Would it be worth applying the same kind of symmetry-exploiting that goes on in AlphaGo (Zero) with respect to king- and queenside? Castling would not be an issue, because it can still be considered a symmetric move in the sense that it involves an unmoved king moving two squares in one direction with the corresponding rook ending up behind it (although this is admittedly different from how it works in Chess960). Just don't forget the castling rights when mirroring a game state.
Would it be worth applying the same kind of symmetry-exploiting that goes on in AlphaGo (Zero) with respect to king- and queenside? Castling would not be an issue, because it can still be considered a symmetric move in the sense that it involves an unmoved king moving two squares in one direction with the corresponding rook ending up behind it (although this is admittedly different from how it works in Chess960). Just don't forget the castling rights when mirroring a game state.