Self repressed from
lichess.org/forum/general-chess-discussion/is-learning-openings-that-important-and-how-can-i-create-my-own-opening-repertoire?page=4#36
@RamblinDave said in #35:
> Maybe a pedantic point, but unless you're sitting down at the board for your first move as white and working out what to do from first principles, you've already got an opening repertoire! And to the OP, looking at your database on here you seem to play pretty consistently, so I'd say that in that sense you've definitely got one.
not pedantic. I expand on that idea
Science-fiction, probably:
In an ideal world all of 960 setups should be our repertoire. I would refer to a deepmind paper using A0 as a model of learning in chess. (my words). Also my words: generalist at birth A0, repertoire biased (familiarity function, or knowledge function, or preference) specialist same A1.. Ak.. and some combination of all biases. Those were not with 960. it was standard, but I think it was through designed books policy biases in training. my understanding, is that while the generalist learns to use the its own previous training data as optimization substracte for its improment of moves given a position policy, and therefore at each RL batch, is improving its range of possible good enough moves given that past experience, it does what RL is boung to do, chooses a compromise through the dilemna of RL learning trajectories, between exploration and exploitation (improvement based on past exploration). As it become expert of its own experience set, it seems that, while it can beat any of the book specialist also trained to their expertise as book biased chess, it actually does not beat some ensemble or combination of the specialists as a new learner object. The details of that combination are still in my fog. But there are many ways, and not just the majority rule. What this shows, no matter how done, is that there are 3 levels of chess wilderness explored. and an ordering in the degrees of how generalist (meaning adapted to as many conditions as possible), or specialist a learner is.
The RL single learning trajectory, albeit starting a full potential generalist, as it has to settled down on ROTS of improvement, is making exploration shortcuts in favor of a tighter better games odds set of move choices.. so that at expertise cutoff, one would have explored a variable degree of leaner exploration degree (less generalist, in my words). The specialist, would have a reduced experience set or bubble of the chess wilderness, compared to that of compromised generalist (on the size of chess region explored over its entire leanring trajectory), still as this latter one at least spent its youth, exploring more.
But, then combining many biases... is better in the end. Think about this. How many of those. and how different are those biases. How does one even measure that. I hope that dimension of experimental design is above board in the paper details, or in the non-published researc behind. I suspect for now emprical trials and errors of book sets..
Could it be that chess theory, or more generally the combination skillset of many indviduals can be a human take home hypothesis? Or that one, should keep doing silly moves againt their own intuition, often enough. Maybe get some mind altering digestible or drunk, often enough to do such thing? Maybe reducing the knowledge game pressure in some class of competitive chess, might allow one playing always within the knowledge game included categories, to keep flexing some exploration muscle, where knowledge obedience would not be a dominating or lack of it, be a deleterious one. One can be wise with artificial selection mechanism, and consider new data, from outside the well established tradition that might act like some kind of tunnel vision. Existence of champion, does not prove much about how to get there. . and what gets there.
End of "science-fiction" probably.
Back to normality, or chess common sense (after five years, I know when I am rowing against the expectations, or I suspect it).
the rest is left in my last post there. as it was I hope not science-fiction to some of that thread possible audience (who knows what lichess users are made of.. are they all of the same cloth as it seems the overwhelming majority of teaching material is as nauseam reinforcing in its blatant assumptions, that have me wonder how the hell are they talking to.. certainly not me.
And why is everyone propagating the myths.. go read some science of development already... are you adult and playing here, studying chess not as your mean of living? Then I would say, they are not really talking to you, or they are hoping you are drinking the cool aid.
lichess.org/forum/general-chess-discussion/is-learning-openings-that-important-and-how-can-i-create-my-own-opening-repertoire?page=4#36
@RamblinDave said in #35:
> Maybe a pedantic point, but unless you're sitting down at the board for your first move as white and working out what to do from first principles, you've already got an opening repertoire! And to the OP, looking at your database on here you seem to play pretty consistently, so I'd say that in that sense you've definitely got one.
not pedantic. I expand on that idea
Science-fiction, probably:
In an ideal world all of 960 setups should be our repertoire. I would refer to a deepmind paper using A0 as a model of learning in chess. (my words). Also my words: generalist at birth A0, repertoire biased (familiarity function, or knowledge function, or preference) specialist same A1.. Ak.. and some combination of all biases. Those were not with 960. it was standard, but I think it was through designed books policy biases in training. my understanding, is that while the generalist learns to use the its own previous training data as optimization substracte for its improment of moves given a position policy, and therefore at each RL batch, is improving its range of possible good enough moves given that past experience, it does what RL is boung to do, chooses a compromise through the dilemna of RL learning trajectories, between exploration and exploitation (improvement based on past exploration). As it become expert of its own experience set, it seems that, while it can beat any of the book specialist also trained to their expertise as book biased chess, it actually does not beat some ensemble or combination of the specialists as a new learner object. The details of that combination are still in my fog. But there are many ways, and not just the majority rule. What this shows, no matter how done, is that there are 3 levels of chess wilderness explored. and an ordering in the degrees of how generalist (meaning adapted to as many conditions as possible), or specialist a learner is.
The RL single learning trajectory, albeit starting a full potential generalist, as it has to settled down on ROTS of improvement, is making exploration shortcuts in favor of a tighter better games odds set of move choices.. so that at expertise cutoff, one would have explored a variable degree of leaner exploration degree (less generalist, in my words). The specialist, would have a reduced experience set or bubble of the chess wilderness, compared to that of compromised generalist (on the size of chess region explored over its entire leanring trajectory), still as this latter one at least spent its youth, exploring more.
But, then combining many biases... is better in the end. Think about this. How many of those. and how different are those biases. How does one even measure that. I hope that dimension of experimental design is above board in the paper details, or in the non-published researc behind. I suspect for now emprical trials and errors of book sets..
Could it be that chess theory, or more generally the combination skillset of many indviduals can be a human take home hypothesis? Or that one, should keep doing silly moves againt their own intuition, often enough. Maybe get some mind altering digestible or drunk, often enough to do such thing? Maybe reducing the knowledge game pressure in some class of competitive chess, might allow one playing always within the knowledge game included categories, to keep flexing some exploration muscle, where knowledge obedience would not be a dominating or lack of it, be a deleterious one. One can be wise with artificial selection mechanism, and consider new data, from outside the well established tradition that might act like some kind of tunnel vision. Existence of champion, does not prove much about how to get there. . and what gets there.
End of "science-fiction" probably.
Back to normality, or chess common sense (after five years, I know when I am rowing against the expectations, or I suspect it).
the rest is left in my last post there. as it was I hope not science-fiction to some of that thread possible audience (who knows what lichess users are made of.. are they all of the same cloth as it seems the overwhelming majority of teaching material is as nauseam reinforcing in its blatant assumptions, that have me wonder how the hell are they talking to.. certainly not me.
And why is everyone propagating the myths.. go read some science of development already... are you adult and playing here, studying chess not as your mean of living? Then I would say, they are not really talking to you, or they are hoping you are drinking the cool aid.