Very interesting approach, I liked it. I was always wondering how sites attributed these brilliancies. I wonder if assessing them with Maia Chess instead of LC0 would yield results more in line with what humans would consider brilliancies.
Good experimental idea. (data analysis experiment, suggestion of brilliancy automatic definition approach).
I would suggest making a clear distinction, perhaps redundantly, to help lighten the cognitive load, so it can be used in a more critical manner along your reasoning, for many readers (I guess I might be meaning myself).
Clarify that the experience is to study a double engine scales set of measures, toward testing what brilliancy could be in some automatic terms, i.e. toward some more objective definition that could hold over very many games. Contrasting that with the historical development of brilliancy annotation attribution.
Both brilliancy notions, being susceptible to critical thinking in your study. For different reasons. I.e. interesting discussions, and furthering research.
I am also curious, if having access to the evaluation function for any input position in near futur, and policy as possibly 2 measures themselves, would allow somehow differentiating the nature of the surprise, w.r.t. so prefix sequence of moves, as part of knowledge or expectations (which might have some effect on the historical definition of brilliancy, as an evolving variable), or about some more position features pattern perception expectation of action.
Lc0 in theory use the state as the position. but I have heard that it has been somehow also considering some antecedent positions (not sure).. I think it is a good approach to use as many measure tapes as we can find to look at the chess game world. I look forward similar experiment about sharpness (which I figure could rely on chess being based on the current position only, and not be historical knowledge evaluation dependent).
Sorry for the assertive tone. I am practicing being confident in my statements. But in fact, I might be asking questions behind the suggestions. The assertive intent for concision and not burying with doubts.
I would suggest making a clear distinction, perhaps redundantly, to help lighten the cognitive load, so it can be used in a more critical manner along your reasoning, for many readers (I guess I might be meaning myself).
Clarify that the experience is to study a double engine scales set of measures, toward testing what brilliancy could be in some automatic terms, i.e. toward some more objective definition that could hold over very many games. Contrasting that with the historical development of brilliancy annotation attribution.
Both brilliancy notions, being susceptible to critical thinking in your study. For different reasons. I.e. interesting discussions, and furthering research.
I am also curious, if having access to the evaluation function for any input position in near futur, and policy as possibly 2 measures themselves, would allow somehow differentiating the nature of the surprise, w.r.t. so prefix sequence of moves, as part of knowledge or expectations (which might have some effect on the historical definition of brilliancy, as an evolving variable), or about some more position features pattern perception expectation of action.
Lc0 in theory use the state as the position. but I have heard that it has been somehow also considering some antecedent positions (not sure).. I think it is a good approach to use as many measure tapes as we can find to look at the chess game world. I look forward similar experiment about sharpness (which I figure could rely on chess being based on the current position only, and not be historical knowledge evaluation dependent).
Sorry for the assertive tone. I am practicing being confident in my statements. But in fact, I might be asking questions behind the suggestions. The assertive intent for concision and not burying with doubts.
@LauOnChess said in #2:
> Very interesting approach, I liked it. I was always wondering how sites attributed these brilliancies. I wonder if assessing them with Maia Chess instead of LC0 would yield results more in line with what humans would consider brilliancies.
I haven't thought about using Maia, it's certainly an interesting idea and I will look into it
> Very interesting approach, I liked it. I was always wondering how sites attributed these brilliancies. I wonder if assessing them with Maia Chess instead of LC0 would yield results more in line with what humans would consider brilliancies.
I haven't thought about using Maia, it's certainly an interesting idea and I will look into it
@dboing said in #3:
> Good experimental idea. (data analysis experiment, suggestion of brilliancy automatic definition approach).
>
> I would suggest making a clear distinction, perhaps redundantly, to help lighten the cognitive load, so it can be used in a more critical manner along your reasoning, for many readers (I guess I might be meaning myself).
>
> Clarify that the experience is to study a double engine scales set of measures, toward testing what brilliancy could be in some automatic terms, i.e. toward some more objective definition that could hold over very many games. Contrasting that with the historical development of brilliancy annotation attribution.
