Game theoretic and machine learning techniques for balancing games

Abstract

Game balance is the problem of determining the fairness of actions or sets of actions in competitive, multiplayer games. This problem primarily arises in the context of designing board and video games. Traditionally, balance has been achieved through large amounts of play-testing and trial-and-error on the part of the designers. In this thesis, it is our intent to lay down the beginnings of a framework for a formal and analytical solution to this problem, combining techniques from game theory and machine learning. We first develop a set of game-theoretic definitions for different forms of balance, and then introduce the concept of a strategic abstraction. We show how machine classification techniques can be used to identify high-level player strategy in games, using the two principal methods of sequence alignment and Naive Bayes classification. Bioinformatics sequence alignment, when combined with a 3-nearest neighbor classification approach, can, with only 3 exemplars of each strategy, correctly identify the strategy used in 55\% of cases using all data, and 77\% of cases on data that experts indicated actually had a strategic class. Naive Bayes classification achieves similar results, with 65\% accuracy on all data and 75\% accuracy on data rated to have an actual class. We then show how these game theoretic and machine learning techniques can be combined to automatically build matrices that can be used to analyze game balance properties.