For neural networks to work well, the feature representation needs to be relatively smooth in how
the input space is mapped to the output space. Positions that are close together in the feature
space should, as much as possible, have similar evaluations.
This means many intuitive and naive representations are not likely to work well. For example,
many previous attempts at using neural networks for chess represent positions as bitmaps, where
each of the 64 squares is represented using 12 binary features indicating whether a piece of each
of the 12 piece types exist on the square or not. This is unlikely to work well, because in this
768-dimensional space, positions close together do not necessarily have similar evaluation...
A much better representation is to encode the position as a list of pieces and their coordinates. This
way, positions that are close together in the feature space would usually have similar evaluation.