Choisir la langue :

Improvements To Monte Carlo Tree Search

onte Carlo Tree Search combines random games with state dependent statistics. We will briefly recall UCT the standard Monte Carlo Tree Search algorithm and we will show various improvements that make use of different kinds of statistics to improve random games. Both game specific and general heuristics enable to improve much the level of play compared to standard UCT. The GRAVE heuristic helps improving the choices of moves for already visited states, the PPAF heuristic improves the playout policy online and the SHOT algorithm makes use of bandits that minimize the simple regret instead of the cumulative regret as UCT does.

Dates: 
Friday, March 31, 2017 - 11:00
Location: 
Inria, room A00
Speaker(s): 
Tristan Cazenave
Affiliation(s): 
LAMSADE, Université Paris-Dauphine