> About models of things, and what models what in the end (still wondering, my self). And the nature of scientific theories in relation to their target phenomenolgy. (and the technological instruments used).
As with all models. It is important to not only think hard on the conclusions or its outputs given inputs, but also the ground assumptions of the model.
So that the sharing of this experiment(s) is not slipping its purpose in our parsing attempts. This is not a critic of blog, article, or your post, as this is very hard, and actually part of the reason, there is science, and peer reviews, and reproducibility and sometimes open-source code and open data requirements. And why it is not just one experiment but a chain of papers or a network across time that can revise each other interpretations by keeping the wheel of theory (models) and experiments going. I think this is not said enough, how science can actually function ideally. It also depends on the time frame between theory testing and theory building arc. Not all sciences can effort such a fast cycle, says as physics but now as the science part of Machine learning, which I guess I can call machine learning models. The market visibility explosion and smoke might make us distracted from the science behind it, which is not complete (ever, but now lagging, at least in visibility).
Now more specific. About how to use A0 to keep thinking about chess theory and chess theory of learning.
So I am not sure that the DeepMind paper using A0 is about how our biological embedded learning dynamics happen. Which your post seems to refer to. And If find it refreshing, as for a long time I need to hear such common sense about learning in general, but coming from the chess community.
A0 might be more human like in that it does not really think that much like a tree, even if the algorithm implementation has to go CPU as this is all we have as a machine model, but even the mathematical form, that prescribes the learning, is not bound by biological rhytms.
And all such Deep NN based modern AI, is about the power of the intuition part of our brain. That is what it is modeling.
So, the way the many of us, as a group, can make helpful concept being shared (theory), to help each other chunk things, and make it human accessible is not this type of model way. As we can't accumulate many games in some batch, all its decisions, turn by turn and back bubbling only the terminal outcomes in a well controlled fashion and statistically vetted attribution to association at position and move chosen probability estimates.
The source code or algorithm might resemble us, as potential, before training. But the data scope and speed and lack of biological limitations (and their hidden other intelligent dynamic solutions as above), is the machine reproducible monopoly (lack of words, good enough). So, it is a historically neglected part of chess implicit unique theory of learning (not your post, it seems to have gone vinyl record "wrong direction", that is good). So we should listen to it. But also be aware of the exact assumptions that matter. (same problem with using SF as we do here on Lichess, btw.) Those are models. And we need to know the "givens" of their conceptions (not the method of their implementation , this expanded above, as edit.
Fortunately, A0 comes with mathematical presentation (in part in first papers), not high level model first, code not even later (we have LC0, but it has it owns recreation, and it does not have the equivalent of the A0 paper (which wsa for go)(***
The A0 concept might not be the ones we need. Unless like SF we can inject such knowlege without preventing the proper exploration of chess (well, SF is perhaps an acrobat trying to have the cake and eat it too, and might not realize if it is still exploring or not).
That is why I think there is need for research beyond our conscious expectations, about more than concepts we can talk about yet. It maybe that A0 is just weeding the terrain for us, to restrict our concept search, but we will need to be creative there.
*** above. just checking which DeepMind paper might have that. or if it is covered by Not disclosing source code, I consider the input encoding to be of scientific target question interest (wink to chess engine of this world.... and people using them blind).
Question to blog author. I have not read any of the chess-specific DeepMind paper method sections. Do they finally present the mathematics of their encoding of the first layers? That would wormhole me to go read that.
@Toadofsky said in #6:
> It's a pity this was done with chess rather than shogi, since chess tactics tend to overwhelm strategic/positional play and I am curious what strategic concepts could be discovered in shogi.
Why should tactics preclude strategy? Ok, maybe from a statistical design for detecting or doing research to find seeds of concepts to keep search further. Chess might be not separating as well the 2 types of board information concepts and one might have the exahaustive serach turn by turn view that all chess is tactical just deepeer and deeper segments displaying the reward logic from upstream decision criticality (to get that deep downstream big enough odds gain).
Perhaps shogi tactical concepts are, easier to separate? In that direction? How long are shogi games. And is the blog here, discussing this? types of concept, or nature of concepts. Same for paper. (questions to myself as well. so I can have some attentions stakes to actually read like i want to).
Also, as often in chess discussion, we tend to jump on words as if everyone had the same meaning. I just gave one direction of possible meaning. The depth between the human pre-decision thinking, imagiin,,g seeing, evaluating, and the stake that decisin will have (with intent? I think if we say strategy, we might be meaning learnable and desirable, which means imaginable and ... intent).
Also why assume we should read the blog before the discussion?
@dboing
"Is that a theory of the type: There are 2 types of people in the world. There are bins missing there :)"
This is not an A or B. Both factors are an individual spectrum.
You can be more pragmatic, or you can be less pragmatic.
You can be more theoretical, or less theoretical.
You can care more about it, or you can care less.
Lol. You have a point though,
if you assume that I'm referring to general people.
"Some people" was a hidden reference to myself :)
@MishaZagreb said in #15:
>
And if apply your first post, recurring denser graph than tree or line of mind processes involved in learning or performing tasks as demanding as chess play over long enough time scales, that would give you some room for such hidden self reference, to visit some of the various characteristics you tried to isolate with that pale shadow means of communicating our thoughts, verbal language but just with ASCII characters.... I think I found refresshing that we don't have to be stuck through the obvious infrastructure assumptions.
The art of using language to share more than it could do, as direct message. Patterns at the spatial vertical space signalling window.
I cut my line at this point.
So this line can be seen in short term memory parallel.
wihtout out working mmeory saturation in vain.
I might be speaking for my single brain evolution experience I could vouch for...
I find this direction of research really exciting, but the example from the paper left me confused. What is the concept? I couldn't identify a common theme between the two positions.
Very good.
@CheckRaiseMate said in #17:
> I find this direction of research really exciting, but the example from the paper left me confused. What is the concept? I couldn't identify a common theme between the two positions.
Maybe something like sacrificing material to create space around the enemy king to use for an attack?
In any case, I think the more fundamental point is that to teach chess concepts is much easier if they have a name.
Artosis (Starcraft caster) once said that the first thing he does when learning a new game is find a glossary, because most of the game's strategic ideas will be reflected there. I think this is true in chess also - pins, skewers, discovered checks, smothered mates, isolated pawns, doubled pawns, etc. are meaningful and identifiable in part because we have given these concepts specific names, and not just some cluster of vague ideas about space.
@CheckRaiseMate said in #17:
> I find this direction of research really exciting, but the example from the paper left me confused. What is the concept? I couldn't identify a common theme between the two positions.
The concept is one "discovered" by AlphaZero, so it's not part of the "human" concepts like space or king safety. The graph I included shows the concepts it's most related to. So it's kind of a mixture of all these easier concepts.