A0 subtlety
Posted: Mon Feb 11, 2019 5:17 pm
Was a set of input planes representing a position at time (t - T + 1) oriented for the side to move at time t or time (t - T + 1)?
Would it matter if they weren’t?Daniel Shawul wrote: ↑Mon Feb 11, 2019 6:08 pmAll the history input planes are oriented for the current side to move, i.e. at time t.
I suppose not but it would make it a little bit harder because the network has to learn to flip the boardchrisw wrote: ↑Thu Feb 14, 2019 4:28 pmWould it matter if they weren’t?Daniel Shawul wrote: ↑Mon Feb 11, 2019 6:08 pmAll the history input planes are oriented for the current side to move, i.e. at time t.
(That’s a serious question btw)
I thought it would depend on how the convolution filters were tracking the data (in what order). Does it track 3x3 across each 8x8 map of chessboard cells, and then combine those outputs - so I suppose it must be best to combine a1 to a1 region and so on, rather than a1 to a8. Difficult to wrap head around.Daniel Shawul wrote: ↑Thu Feb 14, 2019 5:14 pmI suppose not but it would make it a little bit harder because the network has to learn to flip the boardchrisw wrote: ↑Thu Feb 14, 2019 4:28 pmWould it matter if they weren’t?Daniel Shawul wrote: ↑Mon Feb 11, 2019 6:08 pmAll the history input planes are oriented for the current side to move, i.e. at time t.
(That’s a serious question btw)
every step in the history. I do not think history planes matter much anyway to begin with.
I agree with this. However, it's very likely that the history inputs are mostly irrelevant, as Daniel Shawul mentioned.chrisw wrote: ↑Thu Feb 14, 2019 6:44 pmI thought it would depend on how the convolution filters were tracking the data (in what order). Does it track 3x3 across each 8x8 map of chessboard cells, and then combine those outputs - so I suppose it must be best to combine a1 to a1 region and so on, rather than a1 to a8. Difficult to wrap head around.Daniel Shawul wrote: ↑Thu Feb 14, 2019 5:14 pmI suppose not but it would make it a little bit harder because the network has to learn to flip the boardchrisw wrote: ↑Thu Feb 14, 2019 4:28 pmWould it matter if they weren’t?Daniel Shawul wrote: ↑Mon Feb 11, 2019 6:08 pmAll the history input planes are oriented for the current side to move, i.e. at time t.
(That’s a serious question btw)
every step in the history. I do not think history planes matter much anyway to begin with.
well, for me, it is clear that they, or some form of equivalence, are vital. Attempts at explanations have met with uniform blank looks or raised eyebrows, so I’ll just stick with this view, and we’ll see. It won’t be the first time.AlvaroBegue wrote: ↑Thu Feb 14, 2019 11:55 pmI agree with this. However, it's very likely that the history inputs are mostly irrelevant, as Daniel Shawul mentioned.chrisw wrote: ↑Thu Feb 14, 2019 6:44 pmI thought it would depend on how the convolution filters were tracking the data (in what order). Does it track 3x3 across each 8x8 map of chessboard cells, and then combine those outputs - so I suppose it must be best to combine a1 to a1 region and so on, rather than a1 to a8. Difficult to wrap head around.Daniel Shawul wrote: ↑Thu Feb 14, 2019 5:14 pmI suppose not but it would make it a little bit harder because the network has to learn to flip the boardchrisw wrote: ↑Thu Feb 14, 2019 4:28 pmWould it matter if they weren’t?Daniel Shawul wrote: ↑Mon Feb 11, 2019 6:08 pmAll the history input planes are oriented for the current side to move, i.e. at time t.
(That’s a serious question btw)
every step in the history. I do not think history planes matter much anyway to begin with.
I always evaluate for white to move, so the board has to be flipped horizontally when it is black to move and ofcourse all moves are flipped as well