I have four now, and even go out of my way to use them a bit.
My favorites from my collection are a little Dietzgen No 1773, which is surprisingly featureful for something that can fit in a pocket, and a Faber-Castell Novo-Biplex 2/83N, which is kind of one of the big guns for slide rules.
In what appears to be the originating paper for UCT ("Bandit based Monte-Carlo Planning", L. Kocsis and C. Szepesvári, 2006),
* The choice of action to be returned is the one "with the highest average observed long-term reward"
* For simplicity, the payout value used in the paper is 1=win and 0=loss, which will result in the agents maximizing their wins. Presumably one could choose other payout values (i.e. points for that player in games that have that concept) to adjust the priority of the agents. The mathematics does not seem to forbid it.
* This paper uses as their first experiment a multi-player game. They state, "...for P-games UCT is modified to a negamax-style: In MIN nodes the negative of estimated action-values is used in the action selection procedures." It is straightforward and more generalizable to games with N > 2 players to simply record values from that player's standpoint in the first place, instead of manipulating it after the fact in this manner.
Excellent. I did want to eventually write something showing how to modify this algorithm to handle games with randomness, and ISMCTS at a glance looks like it does just that. Thanks!
Actually, I think it should be 3/8, since it's supposed to be the opposing player's win count.
These diagrams were supposed to just be an update of the diagrams from the MCTS Wikipedia article, adding the correct backpropagation values, but I see there are more problems than that.
https://en.wikipedia.org/wiki/Curta