Improvement of State’s Value Estimation for Monte Carlo Tree Search

Monte Carlo Tree Search (MCTS) is an effective search algorithm used in games e.g. the game of Go. MCTS performs several simulations and decides which move is the best based on the mean reward of simulations. However, to find the best move, the maximum expected rewards are needs to be estimated at the successive states, thus it is not obvious whether evaluating moves by the mean rewards are appropriate. In this paper, we propose a MCTS algorithm which uses Simplified Weighted Estimator, an estimator of the maximum of the expected value, instead of the estimation by the mean. We examine its effectiveness in terms of evaluation accuracy, deviations, whether it can choose the best move. The proposed algorithm was better than existing MCTS ones in some settings