Tile Embeddings: A General Representation for Procedural Level Generation via Machine Learning

• Author / Creator

  Jadhav, Mrunal Sunil

• Procedural Level Generation via Machine Learning (PLGML) refers to the application of machine learning techniques to the automated generation of game levels. PLGML researchers have investigated different level generation techniques to generate new game levels matching the style of a training corpus. While the PLGML community has made notable progress in designing impressive level generators, we are still far from achieving the holy grail of generalizability. Generalizability refers to a level generators’ ability to generate a previously unseen, new game level based on the training data used to build the model. A primary reason for this is the limited availability of PLGML datasets and inconsistent level representation practices across different games. Traditionally employed PLGML datasets are hand-annotated by domain experts and fan communities. The process of curating clean datasets is time-consuming. Hence, even though many video games exist, select few have received a disproportionate amount of research attention.

  Towards this goal of generalizability, we propose a representation learning approach for game level design. We introduce tile embeddings, a continuous, unified affordance-rich representation of 2D games. This thesis covers an initial implementation of tile embeddings and their further modification to handle the particular case of skewed tile distribution observed in games like Super Mario Bros.. We then introduce a novel, two-step level generation process that can leverage the flexibility of a discrete representation with the expressivity of continuous tile embeddings. We evaluate our tile embedding representation on its ability to predict affordances for unannotated tiles and to serve as a PLGML representation for annotated games.
  We perform an ablation study for level generation of Super Mario Bros., and further show the ability to apply our approach to level generation for unannotated games. Our outputs cover generative spaces matching the distribution of the original training data, thus demonstrating the potential of tile embeddings for PLGML applications for any tile-based 2D games. The presented thesis attempts to address the core challenges of PLGML around representation and dataset availability. We believe with more work in this direction, our approach has the power to open new horizons for PLGML research.

• Subjects / Keywords

  • Machine Learning
  • Procedural Content Generation via Machine Learning

• Graduation date

  Spring 2023

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-0ygr-kv61

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.