AlphaGo Zero stacks 7 board history along with the current board together to form the input to the network. However, is it possible to use an RNN to replace the input of history and achieve similar performance? Are there any practical reasons for the team to choose stacking history states rather than using a RNN?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
I can't think of any reason why using an RNN wouldn't work in theory.
In practice RNNs are slightly harder to train (they can be unstable, and ever more practically you have to deal with multiple network inputs), so they just went with the easiest solution that worked.
In the followup work MuZero an RNN is used to predict future states from a starting state and played moves, which is a harder task and it does work, so it would be possible to do something similar for AlphaZero too.
