PUCT and Mate Evaluation for Monte-Carlo Alpha-Beta

[dangi12012]

This is actually very interesting because this is an improvement of the canonical mcts that has general applicability.
Alphabeta has a list of 15 domain specific improvements but we now have 2-3 general ones for MCTS already.

I would add here that the authors used NEGAMAX which is not optimal. Even in the non sorted random case - alphabeta will need less nodes to solve the outcome of the tree.

The whole idea of the paper is this: (If there is more to it - let me know!)
Once a subtree is fully expanded the whole subtree could be collapsed into a single leafnode with the outcome of the subtree. (The subtree is actually still there but it wont be visited)