Deep player behavior modeling

Video games have become the entertainment industry’s leading branch, with revenues that surpass even TV, cinema or music. This rapid development goes along with equally skyrocketing consumer demands and expectations, not limited to the continuous production of content, the validation against flaws and gameplay bugs and the preservation of real-time online functionalities – in ever-growing systems and applications. While efforts to overcome these issues primarily involve distinct expenses of intensive manual labor, automated and/or artificial intelligence-driven approaches as procedural content generation, dynamic difficulty adjustment or autonomous testing aim at lifting the burden from the developers’ shoulders. For the simulation of artificial behavior, human-likeness or believability is considered to be one of the main quality criteria, yet most industrial as well as academic approaches focus on generally believable behavior for these purposes. This dissertation introduces the concept, architecture, implementation and evaluation of Deep Player Behavior Modeling, which assesses the atomic decision making of particular players and generates individual behavior representations to be implemented in artificial agents. After the examination through multiple field studies in different games and genres, these agents proved to be able to convincingly display individual strategies and preferences, represent in-game proficiency accurately and became indistinguishable from their original human player. Together with an extensive literature review and expert interviews that point out the case for usable AI in video games, this thesis contributes to the fields of game user research, game AI, machine learning and player modeling within both academia and industry and illustrates significant advances in the application fields of dynamic difficulty adjustment, player substitution, automated game testing and serious games.