Excellent description of one node. And good blog idea. More fresh engine air.
Also, as adult** learners, we come equipped with more self-cognitive awareness.. Best make use of that by being meticulour about own thought process, and tracking own thinking, to that one can maximize the information juice out of any experience.. One can find a whole spectrum of error (not just trinary bins based on one engine score), and with such self-awareness trakcing, even be able to retrace the fine grain thinking from errors (be them own post-game analysis without engine based on the ultimate outcome as only feedback. Where did i really lose, was it a last minute tactical blunder wasting my previous good play? or did i make an earlier more strategic, or positional choice with such consequence or both (e.g. the postional choice, led to sharper territory where not keeping sharp tactical edge meant short term blunder).
I assume that the word documents might have contained all sorts of self awareness notes.. allowing spaced looking back for patterns of persistent less good play.. on many levels...
Defnitely agreeing with your ingredients.
Let the adults many walks of non-chess life and chess-life be creative. They are by far the most expert about their learning or ignorance level status... There is teaching data there.
Horsey for refreshing angles. The vision of that horse must be a unique perspective on things.
**(i have played young though but always for pure pleasure of the game, the one game itself, and never in any social competition layer above that). also long hiatus even a that low intensity level. I am still not competitive level but i can subjectively say i understanding chess increasingly, which is my humble goal, but i am sure my experience is not completely different from perfromance goal minded people using rating improvement as measure.
Also, as adult** learners, we come equipped with more self-cognitive awareness.. Best make use of that by being meticulour about own thought process, and tracking own thinking, to that one can maximize the information juice out of any experience.. One can find a whole spectrum of error (not just trinary bins based on one engine score), and with such self-awareness trakcing, even be able to retrace the fine grain thinking from errors (be them own post-game analysis without engine based on the ultimate outcome as only feedback. Where did i really lose, was it a last minute tactical blunder wasting my previous good play? or did i make an earlier more strategic, or positional choice with such consequence or both (e.g. the postional choice, led to sharper territory where not keeping sharp tactical edge meant short term blunder).
I assume that the word documents might have contained all sorts of self awareness notes.. allowing spaced looking back for patterns of persistent less good play.. on many levels...
Defnitely agreeing with your ingredients.
Let the adults many walks of non-chess life and chess-life be creative. They are by far the most expert about their learning or ignorance level status... There is teaching data there.
Horsey for refreshing angles. The vision of that horse must be a unique perspective on things.
**(i have played young though but always for pure pleasure of the game, the one game itself, and never in any social competition layer above that). also long hiatus even a that low intensity level. I am still not competitive level but i can subjectively say i understanding chess increasingly, which is my humble goal, but i am sure my experience is not completely different from perfromance goal minded people using rating improvement as measure.
Great short thought piece! I'm still skeptical of Stockfish on one node actually playing like a human. What about "Maya", the computer that emulates human play--I believe it studies that database of common moves at certain ratings. I wish there were more such resources. Either way, I like the second approach more. And yes--not only studying chess, but studying "how to study" and then taking ownership has taken me from 1000 to 1800 (Chess.com Rapid) in two years.
@Tilt-Shift said in #3:
> Great short thought piece! I'm still skeptical of Stockfish on one node actually playing like a human.
I think the author was referring to LC0. leela chess zero. not SF. SF needs those in-game node searches. it is in its basic core design. The type A (or is it B) engine. Exhaustive search, with heuristics amendments during is version evolution to get deeper faster (both with computer hardware improving, and the algorithm modficiations themselves).
LC0, is doing its evolution first during training, and then gets tested in trounaments. There might be also the same kind of outer evolution loop about the training heuristics (about optmisiation not about the relationships between knowledge or outcome data and how train within that).
Maya, is leela inert architecture before training, but trained with hand-crafted error models. They are evovling such error models. But not clear to me if they learn the good chess at the same times, so an expansion on lc0 to include some rating variables in their models, or some entrant assumption about what is best play, and some entrant assumption about error model per rting bin.
Leela is a human style good chess. Maya is human error model approximation. as a caricature (my statement is). Trying to model sub-optimal chess with only rating bands might be a bit clunky. as are the sub optimal versions of SF... (playing on some internal parameters).. Like someone wrote.. stretches of good play, and then random blunders. not really the same spectrum of human blunders. in that case, and not really the human good play with errors in the other case.
so using leela without calculation depths would be playing its odds static evaluation on all candidates moves, taking the best. Such static evaluation (one node) is a statistics being fed only by hard terminal outcome data over many training games. not shallower heuristic evaluations. AS far as i understand. It does make tight tactical mistakes though, I have been told.
But my understanding is that all such terminal end points being "integrated" back to early positions, increase in confidence the earlier you start. (as their RL method is starting all its games from the standard initial position, I think still is, meaning, in its uniform initial exploration it is going to do a lot of mistakes and terminal end point near that position, and as the self-play batches evolve, longer games providing deeper and deeper outcomes to feed the statistical grinder (and shape the NN latent space transformation of the basic input vector that is itself lossless encoding of the full position information set).
sorry got more technical and possibly outdated than necessary.. to clarify the big divide in design between SF and leela. 2 species of engine. not really cousins..
warning: the confidence thing while reasonable statements might be too simplest, and is my understanding, of the hypothesis king.. I might have read stuff that support that. others might correct me (with advance gratitude from me, i believe in dialog more than lectures, or I just can't follow one but the other yes). Same goes with maya model, I might be in the wrong about it.
