Search Policies in Multi-Player Games 1

Abstract

In this article we investigate how three multi-player search policies, namely maxⁿ, paranoid, and Best-Reply Search, can be embedded in the MCTS framework. The performance of these search policies is tested in four different deterministic multi-player games with perfect information by running self-play experiments. We show that MCTS with the maxⁿ search policy overall performs best.

Furthermore, we introduce a multi-player variant of the MCTS-Solver. We propose three update rules for solving nodes in a multi-player MCTS tree. The experimental results show that the multi-player variant of the MCTS-Solver is a genuine improvement for MCTS in multi-player games.