Source-linked AI summary
The VGLC: The Video Game Level Corpus
Adam James Summerville, Sam Snodgrass, Michael Mateas, Santiago Ontañón
TL;DR
Machine-learning approaches to procedural level generation need training corpora, but prior work has used varied data sources. The paper presents the VGLC, a 428-level corpus from 12 games in parseable formats with image representations, and aims to support shared research resources.
Problem
Machine-learning methods for procedural video-game level generation require training corpora, while prior work has drawn on synthetic data, annotated images, and video.
Source: INTRODUCTION · p. 1, RELATED WORK · p. 1Method
The paper assembles a corpus of game levels as parseable text files with corresponding images, using Tile, Graph, and Vector annotation formats.
Source: INTRODUCTION · p. 1Results
The VGLC makes available 428 levels from 12 games as parseable text files with corresponding image representations.
Source: INTRODUCTION · p. 1, CONCLUSIONS AND FUTURE WORK · p. 5Takeaways & Limitations
The authors aim for a shared corpus to reduce repeated corpus-building work and invite researchers to contribute games and tools.
Source: CONCLUSIONS AND FUTURE WORK · p. 5Tile annotations lack dedicated representations for each enemy and some special tiles, resulting in information loss.
Source: Annotation Formats · p. 3
Abstract
from arXiv · showhide
Levels are a key component of many different video games, and a large body of work has been produced on how to procedurally generate game levels. Recently, Machine Learning techniques have been applied to video game level generation towards the purpose of automatically generating levels that have the properties of the training corpus. Towards that end we have made available a corpora of video game levels in an easy to parse format ideal for different machine learning and other game AI research purposes.
INTRODUCTION
Procedural level generation has increasingly used machine learning, but these methods require training corpora. The paper assembles a parseable corpus of 428 levels from 12 games, with corresponding level images, for ML PCG and other game-AI research.
INTRODUCTION
- Machine-learning approaches to procedural level generation share a need for a training corpus.The paper situates this need amid prior work using methods including Bayes nets, Markov chains, clustering, matrix factorization, and PCA.Source: INTRODUCTION · p. 1
- The authors assembled a corpus of 428 levels from 12 games, available as parseable text files with corresponding level images, for machine consumption in ML PCG and other game-AI applications.Source: INTRODUCTION · p. 1
RELATED WORK
Prior game-level PCG has drawn on synthetic data, annotated images, and video, while shared corpora are commonplace in other machine-learning areas. The VGLC is available online, and its accompanying table lists the games included at publication.
RELATED WORK
- Level-PCG datasets have included synthetic data, annotated images, and video, with Shaker and Abou-Zleikha (2014) identified as the only known synthetic-data example.The cited work used data generated by other generation systems; Guzdial and Riedl (2015) used Super Mario Bros. long-play videos to learn sprite-group spatial relationships.Source: RELATED WORK · pp. 1–2
- Shared corpora are commonplace in NLP and machine learning, including resources for image classification, handwriting recognition, and speech recognition.The passage names COCA and the Wall Street Journal Corpus as noteworthy NLP examples.Source: RELATED WORK · p. 1
- The VGLC is available online, and Table 1 lists the games included in the corpus as of publication.The paper gives the corpus repository address in its footnote.Source: RELATED WORK · p. 1, Table 1 · p. 2
DATASETS
At publication, the corpus contains levels from 12 games in Tile, Graph, and Vector formats. The formats represent levels as character grids, graphs, or SVG primitives, with Tile annotations acknowledged to omit some information.
DATASETS
- The corpus includes levels from 12 games and uses three annotation formats: Tile, Graph, and Vector.Most games are 2D tile-based side-scrollers; Zelda room-based levels also use Graph, while Doom levels also appear in Vector format.Source: DATASETS · p. 2
Annotation Formats
- Tile levels are two-dimensional w × h character grids, with a JSON legend mapping characters to possible annotation tags.The legend can map multiple images to one character, such as enemies to “E” and solid unbreakable tiles to “X”.Source: Annotation Formats · p. 2
- Graph format represents room-based levels as DOT graphs, chosen for parseability, visualization with Graphviz, and portability.The graph files encode nodes and edges with labels, and each game has a corresponding JSON legend.Source: Annotation Formats · p. 2
- Vector format uses SVG for levels such as Doom, representing walls and doors as line segments and objects as graphical primitives.The annotation for a line or object is determined by its stroke color.Source: Annotation Formats · p. 3
- Tile annotations lack dedicated representations for each enemy and some special tiles, resulting in information loss.The authors propose exploring more lossless representations, including more expressive tile sets and ontologies; original map images are included so users can make their own annotations.Source: Annotation Formats · p. 3
TOOLS
The paper includes processing tools: a tile-based platformer solver and a Doom WAD parser and rasterizer.
TOOLS
- The tile-based platformer solver produces paths through a given level.It takes a JSON legend identifying solid tiles and the game’s dynamics, including its jump arc.Source: TOOLS · p. 3
- The platformer solver uses a JSON legend and game dynamics, such as the shape of a jump arc, as inputs.Its output is a path through the level.Source: TOOLS · p. 3
- The Doom parser and rasterizer convert WAD files into the corpus’s SVG or tile-based levels.Source: TOOLS · p. 3
POTENTIAL USAGE
The VGLC supports several research directions: generating levels, analyzing design patterns, and transferring style across games. Its formats offer distinct ways to represent and process levels.
Corpora Based Procedural Content Generation
- Tile levels support text-generation approaches and can be converted for image-based methods, while graph levels suit graph-based learning and agent traversal studies.The corpus also includes games not yet used for corpus-based generation, leaving them open to exploration.Source: Corpora Based Procedural Content… · pp. 3, 5
Design Analysis
- The corpus adds 34 Super Mario Bros. levels beyond a prior 20-level design-pattern analysis, and includes 48 levels from 2D Zelda games for analysis.The corpus could also support comparisons between three other games and the more extensively studied Super Mario Bros.Source: Design Analysis · p. 5
Style Transfer
- Cross-game style transfer remains unexplored in the authors’ account; proposed directions include restyling one game’s levels as another or generating variants from a sketched level skeleton.Prior work either focused on one game or series, or partitioned work that included multiple games.Source: Style Transfer · p. 5
CONCLUSIONS AND FUTURE WORK
The authors present the VGLC as a shared resource intended to reduce repeated corpus-building work and help researchers focus on more innovative work.
CONCLUSIONS AND FUTURE WORK
- The VGLC is available online, and the authors encourage researchers to contribute games and tools to make it more useful to the community.The shared corpus is intended to reduce repeated effort assembling resources for corpus-based machine learning.Source: CONCLUSIONS AND FUTURE WORK · p. 5