Monte-Carlo Tree Reductions for Stochastic Games

Nicolas Jouandeau (1) , Tristan Cazenave (2)

1 LIASD - Laboratoire d'Informatique Avancée de Saint-Denis
2 LAMSADE - Laboratoire d'analyse et modélisation de systèmes pour l'aide à la décision

Monte-Carlo Tree Search (MCTS) is a powerful paradigm for perfect information games. When considering stochastic games, the tree model that represents the game has to take chance and a huge branching factor into account. As effectiveness of MCTS may decrease in such a setting, tree reductions may be useful. Chance-nodes are a way to deal with random events. Move-groups are another way to deal efficiently with a large branching factor by regrouping nodes. Group-nodes are regrouping only reveal moves and enable a choice between reveal moves and classical moves. We present various policies to use such reductions for the stochastic game Chinese Dark Chess. Move-groups, chance-nodes and group-nodes are compared.

Domaines: Intelligence artificielle [cs.AI]

Nicolas Jouandeau, Tristan Cazenave. Monte-Carlo Tree Reductions for Stochastic Games. 19th annual Conference on Technologies and Applications of Artificial Intelligence, Nov 2014, Taipei, Taiwan. ⟨10.1007/978-3-319-13987-6_22⟩. ⟨hal-02317159⟩