Date of Award

3-14-2014

Document Type

Thesis

Degree Name

Master of Science

Department

Department of Electrical and Computer Engineering

First Advisor

Brett J. Borghetti, PhD.

Abstract

Dominoes is a partially observable extensive form game with probability. The rules are simple; however, complexity and uncertainty of this game make it difficult to apply standard game theoretic methods to solve. This thesis applies strategy prediction opponent modeling to work with game theoretic search algorithms in the game of two player dominoes. This research also applies methods to compute the upper bound potential that predicting a strategy can provide towards specific strategy types. Furthermore, the actual values are computed according to the accuracy of a trained classifier. Empirical results show that there is a potential value gain over a Nash equilibrium player in score for fully and partially observable environments for specific strategy types. The actual value gained is positive for a fully observable environment for score and total wins and ties. Actual value gained over the Nash equilibrium player from the opponent model only exist for score, while the opponent modeler demonstrates a higher potential to win and/or tie in comparison to a pure game theoretic agent.

AFIT Designator

AFIT-ENG-14-M-57

DTIC Accession Number

ADA610929

Share

COinS