After having read through the Deep Mind's Alpha Zero paper, I understood that we are building up a tree and adding a new node to the tree every time we see a new node. For a game like GO (or even CHESS) with such huge state spaces, and such a large training time, we should definitely exceed any practical memory size for such a tree. But as I know, the algorithms have been practically implemented. Where is the gap in my understanding?
Asked
Active
Viewed 122 times
0
-
You assume that the whole state space is held in memory, and that is not the case. – Dr. Snoopy Jul 29 '19 at 06:30
-
Oh okay. Just read a blog on the subject. Realized that the MCTS tree is re-initialized for every game – Vaibhav Gupta Aug 31 '19 at 11:26