Issue with maybe_flip_fen method about chess-alpha-zero

[Akababa]

You're completely right from a chess game perspective, the 180 degree rotation is what you would do when looking at the same game from the other side of the board. However from the game state symmetry perspective, you want to "flip" black into a white-like position, so the neural network doesn't have to learn the same position once for black and once for white. In this case, you have to do the horizontal mirroring to get the kings on the right.

As a beside, it's unclear how much (if any) benefit this provides to the training algorithm, but it can't hurt so why not? ¯_(ツ)_/¯

[simonMoisselin]

Good point!

[simonMoisselin]

Does it mean that the model is making prediction only when this is WHITE to play?

[Akababa]

Yep, this was a while ago but IIRC we actually removed the one-hot color input.

[simonMoisselin]

The more I think about it, and the 180-degree rotation needs to be done correctly, otherwise you will simulate games that does not make any senses? (they cannot corresponds to game in real life)