Examples of chess games and/or chess positions can be offered to "prove" that chess principles do exist or do not exist. Indeed, the very same game or position could be used by advocates of the opposing viewpoints to "prove" their point.
The concept of the "swichen" (or switchen) might come to our aid here. This concept was developed by Ernst Gellner and he invented the term "swichen" for the concept.
"Ernest André Gellner FRAI (9 December 1925 – 5 November 1995) was a British-Czech philosopher and social anthropologist described by The Daily Telegraph, when he died, as one of the world's most vigorous intellectuals, and by The Independent as a "one-man crusader for critical rationalism"." - Wikipedia.
A swichen is basically something which can be viewed in at least two different ways. The empirical evidence (evidence available to our senses and thence to our mind for interpretation) can support two or more conclusions. Simple examples of swichens are optical illusions like the one where viewers can see either a vase or two faces facing each other. A viewer, on becoming aware of both "views" can see or view the image kind of switching back and forth. This example depends on background and foreground switching.
More complex swichens can occur when the thing being viewed is a complex system providing complex and open ended data from which formal, logical (deductive reasoning) conclusions cannot be made: there are too many unknowns or too many uncalculatable/branching possibilities. Instead, processes of inductive reasoning can reach opposite or several conclusions. Perhaps chess is a complex swichen in this sense.
I find myself in the position where the thesis of "The Axiom System" has given me a second way to view the general idea of chess principles. I am stuck in the middle switching my views and seeing inductively the potential feasibility of both views.
finally a good phenomenology of skill acquisition applied to chess by a decent chess player. thanks for writing it up, i guess.
if not yet- the author might find it interesting to look into Dreyfus's model of skill acquisition detailed in https://philpapers.org/archive/DREAPO.pdf
regarding the number of internalised descriminable situations (from the same paper):
> It has been estimated by cognitivists such as Herbert Simon that an expert chess player
> remembers roughly 50,000 types of positions.
@izutsumi said in #62:
> finally a good phenomenology of skill acquisition applied to chess by a decent chess player. thanks for writing it up, i guess.
>
> if not yet- the author might find it interesting to look into Dreyfus's model of skill acquisition detailed in philpapers.org/archive/DREAPO.pdf
>
> regarding the number of internalised discriminable situations (from the same paper):
Thanks for posting this very interesting paper. I hope @DailyInsanity has time to read it soon if he is not already aware of it. There is a lot in that paper that seems to strongly support the direction of DailyInsanity's thesis. I especially think the emergence gap or explanatory gap that opens up between the rule based instruction of the beginner and the eventual intuitive acquisitions by the master is well explained by the thesis of the Hubert L. Dreyfus paper and by the "Axiom System" as thesis-in-progress.
From my standpoint, the Dreyfus paper was difficult to read. It contains "terms of art" that I am not familiar with. A term of art is a term whose use or meaning is specific to a particular field of endeavour. Philosophy and cognitive science are replete with such specialised terms. At first, I found myself in disagreement with the apparent non-representational (non-modelling?) claims of the paper. However, as the paper proceeded I found myself understanding and somewhat agreeing with its position. However, I can't fully agree with any completely non-modelling, non-representational, cognitive theory and I do not think the paper is making a complete non-modelling claim.
Rather I think the paper is saying that any crude claim that the brain is doing logical modelling ONLY will be a claim that is inadequate and incomplete. We do logical modelling in chess when we consciously calculate combinations. And all players from rookie to grandmaster do that (to very different depths). However, that is not all the brain does. It also does "intuitive meta-modelling" or "intuitive goal-seeking" (for want of any better term). A master or grandmaster does mostly "intuitive" meta-modelling or goal-seeking when playing blitz. I think logical modelling and "intuitive" thinking equate to what Kahneman and Tversky called slow and fast thinking respectively.
I don't like the terms "intuitive" or "fast thinking" for this neural network phenomenon. A better term might be extractable from the Hubert L. Dreyfus paper. I will leave this post at this point. I usually tend to write too much.
DailyInsanity has been quiet for a week or so. I hope this means he is busy writing the next part of his paper. Have a good and productive day, everyone!