> Great short thought piece! I'm still skeptical of Stockfish on one node actually playing like a human.
I think the author was referring to LC0. leela chess zero. not SF. SF needs those in-game node searches. it is in its basic core design. The type A (or is it B) engine. Exhaustive search, with heuristics amendments during is version evolution to get deeper faster (both with computer hardware improving, and the algorithm modficiations themselves).
LC0, is doing its evolution first during training, and then gets tested in trounaments. There might be also the same kind of outer evolution loop about the training heuristics (about optmisiation not about the relationships between knowledge or outcome data and how train within that).
Maya, is leela inert architecture before training, but trained with hand-crafted error models. They are evovling such error models. But not clear to me if they learn the good chess at the same times, so an expansion on lc0 to include some rating variables in their models, or some entrant assumption about what is best play, and some entrant assumption about error model per rting bin.
Leela is a human style good chess. Maya is human error model approximation. as a caricature (my statement is). Trying to model sub-optimal chess with only rating bands might be a bit clunky. as are the sub optimal versions of SF... (playing on some internal parameters).. Like someone wrote.. stretches of good play, and then random blunders. not really the same spectrum of human blunders. in that case, and not really the human good play with errors in the other case.
so using leela without calculation depths would be playing its odds static evaluation on all candidates moves, taking the best. Such static evaluation (one node) is a statistics being fed only by hard terminal outcome data over many training games. not shallower heuristic evaluations. AS far as i understand. It does make tight tactical mistakes though, I have been told.
But my understanding is that all such terminal end points being "integrated" back to early positions, increase in confidence the earlier you start. (as their RL method is starting all its games from the standard initial position, I think still is, meaning, in its uniform initial exploration it is going to do a lot of mistakes and terminal end point near that position, and as the self-play batches evolve, longer games providing deeper and deeper outcomes to feed the statistical grinder (and shape the NN latent space transformation of the basic input vector that is itself lossless encoding of the full position information set).
sorry got more technical and possibly outdated than necessary.. to clarify the big divide in design between SF and leela. 2 species of engine. not really cousins..
warning: the confidence thing while reasonable statements might be too simplest, and is my understanding, of the hypothesis king.. I might have read stuff that support that. others might correct me (with advance gratitude from me, i believe in dialog more than lectures, or I just can't follow one but the other yes). Same goes with maya model, I might be in the wrong about it.
Thanks for the lengthy explanation. I didn't know any of that about the computers. But even so, I think I'll probably stick to playing humans! There really isn't any reason not to, unless you really want to grind out a certain opening over and over, or some endgame technique.
!!!!!!!!!!!!!!!!!!!!!!!
The creative process in chess involves contemplation, imagination, and exploration of possibilities. Players must analyze the position, envision different scenarios, and assess the potential outcomes of their moves. This process mirrors the creative thinking employed in other artistic endeavors, such as painting or composing music.
Cool article on how to mix things up in training – I would say that until computers started to play a little more human-like, I was pretty skeptical about using computers as a training tool other than opening prep or blunder checking because the older version of Stockfish tend to beat humans in a very inhuman manner.
However, I am a little hesitant to point to Matthew Sadler as an example of adult improvement. He was already a world class player when he retired, and while I don't doubt he is still a very strong player and that his work with computers is useful, I think his near-2700 FIDE rating is at least a little inflated. He plays primarily in 4NCL tournaments where I believe there's a fair amount of time to prep thoroughly for what looks like a smallish pool of opponents who are mostly on the older side. If Matthew Sadler had to play an open tournament and hold his rating against up and coming junior players more regularly, I think his rating would drop quite a bit.
However, I am a little hesitant to point to Matthew Sadler as an example of adult improvement. He was already a world class player when he retired, and while I don't doubt he is still a very strong player and that his work with computers is useful, I think his near-2700 FIDE rating is at least a little inflated. He plays primarily in 4NCL tournaments where I believe there's a fair amount of time to prep thoroughly for what looks like a smallish pool of opponents who are mostly on the older side. If Matthew Sadler had to play an open tournament and hold his rating against up and coming junior players more regularly, I think his rating would drop quite a bit.
> It’s a very simple technique, yet not one that I had ever heard of before. If you want to try this technique, there are instructions for setting up Leela in the “supplementary material” link on the website for Sadler’s book The Silicon Road to Chess Improvement.
Just to mention, this is not new (unless the article is old), there were even bots running 1-node LC0, like @Leela1Node or @leela-1node
Just to mention, this is not new (unless the article is old), there were even bots running 1-node LC0, like @Leela1Node or @leela-1node
Anyone know how to set up Leela to play with one node? I can't afford a new book right now but am interested to try this method.