拡張モンテカルロ木探索によるブラックジャック戦略

Abstract

This paper focuses on card games, especially Blackjack, which are representative of imperfect information games. AlphaZero, which developed AlphaGo, conquered Go, the most difficult perfect information game. It owes much to Monte-Carlo tree search. The idea of Monte-Carlo tree search, one of the best-first search methods, originates from the UCB algorithm for the multi-armed bandit problem. Using the extended Monte-Carlo tree search, the search tree is improved so as to derive the best action of Blackjack, which is one of the imperfect information games, and the algorithm is constructed. One playout performed at each node of the game tree corresponds to a player winning or losing in that round. The purpose is to realize a Blackjack strategy that can compete with the basic strategy that is effective in Blackjack by searching an extended Monte-Carlo tree. In numerical experiments, the proposed method is evaluated by using the payout ratio.