TL;DR
This paper introduces a Hexagram Tangram puzzle artifact, including code for generating puzzles and 3D models for physical realization. The goal is to move beyond simple puzzle-solving by creating a richer design space with AI-driven constraints and a curriculum to explore it, making puzzles more engaging and educational.
What Problem Does This Paper Address?
The authors were motivated by a commercially available Hexagram Tangram puzzle that, despite having many solutions, lacked features to encourage exploration or deeper engagement. They found the original design disappointing because the colors were purely aesthetic and didn't contribute to the puzzle's challenge. This highlights a broader challenge in procedural content generation for games: creating content that is not just solvable but also interesting and educational. The paper addresses this by proposing a method to design puzzles with richer constraint spaces, making them more amenable to AI-driven exploration and curriculum development.
Methods
The core of the paper's contribution lies in the artifact and the process used to create it. The puzzle design itself was an iterative process involving human design interleaved with Exhaustive Procedural Content Generation (EPCG). EPCG was used to explore the design space, starting from piece selection and culminating in a curriculum of puzzle instances. A key innovation is the introduction of physical constraints on the puzzle board and pieces. Each triangle on the board and pieces has either a "bump" or a "hole." A bump must align with a hole, while holes can align with either bumps or other holes. This physical constraint mechanism is designed to be implementable both digitally and physically. The authors also developed a curriculum of puzzles based on these constraints, organized into sections exploring different types of piece interactions (e.g., touching edges, not touching).
Datasets
The paper does not explicitly mention training datasets. However, it introduces a newly generated dataset of puzzle instances. This dataset consists of all 2,508 possible puzzle instances derived from the Hexagram Tangram design, with symmetric puzzles removed. These puzzles are then analyzed and divided into sets containing solutions and initial states for a curriculum. The data for the puzzles is released under a non-commercial attribution license.
Key Insights and Ideas
- Beyond Solvability: The Importance of Engagement: The paper emphasizes that simply having many solutions doesn't make a puzzle engaging. Introducing meaningful constraints and a structured exploration path (curriculum) is crucial for deeper interaction.
- Physical Constraints for Digital and Physical Puzzles: The "bump and hole" mechanism for piece placement is a clever way to create constraints that are intuitive for physical puzzles and easily implementable in digital environments.
- AI-Driven Curriculum Design: Using EPCG not just to generate puzzles but to create a structured curriculum that guides users through different constraint types and difficulty levels is a novel approach to educational game design.
- Entropy as a Measure of Difficulty: The authors mention using puzzle entropy (derived from Chen, White, and Sturtevant 2023) to name and categorize puzzles, suggesting a quantitative approach to understanding puzzle difficulty.
Key Results
The paper describes the creation of a Hexagram Tangram puzzle artifact that includes:
- Code for generating puzzles and their solutions using EPCG.
- 3D models of the puzzle pieces and board in STL format for physical printing.
- All possible legal puzzle instances in SVG format (2,508 unique puzzles after removing symmetric ones).
- A curriculum of puzzles designed to explore constraints, with 12 puzzles per chapter.
The paper highlights that there are only nine solutions when trapezoid pieces are excluded, motivating their inclusion to increase puzzle space. The curriculum is structured with an introductory chapter and subsequent chapters focusing on color-based constraints.
Limitations and Future Work
The authors acknowledge that the interactive prototypes developed during the project are not included in the current artifact, as the physical prototype was deemed sufficient for testing and the prototypes were not maintained as the design evolved. They also note that the current 3D print files are too large to include but are working on reducing file sizes and may provide higher-resolution files in the future. Future work will involve continuing to refine the design and potentially updating the code.
Summary updated