The anatomy of gameplay : general affect prediction across games and genres

Abstract

General user modelling has long been the holy grail of many domains within artificial intelligence, including affective computing and games research. General affective user models in these fields could advance the understanding of human emotion and serve as a basis for numerous practical applications. However, as affective computing traditionally focuses on passive media, the transferability of affective models to dissimilar tasks is rarely observed. While psychophysiological signals used in affective computing can provide task-agnostic information, they do not scale well outside of a lab environment. In contrast, games research focuses on an inherently interactive medium that provides rich contextual information about the game-state and player behaviour. Unfortunately, most of the work in games user research focuses on game-specific applications, and the handful of studies on general affect modelling in the past have been limited by small, ad-hoc testbeds. Numerous research questions in the field still remain open, and it is clear that a more methodical approach is needed that examines more comprehensive datasets and different levels of generality. To which degree can we predict player affect in unseen games? Is it possible to find transferable characteristics between dissimilar games and genres? This thesis asks these questions and attempts to answer them through a series of experiments on predicting player arousal. Along the way, it presents several contributions to affective computing and game research, including a robust pipeline for first-person affect annotation for interactive elicitors; a comprehensive online platform for data collection; the most extensive dataset to this date that contains affective labels for multiple games and genres; and an ordinal modelling pipeline using dynamic windowing, accounting for player memory. Studies in this thesis highlight the strengths of heuristic general features and present a thorough examination of the modelling of temporal dynamics of arousal through preference learning. The resulting models can transfer between games and genres with a high level of accuracy and the conclusions of this thesis point towards the robustness of time-related game-agnostic features in the modelling of games.