move ranking value += dot(positional embedding of move, dynamically learned mapping from piece_square to vector). Only update weights once every refresh (256 nodes) or when the piece moves...
I'll code it soon. Let's find out.
noble basin
move ranking value += dot(positional embedding of move, dynamically learned mapping from piece_square to vector). Only update weights once every refresh (256 nodes) or when the piece moves...
I'll code it soon. Let's find out.
Okay, sqrt(2) approximation is 239/169.
That would be used in the positional embedding.