Mentionning overfitting sealed the deal. I don't need to read more to support whaever thinking is propose after that. Even if I may not agree with the conclusion (which i have not read) or certain bits left for me to absolutely read, just that means that finally, at least, those who will read your blog will have some basic generalization/recognition of patterns basics under the FOV for some duration enough for such ideas to finally, percolate in the chess community.
Kudos just for that word.
edit: what is a GM norm? sounds intriguiing to me from a mathish point of view
edit: what first came to mind, is did you consider the gamut between exact same and total mix random pairing?
my partial experience based suggestion: the theme (like here on lichess) could provide for thematic guided controlled random novel exploration within theme pool... using environmental feedback from performance rating among enough total mix random to populate with enough numbers theme repetitions (assuming the puzzle categorization was already considered sound or vetted).
edit: (for thought process transparency....:) I noticed or gathered the intent to factor out some variables. Motivation intensity (good point, trying to do that, hard to control or measure externally i would say, but at least trying is good, however, micro-motivation on the visual field of what is the board telling you, might be a selective atttention issue, for some further science-fiction). my intent here was to mention various time scales of learning processes (call that memory times scales, I wonder if there would be only 2 of them, as subjectively, and longitudinally, perhaps with some confirmation bias at introspection, who knows?, i find secons, to differ from minutes , to differ from hours, from days, from weeks, and from monhts (should i add years, that might be a memory slow trimming process).
same edit: i suspect that is already in part (maybe not in complete scope) of the woodpecker idea, it might be right about some of the time scale and duration of exposure chunking and distributed repetitions..
not reading your blog linearly and not done yet.. too many reactions at each attempts to read....
Kudos just for that word.
edit: what is a GM norm? sounds intriguiing to me from a mathish point of view
edit: what first came to mind, is did you consider the gamut between exact same and total mix random pairing?
my partial experience based suggestion: the theme (like here on lichess) could provide for thematic guided controlled random novel exploration within theme pool... using environmental feedback from performance rating among enough total mix random to populate with enough numbers theme repetitions (assuming the puzzle categorization was already considered sound or vetted).
edit: (for thought process transparency....:) I noticed or gathered the intent to factor out some variables. Motivation intensity (good point, trying to do that, hard to control or measure externally i would say, but at least trying is good, however, micro-motivation on the visual field of what is the board telling you, might be a selective atttention issue, for some further science-fiction). my intent here was to mention various time scales of learning processes (call that memory times scales, I wonder if there would be only 2 of them, as subjectively, and longitudinally, perhaps with some confirmation bias at introspection, who knows?, i find secons, to differ from minutes , to differ from hours, from days, from weeks, and from monhts (should i add years, that might be a memory slow trimming process).
same edit: i suspect that is already in part (maybe not in complete scope) of the woodpecker idea, it might be right about some of the time scale and duration of exposure chunking and distributed repetitions..
not reading your blog linearly and not done yet.. too many reactions at each attempts to read....
> Get together a group of friends to do the program at the same time. This will provide accountability and motivation.
I'd like to see puzzle arenas someday.
I'd like to see puzzle arenas someday.
@thibault as seen on this blog's topic of puzles, is it possible to do puzzle arena, or is it already puzzle racer?
about overfitting (and underlying concept of generalization). It demands a testing universe to make sense. Hence my curiosity about the norms. It would not be overfitting if all chess games ever rated would start from the same puzzle starting position.
and it was the same engine/puzzle master robot that was playing the puzzle against you.
one could put clocks calamities in your arenas, but one could also allow non-clock take whatever time (which is already there).
so some of the experiement can be done, assuming enough chess lovers in the pool to drown the non-curious sub-population variable in the statistics.
but in any case there is a huge floating hidden variable of what is the end point measure of learning..
I want to understand above winning, but my bet is that if i had inifinite life time, that could aim at the testing universe being any legal position as starting point, that would be the clearest definition of chess world. But, so my theme suggestion might be biased toward that. Just a warning.
perhaps..... instead of arenas or as complement one could start rating continuations from the puzzle accepted solutions (off topic, but could not help pointing out some fog or opportunity hidden there for learning or improving at "something").
anyway.. The GM norms.... what gives? I should scan the blog for the notion of representative data sets , training and testing...
and it was the same engine/puzzle master robot that was playing the puzzle against you.
one could put clocks calamities in your arenas, but one could also allow non-clock take whatever time (which is already there).
so some of the experiement can be done, assuming enough chess lovers in the pool to drown the non-curious sub-population variable in the statistics.
but in any case there is a huge floating hidden variable of what is the end point measure of learning..
