Theses and Dissertations - UTB/UTPA
Generalizing Agent Plans and Behaviors with Automated Staged Observation in The Real-Time Strategy Game Starcraft
Date of Award
Master of Science (MS)
Dr. Emmett Tomai
Dr. Richard Fowler
Dr. Robert Schweller
In this thesis we investigate the processes involved in learning to play a game. It was inspired by two observations about how human players learn to play. First, learning the domain is intertwined with goal pursuit. Second, games are designed to ramp up in complexity, walking players through a gradual cycle of acquiring, refining, and generalizing knowledge about the domain. This approach does not rely on traces of expert play. We created an integrated planning, learning and execution system that uses StarCraft as its domain. The planning module creates command/event groupings based on the data received. Observations of unit behavior are collected during execution and returned to the learning module which tests the generalization hypothesizes. The planner uses those test results to generate events that will pursue the goal and facilitate learning the domain. We demonstrate that this approach can efficiently learn the subtle traits of commands through multiple scenarios.
University of Texas-Pan American
Copyright 2012 Zackary Gill. All Rights Reserved.