I am a bit confused as the article mentions that the following 2 games have the same concept but it doesn't explain what the concept is. (Later it explains for 1 that Bg5 lures/forces h6 but I don't see how Qe2 in the second position is related to this).
@jk_182 said in #20:
> 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.
I get that, but the exciting thing to me is, what if AlphaZero discovered something that isn't currently considered to be a core chess concept, but once you saw it, you went, "Oh yeah that totally makes sense"?
So far, there doesn't seem to be anything like that. With the example in the post, even after seeing the positions and the related concepts, I can't understand or articulate what the concept is. So I don't see how I can use that to get better at chess.
I'd also argue there's a bit of a philosophical question of, if something is impossible to articulate, is it really a concept, or just sort of a random amalgamation of information inside AZ's weights?
@CheckRaiseMate said in #22:
> I get that, but the exciting thing to me is, what if AlphaZero discovered something that isn't currently considered to be a core chess concept, but once you saw it, you went, "Oh yeah that totally makes sense"?
yep. We need to be curious and allow such experimental statistical instruments to be further studied. But the article has certain audiences to be careful with. The bridging purpose is in the title. But I think you are hitting a ripe nail, there.
It means, humans, this is not just an individual performance obsession pursuit. We might learn from many sources.
chess theory might get some fresh air from more than one expert at a time, not peer reviewed, spilling their guts in sparse books. (and burying that in full games...). ok too much pamphletting. I stop.
I don't think the culture has the habit to listening to reason if not vetted by individual performance title (which is not saying it is non-sensical, but it is not asking explicitly, what do we even want to know about chess, and what chess?.
@jk_182 said in #20:
> 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.
But I think the comment may have been to not only consider the bridging requirement. That there might be recognizable by humans, but not those that each individual humans could share so far in their own pattern leanrning possibly path of leanring dependent trajectory, that we have as a community manage to share with each other. Or that present liviing experts each also having possible learning path** dependent pattern "definitions", that they might not be able to verbalize. Also, big silence on the learning data still sitting in the A0 experiements.. Whey we might get more human clues about many game chess not coming from our historical many humans paths to expertise so far. Well, the comment might not be about that. But related: There might be other patterns for us human to discover from those that might be embeded in the many A0 learning trajectories (not just one). Ok even that just one learned instance. The engine evaultoin is not the only clustering pressure to use (is that correct assumption about the paper?). I understand the bridging question, as important question, but coud there not be new chess theory hypotheses if we drop the restricitno of it has to be already in some 5 individuals representant we hope might have themselves internal model pattern definitions aligned, to recognised. We certainly do not use the same leanrning signals on the outcome side (not only).
But the chess input transformation through the learning NN common stem between evluatoin and policy, itself might contain things that human could research for new hypotheses. This might be my own interpretation of the A0 as source of chess theory hypotheses. They might not be A0 "discovered" depending on the method used in paper to implementat that phrase. A0 discovered concept.
**learning paths in chess position space, for example, and if making sense to others, as it does for me, from long ago, the space support I mean, I am stuck with it, and all the chess machine tools, keep bugging my curiosity there.. sorry that blurted out, and now too late, I want to share that too).
@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.
According to the paper, the concept "possesses both strategic and prophylactic characteristics, involving plans that improve the player’s piece placement while restricting the opponent’s piece activity. This concept contains an additional element of exploiting tactical motifs and weaknesses, combining strategic and tactical play."
Seems a bit too broad to give a simple name to.
@CheckRaiseMate said in #22:
> I get that, but the exciting thing to me is, what if AlphaZero discovered something that isn't currently considered to be a core chess concept, but once you saw it, you went, "Oh yeah that totally makes sense"?
They listed the "known" concepts in the appendix and the tables were over 3 pages long, so I'd guess that the new concepts won't be so easy to articulate. Maybe it's more like different concepts playing together. I think that one would need to see more examples in order to drill down on what the concept might actually be. In general, I thought that the approach to learning the concepts in the paper was a bit questionable.
I actually found the part about extracting existing concepts from AlphaZero more interesting, since it could help searching for games where a specific theme appears and might also help in identifying in which kinds of positions a player makes mistakes.
