search
Table of Contents
<Work In Progress — will update page as I consolidate my findings>
Overview
Atlas is a content-generator for 2D game levels. Given an example level, Atlas reverse-engineers its characteristic style and generates additional levels in that style. Atlas builds upon existing research at the intersection of Computational Creativity and Procedural Content Generation. The defining characteristic of Atlas is that it does not require massive quantities of labeled training data, processing power, or time. Atlas learns style from only a single unlabeled example and is efficient enough to run live on a mobile device in real-time.
Atlas is being developed by Martin Mumford, who you can contact for more information: (arcanemx at gmail.com)
Architecture
The formal proposal for Atlas can be found <here>:
At a glance, Atlas operates in two phases: Analysis and Generation.
Analysis Phase
- Shape Density
- Shape Skeleton
- Shape Flow Network
- Hierarchical Decomposition
- L-System Grammar Induction
- Pattern Recognition
- Global Features
Generation Phase
- Stage I: Sketch
- Stage II: Boundaries
- Stage III: Detailing
- Stage IV: Experimentation
- Global Preferences
- Computational Canvas Representation
- The Decide, Act, Evaluate Loop
- Satisfaction
- Results
Validation
Source
Atlas is written in Swift for iOS, and its source can be found <here>
Roadmap
A high-level overview of the development timeline with critical milestones for completion:
Developer Log
A daily update log on research progress
