Hi, I just read the blog. I did not know the "woodpecker" method and learned it. The only well known woodpecker I knew was the Woody Woodpecker until this blog :)
Anyway, blog reminds me, my own blog I wrote about the topic "is chess a memory game?" I completely agree the writer about this topic : "seeing a new things ,i.e. solving a new puzzle each time is better", however ; I reminded one of the most important exercise tecnique of Kung-fu : mastering the technique by repeating the same movements...
I am not a Kung-fu master but millions of people uses "the same" technique. I like Elvis and heard a good sentence : " 50000000 Elvis Fans Can't Be Wrong"... maybe same can be applied for WoodPecker method.
So as a conlcusion, my opion is that it is worth to try WoodPecker method !
I want to do a new suggestion, a challenge maybe for lichess, in lichess, is there any way/tool to use WoodPecker or other training techniques?
@CheckRaiseMate raises an interesting point. However, overfitting is an undesirable effect of the NN shape, not a consequence of not varying the training data. Besides, it would be almost impossible to overfit any data in a human brain. Overfitting happens because all the information that the NN has access to is the training data. So unless you show only these positions to a person that hasn't played or studied chess before, overfitting should not occur (and even then, we are assuming that the human brain works exactly like a NN, which is at least debatable).
I like that these topics are actively discussed, but I would prefer if the author, who is an FM and therefore better at chess than most people here (myself included), shared the training methods that allowed him to reach his current level. Personally, I would appreciate an article that treated this topic instead of speculating about what should or shouldn't work.
I will keep following the author's posts. Although I disagree with his reasoning's validity in this particular instance, we as a community need more people like him, and it was an entertaining read.
I would not blame the flexibility of a model being fitted for over-fitting automatically. NN comes with learning algorithm. You have to consider the whole package. The exact nature of the global optimization. Is it about maximizing the fit to the data only? or is there a formalism ensuring that testing for generalization over inputs unseen or not part of the training set is also optimized. This is a 2 objective optimization problem.
Machine learning has developed techniques that, given a whole data set assumed to be representative of the general context that is supposed to be learned, will be able to maximize learning which is about both fitting a subset of the dataset and minimizing error on inputs not part of the training. (in essence, cross validation, can measure if the learning algorithm is accomplishing this).
But while the threat of overfitting from the very felixble function basis being optmized in well controlled by such algorithms, and while chess does not have problems of precision, a mate being a mate.. or a SF score being a reproducible SF score if talking about puzzle success from puzzle maker based on SF with specified parameter set values, the other source of bad generalization of over fitting, which is by having the training set being not a good sampling of the general context, might make such particular sample of the reality (say the 1000 cases from the book), representing all the other possibilities.. And the calibrating and precision improvement à la kung fu body motion training (on top of exsiting model of the 3D world we all have developped before learning king fu, or dart or basketball), would be about learning the very precise execution patterns repeated over and over, in cases outside the 1000 puzzles, that would be made to resemble the closest one. It might be seeing nails everywhere after having mastered hammering with a specific hammer and shape of handle (and balance of the tool).
The 1000 puzxles could be too similar to each other... or too far apart compared to the huge chess world of challenge positions... We need to have some way to measuring how they relate to future experience.. how they were chosen.
I would mention that visiting the ELOMETER site might help understand the problem of capturaing all of chess skills and dimensions. via a set of challenge positions.
if the sample of positions turns out to be a bad covering or spread in relation to the general context, then its quirks might be considered part of the law of probability being learned. It might not extrapolate well in positions that are far from the more sample type of positions.. in reality we lack tools to make quantitative statements like we would need to, in order to talk about overfitting.. how does one call the degree of similarity between positions.....
I think there is a difference between sub-optimal training and ineffective training. I think the post showed theoretically why the woodpecker method is sub-optimal, but not why it would be ineffective. And I would love to see a study to see if the theory is true.
ineffective or effective indeed would be the empirical statement. It would need a study. or a group experiment.
There seems to be some confusion as to what the Seven Circles (aka Woodpecker) method is and what it’s designed to do. It’s not an alternative to regular tactics/puzzle training. It’s an intervention meant to be performed once or even once in a while. It’s designed to improve pattern recognition and tactical intuition. If that’s a weak point for you, if you regularly blunder pieces in your games or miss mate in ones, then the method should result in rapid (over a span of months instead of years) improvement in tactical skill. And if your rating is low enough, it should lead to a major increase in rating. If your rating is already beyond the lower classes, it might help you move up a class.
