Available at: https://digitalcommons.calpoly.edu/theses/2673
Date of Award
6-2023
Degree Name
MS in Mathematics
Department/Program
Mathematics
College
College of Science and Mathematics
Advisor
Jeffrey Liese
Advisor Department
Mathematics
Advisor College
College of Science and Mathematics
Abstract
Games can be used to represent a wide variety of real world problems, giving rise to many applications of game theory. Various computational methods have been proposed for identifying game strategies, including optimized tree search algorithms, game-specific heuristics, and artificial intelligence. In the last decade, systems like AlphaGo and AlphaZero have significantly exceeded the performance of the best human players in Chess, Go, and other games. The most effective game engines to date employ convolutional neural networks (CNNs) to evaluate game boards, extract features, and predict the optimal next move. These engines are trained on billions of simulated games, wherein the strategies become increasingly refined as more games are played. To explore the trade-offs inherent in developing CNNs, we will train them to play the game Connect-4, which is relatively small and has known optimal strategies. In this setting, we experiment with a variety of neural structures and other related factors with only a few hundred thousand simulated games. The results will allow us to compare how different aspects of the neural network impact performance. We propose a framework for this training process which generalizes to any two-player board games meeting some necessary criteria.
Cashier’s Receipt for Master's Thesis Processing Fee