UC Santa Cruz

Procedural Content Generation has been a part of videogames for most of their existence. Early use allowed games to present more content than would have been possible to store, but modern use has blossomed in to a field of research with a wide variety of goals. One such goal has been the generation of content matching a particular design style, and most of the research has been focused on classic AI methods -- such as search and constraint satisfaction -- to produce content optimizing for some set of constraints and metrics that a researcher sets forth. Recently, there has been a focus on using machine learning to learn the design latent within the content itself, so that a researcher does not need to encode that design knowledge them self; however, this work still includes a large amount of human annotation and guidance.

This dissertation focuses on two systems, \textit{Mappy} and \textit{Kamek}, which can work together to learn from observation of a game, so as to be able to generate content that has similar properties as that of the original content. \textit{Mappy} is an AI system that uses a wide range of techniques to emulate the process that a human undertakes when they play, observe, and learn from a game. \textit{Kamek} is a machine learning system that learns from game levels and generates new levels of high quality. This dissertation also covers a suite of methodologies for the analysis and presentation of procedurally generated content -- with a specific focus on machine learned generators.