@sicariusnoctis said in #82:
But then, by such a definition, chess is effectively random to the player, with the randomness modeled by probability distribution dependent on candidate moves, which are weighted via a pruned tree search to some depth. (With the typical assumption that the opposing player is trying to win, and will look at the expected candidate moves.)
Or you could start your point of view from the mathematical framework A0 and Lc0 were built on, which does not rely on heuristics of evaluation from positions that are not terminal positions.
It has a full position information lossless model, through all the layers. And uses core ruleset mobility rules and termination rules. no other evaluation posited criteria (AFAI understand or understood once).
make for more seamless thinking about randomness in chess.
Also, true distribution. My question was probably kind of circular.. But positing a distribution true or estimated by player limited experience and or knowledge, is what i think is the definition act for randomness.
It only means which function space are you maximally attributing to legal move set AND position to apply move to combined domain (input of said distribution or density if sloppy like me with words across finite versus continuum ambient spaces).
However it would be nice to be able to have a common big enough domain where we could interpret both hand-crafted and material counting borne engine species development, and the very late introduced more empiriical approach (minimal handcrafting of evaluation as function of non terminal position information).
I have stopped trying to do that.. with the hybridization that came with NNue... But the hand-crafting might still be present or filtered by NNue input feature reduction from the full complete information set preserving A0 and Lc0 input framework.
The big information flow loop for NNue which enables its tournament speed advantages, is however limiting such evolution from previously single parameter optimization history for all the potential features contained in the "repressed" classical evaluation function which I heard was meant to be frozen. I thought the plan was to keep using NNue to include possibly more features, if there would open the information flow loop to be not just about approximating the frozen classical eval SF moderate search output. sorry i am not doing such long sentences on purpose.. the whole loop is needed to understand where things got stuck in my opinion. (i blame tournament cultural inertia to some extent.... priority on speed over true accuracy).
sorry i diggressed. but what is randomness. i meant to say even the true distribution posit is the only meaning of randomness.
other than convention of some never defined perception of what random could mean..
normal distribution assumption or model. is just restricting the uncertainty placeholder function to a few parameter function space spanned by exponential familiy, and intergral operator... i think 2 parameters.
one can make lots of math linking various constraints of interacting random variables and differently characterized respective function spaces.
and yes.. increasing knowledge is about constraining even more the degree of ignorance left using metrics as you suggested... to figure out what is informative or not in the data to distribution shaping evoluation loop (for a0 lc0 that evoluation is whitin training, for SF and fishtest it is across SF versioning, that is where the simpler framework under A0 and lc0. has a hard time in my mind applying).
many levels of possible ignorance models.. choosing a metric also has some implications on how the knowldge ignorance debate evolves. Human opening theories might also have their effect on admissible function space.. but that seems to be part of chess for both humans and engines so far. thanks for bringing your compatible point of view..
engines have their own design constraints on those.. i understand well i think the A0 and lc0 basic learning plan. but i find it difficult to bathe SF evaluation into it.. (some ignorance about it top most functional skeleton, and internal position information data structure which has been ground in many meanders of partial informations to be saving on cost of computation at any code branching points.. over a long history or code amendements... using intricate efficient coding and execution cost tentacular data structures... etc... needs a translation back to math. exhausting to do from outside.
how can i chop this post into on topic and off topic but not that off.. autonomous readable chunks.. i do find all those topics to be related.. but i have a visuo spatial memory support and i am trying to make a path in that to protrude it in this linguistic stream.. but the multiple probably logical connections keep popping and i wonder, did I start from the most reader digestible way.. grab what you can... and dismiss the possible BS.
@sicariusnoctis said in #82:
> But then, by such a definition, chess is effectively random to the player, with the randomness modeled by probability distribution dependent on candidate moves, which are weighted via a pruned tree search to some depth. (With the typical assumption that the opposing player is trying to win, and will look at the expected candidate moves.)
Or you could start your point of view from the mathematical framework A0 and Lc0 were built on, which does not rely on heuristics of evaluation from positions that are not terminal positions.
It has a full position information lossless model, through all the layers. And uses core ruleset mobility rules and termination rules. no other evaluation posited criteria (AFAI understand or understood once).
make for more seamless thinking about randomness in chess.
Also, true distribution. My question was probably kind of circular.. But positing a distribution true or estimated by player limited experience and or knowledge, is what i think is the definition act for randomness.
It only means which function space are you maximally attributing to legal move set AND position to apply move to combined domain (input of said distribution or density if sloppy like me with words across finite versus continuum ambient spaces).
However it would be nice to be able to have a common big enough domain where we could interpret both hand-crafted and material counting borne engine species development, and the very late introduced more empiriical approach (minimal handcrafting of evaluation as function of non terminal position information).
I have stopped trying to do that.. with the hybridization that came with NNue... But the hand-crafting might still be present or filtered by NNue input feature reduction from the full complete information set preserving A0 and Lc0 input framework.
The big information flow loop for NNue which enables its tournament speed advantages, is however limiting such evolution from previously single parameter optimization history for all the potential features contained in the "repressed" classical evaluation function which I heard was meant to be frozen. I thought the plan was to keep using NNue to include possibly more features, if there would open the information flow loop to be not just about approximating the frozen classical eval SF moderate search output. sorry i am not doing such long sentences on purpose.. the whole loop is needed to understand where things got stuck in my opinion. (i blame tournament cultural inertia to some extent.... priority on speed over true accuracy).
sorry i diggressed. but what is randomness. i meant to say even the true distribution posit is the only meaning of randomness.
other than convention of some never defined perception of what random could mean..
normal distribution assumption or model. is just restricting the uncertainty placeholder function to a few parameter function space spanned by exponential familiy, and intergral operator... i think 2 parameters.
one can make lots of math linking various constraints of interacting random variables and differently characterized respective function spaces.
and yes.. increasing knowledge is about constraining even more the degree of ignorance left using metrics as you suggested... to figure out what is informative or not in the data to distribution shaping evoluation loop (for a0 lc0 that evoluation is whitin training, for SF and fishtest it is across SF versioning, that is where the simpler framework under A0 and lc0. has a hard time in my mind applying).
many levels of possible ignorance models.. choosing a metric also has some implications on how the knowldge ignorance debate evolves. Human opening theories might also have their effect on admissible function space.. but that seems to be part of chess for both humans and engines so far. thanks for bringing your compatible point of view..
engines have their own design constraints on those.. i understand well i think the A0 and lc0 basic learning plan. but i find it difficult to bathe SF evaluation into it.. (some ignorance about it top most functional skeleton, and internal position information data structure which has been ground in many meanders of partial informations to be saving on cost of computation at any code branching points.. over a long history or code amendements... using intricate efficient coding and execution cost tentacular data structures... etc... needs a translation back to math. exhausting to do from outside.
how can i chop this post into on topic and off topic but not that off.. autonomous readable chunks.. i do find all those topics to be related.. but i have a visuo spatial memory support and i am trying to make a path in that to protrude it in this linguistic stream.. but the multiple probably logical connections keep popping and i wonder, did I start from the most reader digestible way.. grab what you can... and dismiss the possible BS.