I want to understand above winning, but my bet is that if i had inifinite life time, that could aim at the testing universe being any legal position as starting point, that would be the clearest definition of chess world. But, so my theme suggestion might be biased toward that. Just a warning.
perhaps..... instead of arenas or as complement one could start rating continuations from the puzzle accepted solutions (off topic, but could not help pointing out some fog or opportunity hidden there for learning or improving at "something").
anyway.. The GM norms.... what gives? I should scan the blog for the notion of representative data sets , training and testing...
@dboing said in #2:
> edit: what is a GM norm? sounds intriguiing to me from a mathish point of view
With some exceptions (winning certain major tournaments, for example), a player must attain a FIDE rating of 2500 and three 'norms' to be awarded the GM title. A norm is a tournament performance rating of 2600 or more over a nine (or more)-game tournament, having played multiple grandmasters (at least 1/3 of one's opponents), and the average rating of the players one plays being 2380 or greater. There are exceptions and further rules about norms; I'd check the FIDE rulebook section on title requirements if you want to dig deeper:
handbook.fide.com/chapter/B01Regulations2017
> edit: what is a GM norm? sounds intriguiing to me from a mathish point of view
With some exceptions (winning certain major tournaments, for example), a player must attain a FIDE rating of 2500 and three 'norms' to be awarded the GM title. A norm is a tournament performance rating of 2600 or more over a nine (or more)-game tournament, having played multiple grandmasters (at least 1/3 of one's opponents), and the average rating of the players one plays being 2380 or greater. There are exceptions and further rules about norms; I'd check the FIDE rulebook section on title requirements if you want to dig deeper:
handbook.fide.com/chapter/B01Regulations2017
I am sorry. I read what i wanted, I thought you were proposing some group testing with puzzles over some period of time of the main question opposing many distinct puzzles to many repeats of the same puzzle, while keeping the scheduling the same.. And fast reading with such bias in mind (with some complementation, like we complement the light color when closing our eyes, try it) i thought your were finding a way to get motivation variable under control in the experiment.... I made some science-fiction being enthousiastic about your question.. not the first time.. so again good question to raise... and the rest good not to waste any energy to keep any way of working hard at chess.
since the book talks about 1100 puzles, maybe the difference over the longest time scale, with your question in mind, is how the exposure of diversity is distributed over that time scale.
I would suggest that it might be sequential in one case, and the opposite might be uniformly distributed over the longest time scale. There could be tons of different ways to redistribute diversity with respect to repetition.
The question of overfitting might still be there in both cases, but about the set of 1100 puzzles representativity of future chess experience.. So if talking about opening repertoire repetitious practice, or even some style characteristics present statistically in the 1000 puzzle set, the overfitting might only be a probolem if the repertoire or the style characteristic was not the one(s) in vogue in the duration of the fast improvement experiement.
Still talking about the theoretical learning ideas. no clue about how that applies.
I would suggest that it might be sequential in one case, and the opposite might be uniformly distributed over the longest time scale. There could be tons of different ways to redistribute diversity with respect to repetition.
The question of overfitting might still be there in both cases, but about the set of 1100 puzzles representativity of future chess experience.. So if talking about opening repertoire repetitious practice, or even some style characteristics present statistically in the 1000 puzzle set, the overfitting might only be a probolem if the repertoire or the style characteristic was not the one(s) in vogue in the duration of the fast improvement experiement.
Still talking about the theoretical learning ideas. no clue about how that applies.
When you solve the same puzzles over and over again, you develop a small puzzle database, no? You develop a memory of these positions. The book just about all of the puzzle themes and it has great diversity in positions, stages of the game, motifs, etc. After you memorize a certain position, when you encounter a similar position in a game or a different puzzle, you are able to figure out the tactic, since the positions are similar. Thus, training the same puzzles over and over helps develop a memory that you can access when you reach a similar position OTB.
Instead of doing the same puzzles over and over with increasing speed, as suggested with the Woodpecker Method, perhaps it would be better to do different puzzles that all follow the same theme with increasing speed. By doing different puzzles, you avoid the risk of overfitting, but you also have the ability to focus on a theme to train it more effectively.