I find the boards to have better proportions here. Not significant comment. Thanks for your work and choices of subjects. This is long haul for me. I have other papers from them, that I need to read, too.
Calculatin ability of A0? I read bacward. did conclusion. agreed. now going up.
so. I also wonder about the difference. But calculation, I am not sure it is where the machine not having limitations. My inkling was more in the very evualation probabliity function training (or its policy if still needing that, beyond one node policy, to be beat other engines, that is). That we can't make such long term association statistics, perhaps, with such memory and exactitiude. Although now I wonder about what our subsconsious brain does capture, over different time controls about all the positions and outcomes of many games. That we don,t consciously have recalling memory of them, does that mean our less conscious observer that such ML expeirements are modeling are leaky too.
which might be what you meant about calculation abitily. The statistics from outcome to whole games mamy positoins over many games. That we conscious need to build chunks surrogate plan stepping stones is not necessariyly meaning that our statistical brain is not doing similar long term assocation.
now again having coutner thought. sorry. i think through self debate, i guess ok. if i warn..
Now, even if A0 in in learning of all its layers, and transformation needs from the input layer, which we are ourselves also learning from, even with a bad tranformation of the position for deeper layers feature processing all the way to the last dense layer than can't tranfrom much but which can discrimiate from a well transformed input (my narrative of ideal NN learning in such convolution motifs (not even worrying about which activation function or intermotif cleverness, as ignorant). A certain division of labor (there are papers about such classes of NN and certain tasks that might support this story). My point is that we might have more fading short term memory about the specifics of positions, on top our that long term learning problem (games length assocations), so my point about not knowing subsconscious abilites about whole games range of associations to learn, might be "trumped" by if the game was a blur, our subconscious might not fix the blur, so we might need a lot more training. Which might be in the line of your comment.
With calculation ability I meant the comparison between AlphaZero and humans. Some examples in the paper hinge on variations to really justify them, so I think that humans often struggle more with spotting the details in the variations, rather than seeing the concept in the starting position.
So really calculation. It appears that the A0 concept might be less about the evaluation head and more about dynamics patterns. From still not having read the thing. (mea culpa). I am preparing to read and have enough questions to keep me on task. As humans it is possible that the L0 eval head or as I think a more evaluation agnostic (for us as data exploration) concept search of the common NN part,, might allow more static concepts that imply dynamic ones, if i understand your last sentence, association potential with the root positions (starting the chess problem for the human or machine).
I find the boards to have better proportions here. Not significant comment. Thanks for your work and choices of subjects. This is long haul for me. I have other papers from them, that I need to read, too.
Calculatin ability of A0? I read bacward. did conclusion. agreed. now going up.
so. I also wonder about the difference. But calculation, I am not sure it is where the machine not having limitations. My inkling was more in the very evualation probabliity function training (or its policy if still needing that, beyond one node policy, to be beat other engines, that is). That we can't make such long term association statistics, perhaps, with such memory and exactitiude. Although now I wonder about what our subsconsious brain does capture, over different time controls about all the positions and outcomes of many games. That we don,t consciously have recalling memory of them, does that mean our less conscious observer that such ML expeirements are modeling are leaky too.
which might be what you meant about calculation abitily. The statistics from outcome to whole games mamy positoins over many games. That we conscious need to build chunks surrogate plan stepping stones is not necessariyly meaning that our statistical brain is not doing similar long term assocation.
now again having coutner thought. sorry. i think through self debate, i guess ok. if i warn..
Now, even if A0 in in learning of all its layers, and transformation needs from the input layer, which we are ourselves also learning from, even with a bad tranformation of the position for deeper layers feature processing all the way to the last dense layer than can't tranfrom much but which can discrimiate from a well transformed input (my narrative of ideal NN learning in such convolution motifs (not even worrying about which activation function or intermotif cleverness, as ignorant). A certain division of labor (there are papers about such classes of NN and certain tasks that might support this story). My point is that we might have more fading short term memory about the specifics of positions, on top our that long term learning problem (games length assocations), so my point about not knowing subsconscious abilites about whole games range of associations to learn, might be "trumped" by if the game was a blur, our subconscious might not fix the blur, so we might need a lot more training. Which might be in the line of your comment.
With calculation ability I meant the comparison between AlphaZero and humans. Some examples in the paper hinge on variations to really justify them, so I think that humans often struggle more with spotting the details in the variations, rather than seeing the concept in the starting position.
So really calculation. It appears that the A0 concept might be less about the evaluation head and more about dynamics patterns. From still not having read the thing. (mea culpa). I am preparing to read and have enough questions to keep me on task. As humans it is possible that the L0 eval head or as I think a more evaluation agnostic (for us as data exploration) concept search of the common NN part,, might allow more static concepts that imply dynamic ones, if i understand your last sentence, association potential with the root positions (starting the chess problem for the human or machine).
Stimulating choice of paper (series).
Can we follow any strategy that can lead any chess match a draw?