Many enhancements have been proposed for Monte-Carlo Tree Search (MCTS). Some of them have been applied successfully in the context of General Game Playing (GGP). MCTS and its enhancements are usually controlled by multiple parameters that require extensive and time-consuming computation to be tuned in advance. Moreover, in GGP optimal parameter values may vary depending on the considered game. This paper proposes a method to automatically tune search-control parameters on-line for GGP. This method considers the tuning problem as a Combinatorial Multi-Armed Bandit (CMAB). Four strategies designed to deal with CMABs are evaluated for this particular problem. Experiments show that on-line tuning in GGP almost reaches the same performance as off-line tuning. It can be considered as a valid alternative for domains where off-line parameter tuning is costly or infeasible.
|Title of host publication||Computer Games|
|Subtitle of host publication||6th Workshop, CGW 2017, Held in Conjunction with the 26th International Conference on Artificial Intelligence, IJCAI 2017, Melbourne, VIC, Australia, August, 20, 2017, Revised Selected Papers|
|Editors||Tristan Cazenave, Mark H.M. Winands, Abdallah Saffidine|
|Place of Publication||Cham|
|Number of pages||21|
|Publication status||Published - 2018|