@jk_182 said in #26:
> I actually found the part about extracting existing concepts from AlphaZero more interesting, since it could help searching for games where a specific theme appears and might also help in identifying in which kinds of positions a player makes mistakes.
The point which @CheckRaiseMate and others have made remains: if we can't articulate these 'AlphaZero concepts' in any kind of qualitative, meaningful way--if these concepts are purely quantitative in nature, weights which are correlated in a certain way--then "which kinds of positions a player makes mistakes" is also meaningless to us.
This doesn't mean AlphaZero is incomprehensible to us entirely or that this isn't an interesting research direction! It very much fascinates me.
What's exciting about this research direction is that it might make parts of AlphaZero understandable that were not previously. And @CheckRaiseMate is, I think, fundamentally correct that nothing presented here accomplishes that.
Concepts might be in the subconscious for a long time, and the need to articulate, does not seem required for action; one could review some Carlsen interviews about how he might think about his own moves that challenge concepts already shared by some pioneers of chess theory in the past.
That sparse uncovering through language that I might have mentioned, both in density per position exposure through lists of full games as the norm of sharing, and across experts who could have weighed with their own reasoning in contemporary setting, seems to happen through books at very slow pace, as pioneering from each individual intuition extirpation.
There may have been off-record peer verbal review debates. It would be interesting to find traces, of common problem discussion, and discrepancies of subjectivities being made explicit there. Not one book, one generation at a time. But that might be the stage of the nonexistent science, that most of us don't really care about as such.
I believe there is space (chess is big) for the quirks of historical paths to have shaped of how a group communicates both horizontally and vertically—through the documents of previous generations and their dissemination—as we contemplate the concepts of the board. (AI edited this paragraph, then own editing. somehow that help is double work).
It might stem from that, individually, without a language for some of the concepts to align with conscious (articulately, or arguably, not in the dispute necessarily). We can't really share such non-verbal concepts, wihtout first doing the individual work from performance expertise (or if it were allowed, about the bigger question of learning, as one is learning, about concepts and how they evolve, but that is not really the trend, concept seem to be having some absolute quality...).
If it was pure experience like for A0, and A0s had as much diversity as humans entering chess land for a long duration, allowing some learning of concepts, but only from experience, perhaps we could compare with humans not being able to share past some level.
I think that GMs might evolve beyond the commonly shared concepts, and working hard toward performance goal, unless they take some time to articulate some theories, and maybe if it were confronted to others live before committing to book, we could have some clues about my current blob of questions.
The big space of chess, might also big a big space for different learning trajectories. (I might have themes and repetition, sorry, beyond my editing polishing abilities).
Humans did not just learn by pure individual performance selection and experience. Thy tried to look beyond the mere individual performance learning. Each book that would create new theory must have been, by some of those individual performance experts, trying to articulate their internal model after years of experience.
This sparse population method of uncovering some chess science (pardon the word) might have given us the impression, that the current chess theory might actually be representative of what it would have been for each of us; yes, learning does happen before one becomes an expert. I mean without it, we assume we still would have the same internal concepts, that the words are the concept, might lend to that assumption.
But then since the GM (or anyone that can play well but not explain how or why they did that move, they just had the intuition for it.
The A0 concept that they had to articulate as data scientists out of A0 (not a large language model type of AI... btw)
not being verbalizable (articulately), might be about it being a model of our subconscious learning, not all of our mind learning full model, with conscious and language to help each other.
Perhaps, one A0 instance, as a model of human chess thinking (learned enough=big ELO) is modelling only one naive uneducated chess learning trajectory, as human.
Why I think it would be interesting to look at another interesting Deemind paper about chess. Where that notoin of population of learners, is made a bit more visible. I have been thinking like that for while. if not since I started looking at chess on lichess as something to pursue. The variance. Only we might need so metric about chess being big, but how big (and what chess, and just couniting, or are there other measures to bear on such idea of learning paths and position space being relatable...). I also find the questions applicable to all engine evolution, where does that happen, is not being asked.