The Undergraduate Games Corpus: A Dataset for Machine Perception of Interactive Media

Barrett R. Anderson, Adam M. SmithAAAI 2021 · DOI

Introduction

While machine perception has made significant progress in processing static media such as images and text, the domain of interactivity remains a challenging frontier. Most AI research in gaming focuses on "playing to win" through Reinforcement Learning, often utilizing simplified clones of classic titles like Atari or Super Mario Bros. However, these "lab-grown" environments frequently lack the thematic depth and design polish of games intended for human audiences. To bridge this gap, researchers at UC Santa Cruz have introduced the Undergraduate Games Corpus, a diverse dataset of 755 student-made games designed to help machines perceive the nuances of choice, narrative, and player experience.

The Undergraduate Games Corpus

The primary contribution of this work is a legally redistributable dataset of 755 individual games created by students in introductory game design courses at UC Santa Cruz between 2019 and 2020. Unlike commercial datasets, this corpus includes full source code, assets, and metadata, with over half (54%) provided under Creative Commons licenses.

Distribution of game file sizes across different engines in the corpus.

The corpus spans several popular accessible game engines:

  • Twine (61%): Text-based interactive fiction.
  • Construct 3 (25%): 2D graphical games.
  • Bitsy (11%): Low-fidelity narrative games.
  • Godot (3%): 2D and 3D games.

These games are "authentic" human expressions, reflecting real-world themes from the era they were created, such as the COVID-19 pandemic and social unrest. This authenticity challenges researchers to build perception systems that handle "interestingly messy" real-world artifacts rather than sterile simulations.

Methodology: Graph-Based Narrative Analysis

To demonstrate the utility of the corpus, the authors developed a metadata tag prediction task. The objective was to determine if a machine could analyze a game's source code to predict its genre or emotional themes (e.g., "fantasy," "horror," or "anxiety").

The researchers focused on Twine games, representing them as directed graphs where story passages are nodes and player choices are edges. The pipeline involves several key steps:

Overview of the tag prediction pipeline for Twine games, from source code to story-level vector.
  1. Text Encoding: Individual story passages are converted into 512-dimensional vectors using a pre-trained Universal Sentence Encoder (USE).
  2. Graph Modeling: Graph Convolutional Networks (GCNs) are employed to capture the interactivity of the narrative.
  3. Message Passing: The GCN allows information to flow between linked passages, mimicking "choice poetics"—the concept that the meaning of a scene is conditioned by the choices leading to it.
  4. Classification: The aggregated story-level representation is fed into a multi-task classifier to predict target tags.

Experimental Results

The validation experiments confirmed that structural connectivity is a primary carrier of meaning in interactive media. The results showed that:

  • Structure Enhances Perception: The GCN model, which considers the links between passages, consistently outperformed a standard Multi-Layer Perceptron (MLP) that analyzed text in isolation.
  • Depth of Reasoning: Increasing the number of "message passing" rounds improved performance. The Graph16 model achieved an AUC-ROC of 0.688, compared to 0.645 for the text-only MLP and a baseline of 0.469.

Conclusion and Future Directions

The Undergraduate Games Corpus provides a rare open resource for studying how the structure of choices conveys meaning. While the current scale is modest compared to datasets like ImageNet, it offers a foundation for building better search engines for interactive moments and exploring "General Game Playing" (GGP) systems. Future work will focus on addressing label noise from student-provided tags and developing standardized APIs to allow AI agents to automatically "play" the graphical games within the corpus.

Summary updated

BibTeX
@inproceedings{anderson2021undergraduate,
  author = {Barrett R. Anderson and Adam M. Smith},
  title = {{The Undergraduate Games Corpus: A Dataset for Machine Perception of Interactive Media}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i1.16071},
}

Loading related content...

Figure