Autoencoder for piece placement, active color, castling availability and en passant target square and halfmove clock for the fifty-move rule, to get a single vector to encode these informations, these encode everything you need except the threefold repetition, for that I was thinking of simply encoding the previous reversible halfmoves in the cross attention (the order shouldn't matter), 50 halfmoves is the most you can have online because of the fifty-move rule or it's a draw. This should fix the problem that the model developed by Deepmind has by being blind to threefold repetition, a better objective that consider the length of the checkmate sequence should make the model more decisive in the face of overwhelming victory and it's knowledge of the threefold repetition should discourage the model to going back and forth on a winning position.
Slight but perhaps significant correction, the 50 move rule only takes effect if no piece has been captured and no pawn moved. Online games can definitely be more than 50 moves.
I think you could get pretty far just by learning embeddings for board positions and making moves from similar board positions during a game. Essentially a retriever-only system.
