Comparison of algorithmic solvers for the board game Gobblet

Andreja Kropivšek (2014) Comparison of algorithmic solvers for the board game Gobblet. EngD thesis.

Preview

PDF Download (946Kb)

Abstract

Gobblet is a compelling board game that combines elements of classical tic-tac-toe game with a memory game. It exists in two different sizes of boards: 3x3 and 4x4. The goal of this work was to implement a computer version of the game and to research possible solvers that would implement AI players. We picked two most commonly used solutions for these kinds of problems in artificial intelligence: algorithm minimax and Monte Carlo tree search method, also known as MCTS. We examined the minimax algorithm and its optimisation with the alpha-beta pruning. For an estimation of a given state we developed our own heuristic evaluation function. We used it to develop our first AI player called the Minimax bot. We also examined the pathology in the minimax algorithm that we faced during our work. As an opponent MCTS bot was presented. All the steps of MCTS were introduced and we explained the method used for tuning the best values for its parameters. For both board sizes, the outcomes of play-outs between MCTS and the minimax bot are presented. As an interesting comparison we also present results of play-outs of our bots with an existing implementation.

Actions (login required)