good idea at all. But okay, that's not a good idea from a perspective of engine development. Because you would have score for piece activity and also a score for king safety. And it's best to keep these various scores as orthogonal as possible - this is mostly for optimizing weights - the more orthogonal the easier and more predictable it is. For analysis like this it might be just fine.

Also the sharpness change plot might be better looking if you maybe also added a second axis to show the actual sharpness of the position as an area fill graph on the background - it would allow to contextualize the sharpness change - is the position so saturated in sharpness already or are they keeping the sharpness in balance.

Nice blog post! I wonder about the activity measure. Does it make sense to increase activity for squares on the opponents side? Intuitively it sounds like it does make sense. But what about a bishop fiancheto'ed King side, while both Kings are castled there and the bishop ends up being "very active" on a long diagonal pointing towards an empty Queen side with no action taking place there? About adding extra score for piece activity around the King - I do not think that's a good idea at all. But okay, that's not a good idea from a perspective of engine development. Because you would have score for piece activity and also a score for king safety. And it's best to keep these various scores as orthogonal as possible - this is mostly for optimizing weights - the more orthogonal the easier and more predictable it is. For analysis like this it might be just fine. Also the sharpness change plot might be better looking if you maybe also added a second axis to show the actual sharpness of the position as an area fill graph on the background - it would allow to contextualize the sharpness change - is the position so saturated in sharpness already or are they keeping the sharpness in balance.
#8

@SneakyPil said in #5:

live accuracy scores

?

Otherwise, very good question. I think WDL is based on LC0 continuations when it has reached some kind of good enough, time to go meet the other engines in the horse race, we the trainers can't wait any longer, it has to be good enough (kidding, but some truth in there). So, while the author might answer better, I would say that no, it would not include that more realistic chess information. (but maybe it could.. maia tried but not as you are thinking, which would be the full continuous joint probablility problem that they could have considered).

Rating systems do consider that though, but only about the whole game outcome. Maia and Lc0 look at positions.

@SneakyPil said in #5: > live accuracy scores ? Otherwise, very good question. I think WDL is based on LC0 continuations when it has reached some kind of good enough, time to go meet the other engines in the horse race, we the trainers can't wait any longer, it has to be good enough (kidding, but some truth in there). So, while the author might answer better, I would say that no, it would not include that more realistic chess information. (but maybe it could.. maia tried but not as you are thinking, which would be the full continuous joint probablility problem that they could have considered). Rating systems do consider that though, but only about the whole game outcome. Maia and Lc0 look at positions.
#9

@likeawizard said in #7:

Does it make sense to increase activity for squares on the opponents side? Intuitively it sounds like it does make sense. But what about a bishop fiancheto'ed King side, while both Kings are castled there and the bishop ends up being "very active" on a long diagonal pointing towards an empty Queen side with no action taking place there?

About adding extra score for piece activity around the King - I do not think that's a good idea at all. But okay, that's not a good idea from a perspective of engine development. Because you would have score for piece activity and also a score for king safety. And it's best to keep these various scores as orthogonal as possible - this is mostly for optimizing weights - the more orthogonal the easier and more predictable it is. For analysis like this it might be just fine.

I added the extra activity for pieces attacking squares around the opponent's king to make long range pieces more active if they aim at the castled king instead of the other side of the board. I'm unsure how to deal with a bishop aiming at an empty queenside. In some sense controlling a8 might be useless but it could also prevent some potential manoeuvres from the opponent's pieces. The activity score can certainly be improved by adjusting the weights for the different squares and adding more nuance to the calculation. It's only difficult to specify what makes the score "better".
As for the issue with counting squares around the king higher, I tried to approach the problem from a more human perspective. So I didn't worry about different concepts being orthogonal. I just thought that from a human viewpoint, a piece attacking the opponent's king should be considered to be more active than a piece controlling the same number of squares but on the other side of the board.

Also the sharpness change plot might be better looking if you maybe also added a second axis to show the actual sharpness of the position as an area fill graph on the background - it would allow to contextualize the sharpness change - is the position so saturated in sharpness already or are they keeping the sharpness in balance.

Thanks for the suggestion, I'll add it in the future.

@likeawizard said in #7: > Does it make sense to increase activity for squares on the opponents side? Intuitively it sounds like it does make sense. But what about a bishop fiancheto'ed King side, while both Kings are castled there and the bishop ends up being "very active" on a long diagonal pointing towards an empty Queen side with no action taking place there? > > About adding extra score for piece activity around the King - I do not think that's a good idea at all. But okay, that's not a good idea from a perspective of engine development. Because you would have score for piece activity and also a score for king safety. And it's best to keep these various scores as orthogonal as possible - this is mostly for optimizing weights - the more orthogonal the easier and more predictable it is. For analysis like this it might be just fine. I added the extra activity for pieces attacking squares around the opponent's king to make long range pieces more active if they aim at the castled king instead of the other side of the board. I'm unsure how to deal with a bishop aiming at an empty queenside. In some sense controlling a8 might be useless but it could also prevent some potential manoeuvres from the opponent's pieces. The activity score can certainly be improved by adjusting the weights for the different squares and adding more nuance to the calculation. It's only difficult to specify what makes the score "better". As for the issue with counting squares around the king higher, I tried to approach the problem from a more human perspective. So I didn't worry about different concepts being orthogonal. I just thought that from a human viewpoint, a piece attacking the opponent's king should be considered to be more active than a piece controlling the same number of squares but on the other side of the board. > Also the sharpness change plot might be better looking if you maybe also added a second axis to show the actual sharpness of the position as an area fill graph on the background - it would allow to contextualize the sharpness change - is the position so saturated in sharpness already or are they keeping the sharpness in balance. Thanks for the suggestion, I'll add it in the future.
#10
<Comment deleted by user>