Mastering strategies in a board game of imperfect information for different search techniques

Abstract

To the best authors' knowledge this work is the first to develop a full computer implementation of The Great Turtle Race (GTR), a complex board game characterized by several uncertainties that uses computational techniques to evaluate board positions and select the best move. In the game, a novel combination of popular propagation-based optimization techniques and four playing strategies is implemented. One of the main goals of this study is to determine how to generate opponents that are quick and safe to play against, rather than being necessarily superior. The paper starts by a brief overview of the game and its rules, followed by some analytical results that emerge from its characteristics. It then moves to provide relevant reinforcement learning methods by which Monte Carlo tree search, minimax and alpha-beta pruning were implemented. The validity of the concept is finalized by a series of experiments, in which these algorithms and strategies were successfully verified against each other.