Generalized Experience Management
David Thue · ERA: Education and Research Archive (University of Alberta) · 2015
Computer-based interactive environments present a compelling platform for research in Artificial Intelligence. Using games as its domains, this work has traditionally focused on building AI agents that can play games well (e.g., Checkers, Go, or StarCraft). In more recent years, a parallel line of research has aimed to achieve a different goal: to mimic the abilities of human game designers, extending their reach into the run time of the game. By building an AI agent to gather new information and make decisions as their proxy, designers can ensure that their goals are pursued in a way that adapts to each player automatically, while the game is underway. In this dissertation, I present the Generalized Experience Management (GEM) framework, the first mathematical formalization of modifying the dynamics of an interactive environment during end-user play. Moving beyond traditional, ad hoc methods for creating AI agents that manage player experiences, GEM is grounded in the theory of Markov Decision Processes while still remaining practically applicable in both industry and academia. To evaluate the framework and demonstrate its versatility, I present four adaptive systems as instances thereof: two that I designed and tested through controlled user studies, one that was created independently in a commercial video game, and one that was seminal in the domain of Interactive Drama. Finally, I propose and demonstrate a detailed method for evaluating GEM systems, including a new way to distinguish between the effects of player-specific and player-independent adaptation.