This article didn't tell me anything genuinely new or interesting, it's just a proposal about how getting out of opening theory and what it does to ELO. That's not very interesting because there are better ways to get out of opening theory than blundering an exchange. For example, you could just play 2.Ra2, and your opponent would be just as clueless to opening lines. Speaking of which, the importance you put on openings is too much. Openings matter very little in the grand scheme of who wins a chess game, at least for amateurs or above. Additionally, deep preparation into 2.Ra3 is very unlikely to lead to an advantage for white because white rarely gets genuine counterplay, and black has many variations to choose from so white will also likely be out of theory by move 6. In blitz especially, a -2 disadvantage is the far from lost when considering the wildness that a blitz game can turn out to be.
@nickpierce88 said in #21:
> This article didn't tell me anything genuinely new or interesting, it's just a proposal about how getting out of opening theory and what it does to ELO. That's not very interesting because there are better ways to get out of opening theory than blundering an exchange. For example, you could just play 2.Ra2, and your opponent would be just as clueless to opening lines. Speaking of which, the importance you put on openings is too much. Openings matter very little in the grand scheme of who wins a chess game, at least for amateurs or above. Additionally, deep preparation into 2.Ra3 is very unlikely to lead to an advantage for white because white rarely gets genuine counterplay, and black has many variations to choose from so white will also likely be out of theory by move 6. In blitz especially, a -2 disadvantage is the far from lost when considering the wildness that a blitz game can turn out to be.
Thanks!
> This article didn't tell me anything genuinely new or interesting, it's just a proposal about how getting out of opening theory and what it does to ELO. That's not very interesting because there are better ways to get out of opening theory than blundering an exchange. For example, you could just play 2.Ra2, and your opponent would be just as clueless to opening lines. Speaking of which, the importance you put on openings is too much. Openings matter very little in the grand scheme of who wins a chess game, at least for amateurs or above. Additionally, deep preparation into 2.Ra3 is very unlikely to lead to an advantage for white because white rarely gets genuine counterplay, and black has many variations to choose from so white will also likely be out of theory by move 6. In blitz especially, a -2 disadvantage is the far from lost when considering the wildness that a blitz game can turn out to be.
Thanks!
@dboing said in #20:
> By curiosity for the first statement, how many games over what kind of duration, did you get that association?
Dunno you can count or use an opening book to find out. But my last blitz and bullet ratings were done by playing that opening. My record ratings aren't that much better.
> By curiosity for the first statement, how many games over what kind of duration, did you get that association?
Dunno you can count or use an opening book to find out. But my last blitz and bullet ratings were done by playing that opening. My record ratings aren't that much better.
@Stillmtndewfan said in #18:
> In lieu of stockfish I would actually recomend using the maia bots as the benchmark. They're nueral networks trained to play more like a human would. Lichess hosts @maia1 @maia5 and @maia9 with all levels available from github to play against offline.
I meant to reply sooner, but do you know any good places to start reading up on Maia? I'm interested to find out more about playing more like a human and how the models differ - could be a neat starting point for another post. Thanks in advance!
> In lieu of stockfish I would actually recomend using the maia bots as the benchmark. They're nueral networks trained to play more like a human would. Lichess hosts @maia1 @maia5 and @maia9 with all levels available from github to play against offline.
I meant to reply sooner, but do you know any good places to start reading up on Maia? I'm interested to find out more about playing more like a human and how the models differ - could be a neat starting point for another post. Thanks in advance!
@NDpatzer checkout their website maiachess.com/
They made a version for every 100 lichess ratings band from 1100 to 1900. Each version learned from 12 million human games, and learned how chess is typically played at its specific level.
They made a version for every 100 lichess ratings band from 1100 to 1900. Each version learned from 12 million human games, and learned how chess is typically played at its specific level.
@NDpatzer said in #24:
> I meant to reply sooner, but do you know any good places to start reading up on Maia? I'm interested to find out more about playing more like a human and how the models differ - could be a neat starting point for another post. Thanks in advance!
You could start with the latest paper, well, that I am aware of, which is about trying to catch individual error models, which made them improve by one parameter the error part of the model function superimposed to some best play benchmark. My current understanding. I am not sure where I read that, if it was directly from the paper. Then you could work your way back. That would have the advantage of their own thinking back of their own previous work. It might make a more efficient reading of the first paper.
I am also curious about that topic. I would not say that I have been interested before by the engine part of the work, for the first paper, but I was impressed by an early version of the first paper, where I seem to recall a lot of preliminary data analysis of Lichess data over the spectrum of ratings from which they did some binning.
It has been a while. It may have been the first effort of human data-based conversion curves data fitting using Lichess data, maybe Lichess did some for its transformation of SF whole game analysis feedback, I know it does now, but theirs was not limited to games.
