[Paper Review] Automatic Generation of Level Maps with the Do What's Possible Representation

Daniel Ashlock, Christoph Salge|arXiv (Cornell University)|May 23, 2019

Video Analysis and Summarization20 references4 citations

TL;DR

This paper proposes a scalable, open-ended approach to procedural dungeon generation using the Do What's Possible (DWP) representation, which combines a deterministic complex string generator (SDA) with a generative possibility filter to iteratively build level maps. The method enables infinite map growth, adapts to spatial constraints, and supports evolutionary optimization of map generators for high-quality, customizable dungeon layouts with two-tier adaptivity to both design patterns and environmental obstacles.

ABSTRACT

Automatic generation of level maps is a popular form of automatic content generation. In this study, a recently developed technique employing the {\\em do what's possible} representation is used to create open-ended level maps. Generation of the map can continue indefinitely, yielding a highly scalable representation. A parameter study is performed to find good parameters for the evolutionary algorithm used to locate high-quality map generators. Variations on the technique are presented, demonstrating its versatility, and an algorithmic variant is given that both improves performance and changes the character of maps located. The ability of the map to adapt to different regions where the map is permitted to occupy space are also tested.

Motivation & Objective

• To develop a scalable, open-ended representation for procedural level map generation that supports arbitrary map size and continuous growth.
• To apply evolutionary algorithms to optimize complex string generators (SDAs) that produce high-quality dungeon layouts.
• To investigate adaptivity in map generation by enabling the system to respond to spatial constraints and existing content.
• To evaluate the versatility and performance of the DWP representation across different map generation scenarios and parameter configurations.
• To explore extensions such as special room insertion and probabilistic placement for practical game design applications.

Proposed method

• The DWP representation uses a self-driving automaton (SDA) as a complex string generator (CSG) to produce deterministic bit sequences that drive map generation decisions.
• The generative possibility filter (GPF) interprets these sequences as commands to place new rooms, rejecting placements that conflict with occupied space.
• Room placement decisions are made iteratively: selecting a parent room, side, orientation, and dimensions, with placement only allowed if space is available.
• The evolutionary algorithm optimizes the SDA's transition rules to maximize fitness based on map quality metrics such as connectivity and room count.
• The system supports incremental map growth, allowing generation to be stopped at any time with a valid, partially generated map.
• Extensions include inserting special rooms via a list and softening the GPF to allow probabilistic choices (e.g., preferring flat roads on even terrain).

Experimental results

Research questions

• RQ1 Can the DWP representation generate arbitrarily large dungeon maps with minimal representation overhead?
• RQ2 How do evolutionary parameters affect the quality and diversity of generated maps?
• RQ3 To what extent can the DWP system adapt to spatial constraints and existing obstacles during map generation?
• RQ4 How does the two-tier adaptivity—evolved CSG patterns and dynamic GPF filtering—improve content generation for existing environments?
• RQ5 Can the DWP framework support practical enhancements like special room insertion and probabilistic placement for terrain-aware design?

Key findings

• The DWP representation successfully generated self-avoiding walks up to 40×40, exceeding prior limits of 12×12 with adaptive representations.
• The method enabled the evolution of an 8-bit Gray code with 256 unique values, demonstrating strong generative power for complex patterns.
• Evolved SDA-based generators produced maps with high connectivity and structural diversity, even without explicit fitness shaping for aesthetics.
• The system achieved scalable map generation, capable of growing indefinitely with minimal data structures and no predefined size limits.
• The two-tier adaptivity allowed the same evolved CSG to generate different map styles when applied to different obstacle patterns, enabling context-sensitive content.
• Modifications such as special room insertion and probabilistic placement enhanced practicality for terrain-aware generation, e.g., favoring flat roads on even terrain.