>
> Both brilliancy notions, being susceptible to critical thinking in your study. For different reasons. I.e. interesting discussions, and furthering research.
>
> I am also curious, if having access to the evaluation function for any input position in near futur, and policy as possibly 2 measures themselves, would allow somehow differentiating the nature of the surprise, w.r.t. so prefix sequence of moves, as part of knowledge or expectations (which might have some effect on the historical definition of brilliancy, as an evolving variable), or about some more position features pattern perception expectation of action.
>
> Lc0 in theory use the state as the position. but I have heard that it has been somehow also considering some antecedent positions (not sure).. I think it is a good approach to use as many measure tapes as we can find to look at the chess game world. I look forward similar experiment about sharpness (which I figure could rely on chess being based on the current position only, and not be historical knowledge evaluation dependent).
>
> Sorry for the assertive tone. I am practicing being confident in my statements. But in fact, I might be asking questions behind the suggestions. The assertive intent for concision and not burying with doubts.
My goal was to see if LC0's policy can be used to detect brilliancies, not to introduce a new scale. But I should have probably made my intention clearer.
Regarding the knowledge of prior moves, I'm also not sure if LC0 uses this, but it's an interesting idea to test in the future.
My posts on sharpness are already on my website: www.chess-journal.com/evaluatingSharpness1.html and www.chess-journal.com/evaluatingSharpness2.html
The first one is about the idea and a sharpness function I came up with and the second one compares this function to one that the LC0 used and also looks at the sharpness of different openings.
> Good experimental idea. (data analysis experiment, suggestion of brilliancy automatic definition approach).
>
> I would suggest making a clear distinction, perhaps redundantly, to help lighten the cognitive load, so it can be used in a more critical manner along your reasoning, for many readers (I guess I might be meaning myself).
>
> Clarify that the experience is to study a double engine scales set of measures, toward testing what brilliancy could be in some automatic terms, i.e. toward some more objective definition that could hold over very many games. Contrasting that with the historical development of brilliancy annotation attribution.
>
> Both brilliancy notions, being susceptible to critical thinking in your study. For different reasons. I.e. interesting discussions, and furthering research.
>
> I am also curious, if having access to the evaluation function for any input position in near futur, and policy as possibly 2 measures themselves, would allow somehow differentiating the nature of the surprise, w.r.t. so prefix sequence of moves, as part of knowledge or expectations (which might have some effect on the historical definition of brilliancy, as an evolving variable), or about some more position features pattern perception expectation of action.
>
> Lc0 in theory use the state as the position. but I have heard that it has been somehow also considering some antecedent positions (not sure).. I think it is a good approach to use as many measure tapes as we can find to look at the chess game world. I look forward similar experiment about sharpness (which I figure could rely on chess being based on the current position only, and not be historical knowledge evaluation dependent).
>
> Sorry for the assertive tone. I am practicing being confident in my statements. But in fact, I might be asking questions behind the suggestions. The assertive intent for concision and not burying with doubts.
My goal was to see if LC0's policy can be used to detect brilliancies, not to introduce a new scale. But I should have probably made my intention clearer.
Regarding the knowledge of prior moves, I'm also not sure if LC0 uses this, but it's an interesting idea to test in the future.
My posts on sharpness are already on my website: www.chess-journal.com/evaluatingSharpness1.html and www.chess-journal.com/evaluatingSharpness2.html
The first one is about the idea and a sharpness function I came up with and the second one compares this function to one that the LC0 used and also looks at the sharpness of different openings.
@jk_182 said in #5:
> My posts on sharpness are already on my website: www.chess-journal.com/evaluatingSharpness1.html and www.chess-journal.com/evaluatingSharpness2.html
> The first one is about the idea and a sharpness function I came up with and the second one compares this function to one that the LC0 used and also looks at the sharpness of different openings.
Don't worry about reposting here. It would make a complete series on lichess. Others routinely repost.
I will look later though.. because sharpness has been used a lot in many conversations I have been aware of on lichess forum. I think making definition refinement needs such attempts. But keeping the difference between the convention notions and the proposed implementation using such automatic tools, i.e. not merging the 2 too soon, or emphasizing the distinction so the experimental information flow becomes salient, would help allow some room of imagination, in otherwise different definition understandings of the same words. Part of the proposal is that new empirical definition as candidate refinement or better definition. But one has to let the audience consider both versions, and whether they are the same or not.