I am a bit curious about the same aspect, the definition of complexity of a position in that conversion curve, and then in the error model (and binning of game pair of rating avg.) on top of which benchmark. Then my question, is why not involve more of the available information. How dependent on the position information is the error model. Are all the positions of all the games within the range of Lichess data used, being modelled for their own error model differently, or it is a move-based model of error. Since they use the same NN function basis as LC0, but not the reinforcement learning, I am curious about that departure, how is the model being learned, what is the NN learning over what data. I assume they use the same basic input vector (or input planes) which is position information without loss of information, and they could have explicit position and move as part of their model or error.
How is the NN objective function constructed. That is the art and the most chess informative often missed "detail" in many of the non-dev audience intended blogs, i.e. people who would benefit from understanding what the engine goggles might be (and also, not much is dev places, well maybe now they do). I guess, using maia, might bypass the lack of curiosity that a best in its class of ending ELO reaching. The oracle vernacular is not even my editorial choice of word. It is part of the parlance in machine learning. It distinguishes the learning set up and source of training information to digest into the NN, between supervised learning and the learning from self-play, i.e. exploring is left to the learner, as well as the teaching left to the learner produced game outcomes.
In SL mode. The teacher provides all the position input and, for each of them, also all their target actions behavior. So the model fitting here might be at outside the NN training loop. Otherwise, I would be curious about the method of integration of the error model into the NN architecture.
It could be some kind of oracle policy plus some error model. I would love to see an presentation of yours based on that. I gather none of the above is foreign to your knowledge set. So you might be able to correct my guesses.
Then the individual fitting. I think the individual fitting subsequent paper, might be more insightful. I just don,t have the energy to answer my questions myself. as I don,t think that the papers present the preliminary work anymore, that might explain why they went for the simple error model. I don't know how much I believe in bulk similarity of rating determined errors per position, but it seems there is some 50% explanation power, over some data representation. I am missing a lot.. I know.
> I meant to reply sooner, but do you know any good places to start reading up on Maia? I'm interested to find out more about playing more like a human and how the models differ - could be a neat starting point for another post. Thanks in advance!
You could start with the latest paper, well, that I am aware of, which is about trying to catch individual error models, which made them improve by one parameter the error part of the model function superimposed to some best play benchmark. My current understanding. I am not sure where I read that, if it was directly from the paper. Then you could work your way back. That would have the advantage of their own thinking back of their own previous work. It might make a more efficient reading of the first paper.
I am also curious about that topic. I would not say that I have been interested before by the engine part of the work, for the first paper, but I was impressed by an early version of the first paper, where I seem to recall a lot of preliminary data analysis of Lichess data over the spectrum of ratings from which they did some binning.
It has been a while. It may have been the first effort of human data-based conversion curves data fitting using Lichess data, maybe Lichess did some for its transformation of SF whole game analysis feedback, I know it does now, but theirs was not limited to games.
I am a bit curious about the same aspect, the definition of complexity of a position in that conversion curve, and then in the error model (and binning of game pair of rating avg.) on top of which benchmark. Then my question, is why not involve more of the available information. How dependent on the position information is the error model. Are all the positions of all the games within the range of Lichess data used, being modelled for their own error model differently, or it is a move-based model of error. Since they use the same NN function basis as LC0, but not the reinforcement learning, I am curious about that departure, how is the model being learned, what is the NN learning over what data. I assume they use the same basic input vector (or input planes) which is position information without loss of information, and they could have explicit position and move as part of their model or error.
How is the NN objective function constructed. That is the art and the most chess informative often missed "detail" in many of the non-dev audience intended blogs, i.e. people who would benefit from understanding what the engine goggles might be (and also, not much is dev places, well maybe now they do). I guess, using maia, might bypass the lack of curiosity that a best in its class of ending ELO reaching. The oracle vernacular is not even my editorial choice of word. It is part of the parlance in machine learning. It distinguishes the learning set up and source of training information to digest into the NN, between supervised learning and the learning from self-play, i.e. exploring is left to the learner, as well as the teaching left to the learner produced game outcomes.
In SL mode. The teacher provides all the position input and, for each of them, also all their target actions behavior. So the model fitting here might be at outside the NN training loop. Otherwise, I would be curious about the method of integration of the error model into the NN architecture.
It could be some kind of oracle policy plus some error model. I would love to see an presentation of yours based on that. I gather none of the above is foreign to your knowledge set. So you might be able to correct my guesses.
Then the individual fitting. I think the individual fitting subsequent paper, might be more insightful. I just don,t have the energy to answer my questions myself. as I don,t think that the papers present the preliminary work anymore, that might explain why they went for the simple error model. I don't know how much I believe in bulk similarity of rating determined errors per position, but it seems there is some 50% explanation power, over some data representation. I am missing a lot.. I know.
Here's how you can challenge your own idea
> "Assign Stockfish to play White at very low strength."
