Machine learning algorithms for automated analysis of player behavior in next-generation video games

点击翻译按钮获取中文摘要

In collaboration with the video game manufacturer Ubisoft (Montreal) we will develop machine learning algorithms which improve the state-of-the-art in immersive video games. Our goal is to build an informative, predictive player state that incorporates past gaming behavior, tastes, demographic information, and current game play trajectory. This state will be used to personalize and otherwise improve game aspects such as dynamic difficulty scaling, non-player character artificial intelligence (AI) and multiplayer matchmaking. We place particular focus on the collection and analysis of data from next-generation controllers such the Natal 3D camera system from Microsoft. We propose to build upon recent discoveries in training so-called deep architectures, which can take advantage of large amounts of data (even without manual labeling). These architectures are able to learn powerful representations of the main explanatory factors of variation present in the data, and once these are pulled apart, it is easier to learn to make the predictions useful for direct in-game feedback. In concert with Ubisoft employees, we will integrate our new algorithms into Ubisoft's own development framework so as to provide game designers with powerful new tools for constructing the next generation of adaptive games.