Optimization-based Approaches for Structured Content Creation in Game Design

Abstract

Video games are a popular modern form of entertainment, and can serve as effective platforms for research and education at the same time. Content creation is a crucial part of game design and development, and has been receiving academic and industrial attention. This thesis focuses on my research on procedural content generation (PCG), the use of AI techniques in automated content creation, primarily in the context of games. The thesis discusses around the central idea of PCG problems being multi-objective where two key aspects, quality and variety, are important but conflicted, and well-structured models being able to help express and optimize for better quality, variety or a balance between them. It demonstrates how my Ph.D. works address relevant parts of the idea, including optimizing the simulation parameters in creating a scientific game about a type of useful chemical for better accuracy (quality), and providing a domain specific language for PCG with grammatical structures, as a tool for authoring generators for variety within structured models and compile time integration with a general purpose programming language. In addition, my final Ph.D. project is an overarching work on addressing all parts of my central idea, which creates a digital card game supporting diverse PCG cards with grammatical structures, then applies a deep learning model based on recursive neural networks and conditional variational encoders, along with simulation-based machine learning techniques within a feedback loop to improve the game balance of the card game, providing insights on combining the considerations on both quality and variety via carefully designed structures and optimization processes, within an interesting and challenging problem domain.