UCT

Now that completely naive UCT can select objectively good moves in go 囲碁 albeit slowly, we can calmly return to the task of finding heuristics to place prior probabilities on the UCT search tree. We no longer have the urgency that the heuristics be very good: the probabilistic nature of UCT can robustly handle a little bit of error in the heuristic.

We may derive heuristics automatically using machine learning.

How well does UCT do on the combinatorial game theory endgames?

I would like to see UCT applied to chess, with a limited depth search as the node prior generator. It should do well in "strategic" positions.

How well does UCT do on chess endgames, where computers these days play especially poorly?