I meant 2 scales about the 2 engine types.. I know ELO in some uncharacterized engine pool seems to be without any doubt about the supremacy over whole game statistics in the specific engine tourneys constraints, but for me they are still goggles on chess. Each having their own scale. I agree the way I wrote, in hindsight could have been about policy versus evaluation. I would like to remind that ELO is about whole game outcomes, and that we only assume it means that the accuracy per position is uniform across all chess positions.
Thanks for confirming plausibility of that history dependency, or redefinition of state in lc0.
Edit: I see a lot more space control over in that journal, as writer.
> My posts on sharpness are already on my website: www.chess-journal.com/evaluatingSharpness1.html and www.chess-journal.com/evaluatingSharpness2.html
> The first one is about the idea and a sharpness function I came up with and the second one compares this function to one that the LC0 used and also looks at the sharpness of different openings.
Don't worry about reposting here. It would make a complete series on lichess. Others routinely repost.
I will look later though.. because sharpness has been used a lot in many conversations I have been aware of on lichess forum. I think making definition refinement needs such attempts. But keeping the difference between the convention notions and the proposed implementation using such automatic tools, i.e. not merging the 2 too soon, or emphasizing the distinction so the experimental information flow becomes salient, would help allow some room of imagination, in otherwise different definition understandings of the same words. Part of the proposal is that new empirical definition as candidate refinement or better definition. But one has to let the audience consider both versions, and whether they are the same or not.
I meant 2 scales about the 2 engine types.. I know ELO in some uncharacterized engine pool seems to be without any doubt about the supremacy over whole game statistics in the specific engine tourneys constraints, but for me they are still goggles on chess. Each having their own scale. I agree the way I wrote, in hindsight could have been about policy versus evaluation. I would like to remind that ELO is about whole game outcomes, and that we only assume it means that the accuracy per position is uniform across all chess positions.
Thanks for confirming plausibility of that history dependency, or redefinition of state in lc0.
Edit: I see a lot more space control over in that journal, as writer.
I like this idea. I think you are working off this definition of brilliant:
"A brilliant move is a move only a brilliant player would make. Even a merely good player would mess up the position."
This definition makes sense to me. And LC0 pure policy seems a decent stand-in for the non-brilliant player.
Perhaps should you should confirm that the "merely good" player wouldn't save the position in another way. One way to do that would be to make sure that ALL good moves are brilliant, if any are.
"A brilliant move is a move only a brilliant player would make. Even a merely good player would mess up the position."
This definition makes sense to me. And LC0 pure policy seems a decent stand-in for the non-brilliant player.
Perhaps should you should confirm that the "merely good" player wouldn't save the position in another way. One way to do that would be to make sure that ALL good moves are brilliant, if any are.
Great post, like the idea itself. #2 It will be really great to use maia here cause it almost thinks like a human. Also wanna know how chess.com ‘s linking brilliant moves - komodo, sf, torch ai or special algorithm? Ra2 is kinda cool <3
@LauOnChess said in #2:
> Very interesting approach, I liked it. I was always wondering how sites attributed these brilliancies. I wonder if assessing them with Maia Chess instead of LC0 would yield results more in line with what humans would consider brilliancies.
badgyal was trained on human games and maia is terribly weak so it wouldnt be good. it should be noted chess.com just considers a sacrifice near or at top engine move as a brilliant move.
> Very interesting approach, I liked it. I was always wondering how sites attributed these brilliancies. I wonder if assessing them with Maia Chess instead of LC0 would yield results more in line with what humans would consider brilliancies.
badgyal was trained on human games and maia is terribly weak so it wouldnt be good. it should be noted chess.com just considers a sacrifice near or at top engine move as a brilliant move.
what it seems you're finding is low policy moves, usually strong moves are difficult to find, and generally you'll find low policy on historical brilliant moves, because leela, similarly wouldn't see the point at 1 node, but perhaps at 1000