@izutsumi said in #62:
Thank you very much for your comment and for linking that paper. I think I've heard that name mentioned when reading some of the works/books/papers of Fernand Gobet, who has done some research regarding decision-making in chess. I'll definitely check that paper out when I get the time!
@izutsumi said in #62:
> if not yet- the author might find it interesting to look into Dreyfus's model of skill acquisition detailed in philpapers.org/archive/DREAPO.pdf
> A Phenomenology of Skill Acquisition as the basis for a Merleau-Pontian Non-representationalist Cognitive Science
Hubert L. Dreyfus, Department of Philosophy, University of California, Berkeley
pdf file date has 2002 as the year. There are references in some section at the end (or they might be page notes, in a more philosophy publication style).
It does use code words from that field. At least words I am not sure to even have a first guess (for I have been wrong so many times with established terminologies borrowing from natural language, through coiner idiosyncratic preferences).
I have difficulty figuring out the last pages context. if those are a reference section with notes, or if the last paragraphs act as conclusion.
But this is creating interest as reference, whether used as opposing model or not.
> 7 Richard S. Sutton and Andrew G. Barto, Reinforcement Learning, Bradford, MIT Press, Cambridge, MA., 1998.
That is the background behind the deep learning implementation used in A0 and Lc0. (sorry if other pages here already mentioned that).
well. yes.. these seem to be notes. maybe for the whole paper. anyway. my chunky fast read so far, looking for things I can get some sense of that is not a philosophy author backreference (a philosophy text habit, very keen on tracking the lineage of the concepts, danger being, that we don't get the full set of assumptions, de novo, or de visu).
However, the level of discussion in the notes is promising in my eyes.
> finely discriminated situations.
I like this for opening the door about inside the previously used as black box of patterns. or linguistic "chunks" from a somity in the field of chess psychology, used as models and terminology for chess patterns. But I am not sure there were actual research or even model in relation to the position information. Here the words seem to let me believe they might be going there.
Also, my pet topic: the question of the evolution of patterns in relation to this blog discussion (some question of improvement from some state of skills or performance or strenght level to some other, invovling the awareness that previous expereince set is smaller than the future experience set of positions to decide a move upon (or situations here, which might be not exactly the position sets of axiom1).
It seems to finally involve the idea of evolving boundaries in the chess world of "situations" (need to read, of course to get rid of the quotes). Those are my words. As I do use some mind'eye ambient space projection of my own thinking habits from outside of chess experience. I see clusters and fog, and pattern recognitoin but also discovery (the missing cog in the thinking only models based on expertise only focus, and the rest as control group to make contrasting insights). All in evolution in relation to experience and other learning sources.
What seems to be missing, in my eyes, that their proposition, a constructed model, seems to be concerned with is what of the position signals (situation?) can be discrimated apart from each other, and how does experience change that.
From my corner of past expertise, which is more math. mouthful, I would use the notion of distance between positions.
Beyond similarity intuition and less similar going into not similar any more, to a more global measure of inter position distances.
I am hoping that the reading of this by others, would make such a concept not as alien, as I am under the impression that it is.
Similarity from one single positoin is not enough to construct a distance. One needs to get an idea of the large set of position. Is my hunch. Glad about this pdf. having possibly more credibliity or influence on interested participants here. than my ramblings.
Now, if some one could make glossary about the many proper names used as pointer to concepts or models in there. I would be "buying". Not the same mindframe as I am used to. Or I forgot... (backreferences, and proper names based concept nomenclature, tend to saturate my working memory before any logic can be ground and long term memory take over.. reading chunks discrimination that might be.)
> I believe everyone can agree that for chess rules and principles to be useful in decision-making, they must influence us to play moves we wouldn't have chosen without consulting them.
more random chunk readings. I am actually looking for when my understanding of the learning question might depart from yours. Our respective unspoken premises left, even through blog 1 to blog 3 precautions which I have followed, and the set of axioms, which still has room for my understanding. It might go back to axiom 1 unspoken or undefined (my current working hypothesis). I am also being distracted by your youtube video about woodpecker. jomega has had a look at it, not me. But then it seems that it might be where you would be headed, or have already been trhough in blog 4 parts i have not yet read, or any of previous post blog discussions.
