Application of Abstraction Hierarchies to Incorporate Human Knowledge for Machine Learning

Abstract

We evaluate the use of the Abstraction Hierarchy (AH) modeling technique from Work Domain Analysis (WDA) to generate policies suitable for use by a reinforcement learning agent. The paper extends previous work utilizing the classic arcade game, Pac-Man to the more complex world of Mario Bros. This paper presents the investigation of AHs for Mario Bros. to generate options and constraints for a reinforcement learning agent. For both games, AHs were generated for high and low performers using qualitative data. The differences between high and low performers’ AHs were translated into sets of options and constraints as suitable information into machine learning algorithms. A general AH form, which was built to enable researchers to generate an AH for simple game domains using the results of the previous study and this study, supports that an AH for another simple game might be generated using only quantitative data collected from game play and the observations of a researcher.