Section: New Results

Extensions of Upper Confidence Trees

We developed extensions of Upper Confidence Trees to continuous or large domains (states and/or actions) and to domains with high expertise or strong structure[37] , [31] , [38] (incidentally realizing performances on MineSweeper); we recently submitted a proof of a variant of UCT with consistency proof in the continuous domains (both actions and random variables are allowed to be continuous). Another extension is to the difficult setting with no possibility to “undo” a decision or duplicate a state; see [63] . Yet another extension aims at multi-objective optimization [56] .