About the quote above first: It seems you may have restricted the purpose of chess theory to the in-game decision process of an already learned enough or expert enough player. That this might be a hint for me about the scope of axiom one, that is not spelled out. For me, it is clear that chess theory is about communicatoin and learning or teaching first, and is not for performance at high level. I think we all agree that once internalized chess concepts are better left in execution mode to that subconscious fast acting brain (but slow learning btw).
I have a wider interpretation of the axiom one problem. which is about for once not conflating the expert decision model with the learning of it from any stage of learnedness in the more general form. So, can you specify axiom1 a bit more with respect to the external science phenomenology that we might not have the same common sense about.
Now going back into my thinking in relatoin to your proposal that I could read so far.
Few words in your replies to me, gave me some hint that we are departing on axiom 1. about the very question of improving or learning specifics. Since as worded there is no restriction I can point to, I might have to use hypotheses about your meaning.
It appears that all you are saying about the relation between training set and performance set, is that it is bigger.
In the limit, it could be the training set or previous "digested" experience set, plus one new position never visited.
Tangent: might I had, that this might be a dynamical thing is self, how does the new experience become the new training set, and does any new instance of single exposure to a seed position and one continuation from it, means it is equally now part of the training set as the previous training set).
From my current preferred existing machine model of learning using representation likely full deep NN latent space in A0 and LC0, as at least one model to consider, I would say, that the question is lacking definition, or specifics, that might be helpful.
The question of position similarity, constrast, difference, or more generally and more general, consistent and potentially complete questino of distance between positions have been thoroughly avoided in chess culture. This, in spite of the sprawling growth of "opening theory" since that started being tallied for all to find secret winning fasct execution recipes, so that they don't need to think about it until out of book.. book growing being that point.
Sometimes we don't ask because it is not helpful. Sometimes we don't or stop asking, because we do not have yet the tools that would sustain the well formed question.
As you said, possibly misusing the word "volume" as possibly from a database or economical language point of view to me still count or cardinality of a countable (ahh.. words, pain in the ...) finite coutable set.
That a set of position size is only about how many distinct FEN strings are in it. So that, given the couting nounds of the past, this is not within our grasp to go specific in your axiom 1. The number of stars in the universe thing.
I should say, that the chess culture avoidance seems so strong that even LC0 and A0 have been oblivious to it for some time. It possibly recently changed: they did produce a paper where they realized that the A0 champion (and LC0 as the open source version or off-shoot, might be ever better than A0, if not taking winning shortcuts based on undefined ELO pools of engine to beat as priority of optimization, always a danger, at the core of the RL dilemna of exploratoin versus exploitation. (this is the dynamic version of the problem of generalization in supervised learning). They found out that their expert zero prior chess champion, did make compromission in its chess position exploration to become better. I would love a discussion on the paper with anyone willing to participate in a forum thread with some plan of reading.. I need the social motivation for the remaining things I did not read or might have guessed over too fast. (the methodology of combining the repertoire biased or as they say specialists, to beat the zero-prior one RL trajectory champion).
But the mathematical model that is being implemented in those machines, allows tools, that do not require that compromission. I guess I will leave this at that. In short. There are other ways, to talk about size. and you were saying that surely size is not all that matters, and of course. whereabouts of the positions with respect to all positions in some space, ..., well that is already formalized that space in the mathematical model...
and it is not blasphemy to imagine it in our visual internal support imagination. even if projecting 32 planes into some 3D cloud of positions, as points. I wonder why people trust one number from the SF magics without blinking an eye, and have trouble with having a stable external non competitive agnostic of evaluation metric system for their position... Is it that the tools are not part of the extended chess math culture, save A0 and LC0? and now some of SF activityies. Note: they can use deep NN as black boxes forever.. and would still get great ELO to show for. I suggest that generalist verus specialist deepmind paper. sorry i lost the exact links and titles. my usual compacting long term memory processes that only retain assocations with some logical probabliity aspect.
sorry for the english. It takes me at least a day to be able to read my bad english as such. otherwise my thoughts fill in the errors... bliss of fading memory.