But why is this so? Players below Class D level simply don’t have good board and tactical vision. They are told by trainers and books to follow steps like look for checks and threats before and after each move but looking for something and seeing it are two different things. Regular puzzle training can definitely improve tactical vision but it works slower. You have to see the same pattern several or even many times before it gets ingrained into your subconscious. When that happens you see the threats and mates automatically: they jump right out at you. You no longer have to consciously look for them. And this is not just for beginning players. Even at higher levels, everyone has blind spots that can be eliminated with such training.
The method is simply to ingrain patterns into your subconscious by repeating a group of puzzles (say 1000) a set number of times (7 according to the method) and each time trying to find the solutions faster. If you see the same puzzle too soon, you run the risk of just memorizing it. And if it takes too long before you see it again, you run the risk of having completely forgotten it and starting from scratch. So the set size of 1000 puzzles makes sense (though arbitrary and so perhaps a different number might be more ideal). After the cycle is complete, you continue with regular puzzle training. And during and after, you should also be playing practice games to integrate the patterns.
I myself have used this method so I know it works. I was stuck at the class C level for many years until I came across Michael de la Maza’s book. I decided to try the method using tactics software over a period of a couple of months, combined with practice games against the computer. I returned briefly to playing tournament chess and quickly moved up a class. So for someone who was not making progress for a really long time, this intervention worked. If your progress is slow or non-existent with other training methods, it might just work for you too.
Some two cents ideas
Chess training can indeed be compared to sports training. And some automatisms can probably be acquired only if we repeat the same situation.
Is it useful to work on complex combinations if we are already unable to exploit/recognize a simple pin or skewert?
The Woodpecker method may be useful if
1. we don't work only with this method
2. we combine it with the spaced repetition techniques
The articles and advices on this method do not help us much on the way to work on tactics. We arrive at conclusions such as "if it suits you, do it" or "it allowed me to win 150 elo in 6 months".
In short, woodpecker or not? But what is the alternative?
Good post.
Continuing on the ML analogy, it is true, that for some learners, repeated exposure to the same examples helps solidify the pattern, say, if it's in a different ensemble, and not always the same group (think epochs in NN).
We don't know what it means for us humans, but changing the order (back to front, instead of the natural order) did good. I memorized specific neighbors for some puzzles, and was recalling the answer for those even before I saw the position. I imagine something like a randomizer would work better to get the patterns even better
I had forgotten this blog
It is one that does ask the generalization question up front
Maybe a blog tag for generalization could be used by authors having that concern high on their blog thesis intent
That would help the theme itself
Even arguments dismissing its need
Although that is more by lack of concern than argument. Possibly because we have no way to measure what that means
New position. Is any distinct Fen new compared to any other
How much but information or character change makes for a new posituin
We all have our own evolving subjective notion
I would bet that experts from the same cloth of experience might converge to an objectively 3rd person sense how new that they would each agree to
Although I think there would still be some human trajectory diversity even at the highest levels. This is all research opportunity
I meant « bit « of information not: but.
Thumb device typing above and now.
Also given a position and all possible different position, which would be considered too similar to constitute a new position for the continuation problem solving to best outcome? Or which would?
I think a new position in some player experience learning trajectory (bundle), as a question, is related to that not yet seriously researched question which is about objectifying the currently subjective only notion.
Trick question of sorts. That is what I was trying to explain. As it had not been made objective. Novelty by move sequence seems to be about the one ply that differs by any different choice than the previously commonly known.
The problem of characterizing experience over some wilderness large set of positions is in turn related.
Which in turn if using machine learning reproducible and further research able as seed models, or starting working hypothesis of the current black box we have been calling intuition or subconcious or digested experience, would relate to opening the dataset of training relation to the target wilderness
Over under fitting are well defined in context where the sample is assumed representative of the target problem or task or scientific subset of external reality for which we seek to capture the law of proba.
Back to move novelty: ...modulo quality of play following up. Even notions of surprise are like that. Sorry editing thumbly on touchscreen. Small fov.
Now completing aborted sentence from previous post.
Even blogs that would seem to imply or suggest that generalization is not really a chess skill to teach or learn or hone
That is there or not or that random or unconcerned by such notion sufficient repeated exposure to samish positions will eventually provide as much as having had a concern , well those could also use the tag
And be useful to a calm understand pursuit through discussion across blogs or within their post blog discussions
Or new forum integrative discussions could refer through such blog
I hope this cramped thumb thing is not adding to my ramblings apparent dispersion. Edit: it was. And it has been painful. This means I am really wanting to share this. Been for ever
Glad not alone anymore. At least that I can read without digging or finding how to go about it with intelligence angle without having the title that would help the diffusion
I have only rambling of my transparent internal debate
So readers can bring their own brain to the quest or understanding pursuit
Now I can bring my humble support still rambling and now really tired of fighting with thumb typing and autocorrect. To those that have many more types of experience to be heard efficiently.
The end
