HQ-DoG

Abstract

This thesis presents HQ-DoG, an algorithm that uses hierarchical Q-learning to learn a policy that controls a team of ten soldiers in a video game environment to compete in the gametype known as domination. Domination is a game where two opposing teams compete for possession of three different positions on a map. The team that holds the most positions for the longest amount of time wins the game. HQ-DoG uses three Q-tables that represent a captain and two subordinate lieutenants connected in a hierarchical structure. Together the captain and lieutenants learn where to deploy soldiers across the environment in different circumstances and how many soldiers should be sent to different targets. HQ-DoG is tested against a series of strategies that exhibit different levels of sophistication in playing domination, and the results show it is able to learn a competitive strategy given enough time.