lichess.org
Donate

The xG of Chess: Shark Points

Great one! Probably my favorite blog of all Time! Thanks, it's interesting !

Great one! Probably my favorite blog of all Time! Thanks, it's interesting !

Wow, actually relevant and interesting analysis on a lichess blog! Good work. This is a cool concept.

Wow, actually relevant and interesting analysis on a lichess blog! Good work. This is a cool concept.

This is my favorite article I've read on here, super interesting idea. I wonder if there's an inverse of this - an ability to swindle their way into a draw or win when they're defending a worse position.

This is my favorite article I've read on here, super interesting idea. I wonder if there's an inverse of this - an ability to swindle their way into a draw or win when they're defending a worse position.

I like the idea of building metrics with some angles of pressure like that. But here while it might not matter for the mechanics and usage, I am not sure to get the direction of "being a shark". Is it about turning around in to find the source of blood before striking. like toying with the food? from other points of view, or the point of view of the food?

Because you are talking about delaying conversion, until sure to strike. Enjoying the control of the situation...

Conversion meaning from positional features advantage (anything of the board that has not been consumed materially, i guess would be a robust definition of positional, so even a tactical mechanism left in tension mode.... with some current gain, but wait, maybe it could be leverage even more... things like that).

There is the other human "shark" behavior which is more about callousness and only there for the win, and taking all opportunities. One could talk about cheaters having a shark mentality, whether with engine, or just how they might surf on some rating technicalities Not saying there are cheaters, just talking about the notion of sharky.. No affect. just hungry.

here different notion... about the method of approach... (it could have been cat approach.... but I guess "shark" looks like more pointee teeth.. more aggressive, in line with some sport motivational self-priming.. Give no quarters.... I am derailing.. had forgotten the sport introduction.. Footbal and cat behavior.... hmm. does not compute! not a winner motivation hyperbole.

I like the idea of building metrics with some angles of pressure like that. But here while it might not matter for the mechanics and usage, I am not sure to get the direction of "being a shark". Is it about turning around in to find the source of blood before striking. like toying with the food? from other points of view, or the point of view of the food? Because you are talking about delaying conversion, until sure to strike. Enjoying the control of the situation... Conversion meaning from positional features advantage (anything of the board that has not been consumed materially, i guess would be a robust definition of positional, so even a tactical mechanism left in tension mode.... with some current gain, but wait, maybe it could be leverage even more... things like that). There is the other human "shark" behavior which is more about callousness and only there for the win, and taking all opportunities. One could talk about cheaters having a shark mentality, whether with engine, or just how they might surf on some rating technicalities Not saying there are cheaters, just talking about the notion of sharky.. No affect. just hungry. here different notion... about the method of approach... (it could have been cat approach.... but I guess "shark" looks like more pointee teeth.. more aggressive, in line with some sport motivational self-priming.. Give no quarters.... I am derailing.. had forgotten the sport introduction.. Footbal and cat behavior.... hmm. does not compute! not a winner motivation hyperbole.

Also, chess being big, I agree that one would need many metrics.. Why does LC0 have so many neurons.. and such an exploded input space. And why there were so many components in the SF leaf classical evaluation. While SF is erring on the cheap end of position evaluation, the cheapest that keeps it on top of certain pools of other engine ecosystem, it is also going toward improving its leaf evaluation quality. It might end up transferring components that were in the classical hand crafter as reduced features set (compared to LC0 input as NN) with discrerning power, as it keep going deeper or preferably more complete).

I like the metric direction, compared to a mere 1D score advantage difference.. The more dimensions we consider the more we might be able to start asking the question of how big is chess, awas from just how is one players better number than another on average of its pool of games over some duration. We can start to have some handle on the world of chess positions. Quantitative (and therefore also qualitative).

just I would not pile up too much of metrics on the one engine type output... Propagation of biases and error kind of prudence of mine.

Also, chess being big, I agree that one would need many metrics.. Why does LC0 have so many neurons.. and such an exploded input space. And why there were so many components in the SF leaf classical evaluation. While SF is erring on the cheap end of position evaluation, the cheapest that keeps it on top of certain pools of other engine ecosystem, it is also going toward improving its leaf evaluation quality. It might end up transferring components that were in the classical hand crafter as reduced features set (compared to LC0 input as NN) with discrerning power, as it keep going deeper or preferably more complete). I like the metric direction, compared to a mere 1D score advantage difference.. The more dimensions we consider the more we might be able to start asking the question of how big is chess, awas from just how is one players better number than another on average of its pool of games over some duration. We can start to have some handle on the world of chess positions. Quantitative (and therefore also qualitative). just I would not pile up too much of metrics on the one engine type output... Propagation of biases and error kind of prudence of mine.

If I understand correctly, what you propose is to eliminate the concepts of blunders and mistakes and dubious moves and instead rely on a continuous metric that is calculated based on nonlinear transform of the evaluation.

But isn't that exactly what the analysis charts are? They have to use nonlinear transform because the eval can be very large, so that's what eval charts and gauges are usually showing. And the end value... shark points, how is it different from the accuracy calculation?

I am sure you know what you're talking about, I just don't get it at the moment. Can you go into the math of your metric and how it differs from the ones we already have?

If I understand correctly, what you propose is to eliminate the concepts of blunders and mistakes and dubious moves and instead rely on a continuous metric that is calculated based on nonlinear transform of the evaluation. But isn't that exactly what the analysis charts are? They have to use nonlinear transform because the eval can be very large, so that's what eval charts and gauges are usually showing. And the end value... shark points, how is it different from the accuracy calculation? I am sure you know what you're talking about, I just don't get it at the moment. Can you go into the math of your metric and how it differs from the ones we already have?
<Comment deleted by user>
<Comment deleted by user>
<Comment deleted by user>