25,161
edits
Changes
no edit summary
Since the 9x9 [[Shogi]] board has a centered king file and [[Castling|castling]] is not known in Shogi, [[Color Flipping|color flip]] versus[[Flipping Mirroring and Rotating#Rotationby180degrees|180 degree rotate]] differs in a [[Horizontal Mirroring|horizontal mirrored]] position from the other side's point of view, with otherwise identical playing options. The NNUE is trained and probed from the side to move point of view, where the used 180 degree rotation (xor 63 instead of 56) to flip sides looks rather strange for chess <ref>[https://github.com/official-stockfish/Stockfish/blob/615d98da2447e79ceceae205e0cd4e878115acc3/src/types.h#L323 Stockfish/types.h at 615d98da2447e79ceceae205e0cd4e878115acc3 · official-stockfish/Stockfish · GitHub]</ref>. i.e color flipping the white king from e1 to d8 rather than e8. Does it consider castling short on the queen side?
It is a little bit unclear, how that rotation rather than flip affects the playing strength <ref>[https://github.com/official-stockfish/Stockfish/issues/3021 NNUE eval rotate vs mirror · Issue #3021 · official-stockfish/Stockfish · GitHub] by [[Terje Kirstihagen]], August 17, 2020</ref> and whether NNUE for chess or its developers suffer from [https://en.wikipedia.org/wiki/Associative_visual_agnosia associative visual agnosia].
Maybe Fishtest needs to play many games with color-flipped openings, i.e. 1.e4 e5 and 1.e3 e5 2.e4, to look whether results differ or not.
Anyway, a fix from rotate to flip has to be done from producer and consumer sides, and is likely to void some training sessions.