It should not be that low. For your experiment to make sense, it should be a strength that YOU can NOT beat 50% as white. And then you try your "Meltdown" setup at that strength. See if you now win more. I doubt find such a level.
Or the same principle: find a level where you do win well over 50% with the Ra3 stuff. Then play regular whites. I do bet you will still win +50% in any opening.
> "Assign Stockfish to play White at very low strength."
It should not be that low. For your experiment to make sense, it should be a strength that YOU can NOT beat 50% as white. And then you try your "Meltdown" setup at that strength. See if you now win more. I doubt find such a level.
Or the same principle: find a level where you do win well over 50% with the Ra3 stuff. Then play regular whites. I do bet you will still win +50% in any opening.
@MillenniumBug said in #27:
> Here's how you can challenge your own idea
>
>
> It should not be that low. For your experiment to make sense, it should be a strength that YOU can NOT beat 50% as white. And then you try your "Meltdown" setup at that strength. See if you now win more. I doubt find such a level.
I'm not sure I follow. Are you saying it's better for the human player to sac the exchange with Ra3? If so, I don't think I get what that would tell you - you'd be trying to find an engine strength you could beat with an opening disadvantage you have no prep for.
>
> Or the same principle: find a level where you do win well over 50% with the Ra3 stuff. Then play regular whites. I do bet you will still win +50% in any opening.
There's a good idea here, but I think also a point of confusion: When you say "find a level," how will you do that? The motivation to starting with a weak engine strength is to implement the staircase procedure I described, which is just a systematic way to "find a level." I don't think that part of the design is a matter of the experiment making sense or not - it's a technique for finding a threshold, which you will need too. The good idea is to compare the threshold you find against Ra3 to a different threshold measured with an opening you know well - great way to calibrate the effect.
> Here's how you can challenge your own idea
>
>
> It should not be that low. For your experiment to make sense, it should be a strength that YOU can NOT beat 50% as white. And then you try your "Meltdown" setup at that strength. See if you now win more. I doubt find such a level.
I'm not sure I follow. Are you saying it's better for the human player to sac the exchange with Ra3? If so, I don't think I get what that would tell you - you'd be trying to find an engine strength you could beat with an opening disadvantage you have no prep for.
>
> Or the same principle: find a level where you do win well over 50% with the Ra3 stuff. Then play regular whites. I do bet you will still win +50% in any opening.
There's a good idea here, but I think also a point of confusion: When you say "find a level," how will you do that? The motivation to starting with a weak engine strength is to implement the staircase procedure I described, which is just a systematic way to "find a level." I don't think that part of the design is a matter of the experiment making sense or not - it's a technique for finding a threshold, which you will need too. The good idea is to compare the threshold you find against Ra3 to a different threshold measured with an opening you know well - great way to calibrate the effect.
In my opinion, it is not just a question of preparation, but a question of psychology: the fear of the unknown.
a reflex that we always have when we discover something that could put us in danger.
we cannot erase this, but we must try to keep a cool head and proceed logically. just say that he has a severely weakened side, one less rook and a less dynamic knight. so we have to tell ourselves not to make a mistake by going too offensively or having excess confidence. never consider yourself superior.
therefore, playing defensively nevertheless without excess prophylaxis and knowing how to play logically, by provoking exchanges
a reflex that we always have when we discover something that could put us in danger.
we cannot erase this, but we must try to keep a cool head and proceed logically. just say that he has a severely weakened side, one less rook and a less dynamic knight. so we have to tell ourselves not to make a mistake by going too offensively or having excess confidence. never consider yourself superior.
therefore, playing defensively nevertheless without excess prophylaxis and knowing how to play logically, by provoking exchanges
What I was trying to say was this:
Your theory appears to be that preparation of Ra3 will give you an edge. And you wish to prove your theorem by showing better results, against an engine, with 2. Ra3 than with regular play. Perhaps I misread you?
Forget my "find level" ... I do not think you will perform better with this opening at any level against any engine.
The "edge" against humans, comes from
1. Psychology as already mentioned, opponents may over-push to punish
2. Clock, as perhaps mentioned? Since many moves can be played automatically, you gain a clock advantage. This is mentioned (on the fly) by GothamChess in one of his YT videos on the Naroditsky - Ra3 games
Your theory appears to be that preparation of Ra3 will give you an edge. And you wish to prove your theorem by showing better results, against an engine, with 2. Ra3 than with regular play. Perhaps I misread you?
Forget my "find level" ... I do not think you will perform better with this opening at any level against any engine.
The "edge" against humans, comes from
1. Psychology as already mentioned, opponents may over-push to punish
2. Clock, as perhaps mentioned? Since many moves can be played automatically, you gain a clock advantage. This is mentioned (on the fly) by GothamChess in one of his YT videos on the Naroditsky - Ra3 games