@DailyInsanity said in #64:
> Fernand Gobet, who has done some research regarding decision-making in chess
thank you for mentioning it, his works completely went under my radar for some reason. will be checking "The Psychology of Chess" out
Any link for Fernand Gobet public access pdf. If there are articles. or preprints. The wikipedia page, seems all inert references. I did not try to make google searches of them yet. but if a good pointer already exists, please share it here..
I value not relying on one author only is such non-reductionist or non-derived models. Such lack of direct emprical reprocible empirical methods, might be unavoidable, but it is still highly susceptible to school of thought, which might not be aware of itself as such, and might not then use the prudence and debating method of progress, that should go with such scientific method, not rich on the empirical-theory tighnest of interaction. (I do not know the official epistomological words, I am of wild education, more like hands-on there too, and my wandering hypothesis engine).
Chess theory should also proceed that way. There were shcools of thoughts, but I did not seem a culture of debate. mais across long times scales. . from FINE to Kendricks.. not sure that was constructive debate. Not the authors doing either. I think the one man show and lack of individual humility that might be nurture along the hard lifelong needed word commitment and success frequency might have left some trace in the culture of coopertaive endeavors such as any thing to share outside of the single game board. I am thrwoing this hypothesis.
I do not thing this is the case here. though. So.
But chess learning theory, might also, speed up, and look below. Prudence in jumping on consistent looking constructions. It seems to me that chess culture does not have a strong tradition of peer debate. Except here. Why I find this rare. peer being about the common accepted target of intellectual pursuit, made clear from the begginning (or trying to, slipping or credentials otherwise might really put noise in the comunication channels, and I propose that it has for a while). Not here.
I think we here are awaiting Part 5 with considerable interest. Meanwhile, here are some ideas and anecdotes. Yes, anecdata is of limited value but it might suggest some lines of enquiry. First, some concepts which may be useful. All of these concepts are very likely already known to @DailyInsanity.
1. The Forgetting Curve - https://en.wikipedia.org/wiki/Forgetting_curve
2. Overlearning
3. Spacing effect and spaced repetition
As a very old adult improver (nigh on 70) I wonder how much improvement I can ever expect. Clearly, I am going to have some problems with learning and retention compared to a younger person (and especially compared to children, teens and even anyone under 30). So I presume that special attention to overlearning, spacing of learning /spaced repetition will be required. Another issue is energy and stamina or lack thereof. I feel I can see from personal observation that physical and mental stamina are linked.
I take DailyInsanity's point that repetition of motifs rather than a mix of puzzles will be more effective for retention. The thing is this. There are so many motifs that even a motif a day with 2 or 3 spaced sessions in the motif (on that day) will take at least a fortnight or maybe a month to run through. The time logistics seem fraught even if one trains say 6 days a week x 3 times a day at 1 motif per day. I feel that a given motif will also need to be done 3 times a week at 3 times a day for the first week, then 2 times a week, then once a week, then once a fortnight. Thus allowing new motifs to be introduced and worked on and to follow the same trajectory of lessened frequency and increased spacing.
Now, how should chess students decide which motifs to target first? They might think, target first what I am bad at. To the beginner or rookie (just above a beginner) the answer will be "But I am bad at everything." Even so, there will be mistakes that particularly irk them. I was annoyed when I was falling for simple pawn forks so I targeted forks. ( I couldn't refine down to pawn forks on Lichess IIRC.) I also made a mental note to check that I am not putting pieces on the same rank one space apart without checking for potential pawn forks. This worked to some reasonable extent. I was missing chances to trap opponent pieces so did trap pieces puzzles. This worked quite a bit. I noted that once I did hook mates I started noticing hook mate possibilities in game, which I never noticed before. I was able to use a hook mate threat in a game. The opponent saw it and avoided it. I still got a better position but blew it later and lost for other reasons. Nevertheless, hopeful signs.
The real deal will be "What motifs and what techniques" will give me the best percentage return of wins or at least technically won positions? I think there needs to be research into why players (and individual players) lose. What mistakes do they make most at what skill levels? What motif chances (tactical motif chances) do they miss most when finding the motif or the potential for it would improve the position or lead to a technically won position? Knowing this, the student can prioritise motifs and tactics to start with. Failing that, the student I guess should simply make an educated guess and start somewhere in the round of motifs and then go round and round them with reinforcement and spacing.
Thoughts anyone?