💭 At this point my plan for handling arbitrary board games is

• represent the game state in a heterogenous graph. A graph that has different node and edge types..
• Use stochastic muzero with a graph transformer model with type aware attention
• Make the set of legal moves part of the game state rather than being a logical mask applied onto a fixed size policy vector. You just can't put complex board games' choice spaces into a fixed size vector.