New DeepMind paper

[Madeleine Birchfield]

If I understand it correctly then MuZero uses some tree search where every move it makes correspondents to a move made on the board. However, the interesting idea is that this is not necessarily the only possibility. One could imaging that MuZero makes sometimes an abstract move that doesn't need to correspond with a single move (a single timestep) on the board but an "idea" that, if executed, leads to another internal state that can be evaluated. I can't immediately think of a way how to train this, but this could be very interesting. It might not be a huge improvement in chess, but engineering an "idea" is something I find very interesting.

MuZero's big innovation is using neural networks for board (state) representation and move generation. This works well in more complex settings such as in video games, where it is either very hard to define what a move is, there are too many moves to handcraft the move generation code, or too much data to handcraft the representation, On the other hand, chess is simple enough that one could easily write the board representation and move generation code by hand and have it be fast, and MuZero's algorithm would only yield a slowdown in training and runtime.