Teaching Game AI as an Undergraduate Course in Computational Media

Adam M. Smith, Daniel G. ShapiroAAAI 2020 · DOI

Overview

This paper details a five-year experimental effort at UC Santa Cruz to reboot the "Game AI" curriculum for a new generation of "design-literate engineers." By shifting the focus from traditional "smart NPCs" to a broader set of roles—such as AI as a design assistant, storyteller, or choreographer—the authors successfully taught complex engineering concepts to a mixed audience of computer science and game design students.

The Disciplinary Silo Problem

Traditional AI education often suffers from a "disciplinary silo" problem. Computer science students are typically well-versed in algorithms but often lack the design intuition to apply them to human experiences. Conversely, game design students may have creative vision but struggle with the technical complexity of modern AI systems.

In the gaming industry, this tension results in AI being narrowly applied, mostly to pathfinding and combat. The authors argue that as AI matures, there is a pressing need for a curriculum that treats computation as a medium for creative expression, preparing students to invent entirely new roles for AI in interactive media.

Curriculum Design and Methods

The authors designed an 11-week undergraduate course structured around four central themes:

  1. Core AI Metaphors: Teaching algorithms as tools with specific "affordances" for games, including forward search (A*), Monte Carlo Tree Search (MCTS), Behavior Trees, and Genetic Algorithms.
  2. AI as "X": Expanding the student's vocabulary of what AI can be. Instead of just an "Adversary," AI is presented as an Actor (readable behavior), an Author (narrative assembly), a Design Assistant (providing feedback to humans), or a Drama Manager (shaping the player's experience).
  3. Unpacking Existing Games: Using industry "postmortems" and guest lectures from AAA developers to ground theoretical lectures in real-world practice.
  4. Creative Application: A major 5-week final project where students demonstrate a novel role for AI in a prototype system.

The programming assignments were designed to be "high-level," using Python to allow students to quickly prototype complex behaviors—like a Minecraft-inspired crafting planner or an evolutionary level generator—rather than getting bogged down in low-level implementation details.

Key Insights from Student Projects

Over five years, the authors analyzed a dataset of 102 student-initiated final projects. A key finding was the "native appeal" of specific technologies. Genetic algorithms and Behavior Trees were the most popular choices, likely because they are highly visual and produce immediate, often surprising, results.

Figure 3: Distribution of AI technologies used across 102 student projects, showing a strong preference for Genetic Algorithms and Behavior Trees.

Other significant insights included:

  • AI as a Design Tool: One of the most successful shifts was teaching AI for "design-time" rather than "run-time." Students built tools that used pathfinding to tell a designer, in real-time, whether a level they were currently drawing was actually beatable.
  • Breaking the A* Association: By tasking students to build a "Minecraft Crafting Planner," the course taught them to see graph search as a way to navigate logical spaces (e.g., crafting recipes) rather than just physical ones.

Results and Role Invention

Students successfully moved beyond the "Adversary" trope. Projects included AI that generates tutorials for novice players, AI that "playtests" puzzles to ensure they are solvable, and AI that evolves creatures based on environmental factors.

Figure 4: Creative purposes and roles served by student projects, highlighting a shift toward tutoring, level design, and smart adversaries.

The course proved that complex engineering (like MCTS) could be taught to a mixed audience by adopting an "applications perspective" rather than a purely theoretical one.

Limitations and Future Work

Despite the success, the authors noted a "Machine Learning Gap." During the five years reported, neural networks were underrepresented in student projects due to a higher barrier to entry compared to symbolic AI. Additionally, finding technical details on commercial games remains difficult due to industry secrecy. Future iterations of the course aim to integrate more machine learning, particularly in the context of procedural content generation and narrative systems.

Summary updated

BibTeX
@inproceedings{smith2020teaching,
  author = {Adam M. Smith and Daniel G. Shapiro},
  title = {{Teaching Game AI as an Undergraduate Course in Computational Media}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i09.7064},
}

Loading related content...

Figure