Two-Step Constructive Approaches for Dungeon Generation

Key insights

6 takeaways

01
A two-step procedural generation framework—decoupling architectural layout from object placement—enables efficient, controllable level design for mobile environments.See it on page 1
02
Testing nine combinations of three layout creators (constraint-based, cellular automata, agent-based) and three furnishers across 9,000 generated levels confirms that layout creators dictate architecture while furnishers control object density.See it on page 2
03
Constraint-based layout creators consistently produce superior architectural quality, characterized by longer paths and fewer isolated wall segments compared to other methods.See it on page 4
04
The combination of constraint-based creators and furnishers creates the highest level of difficulty for artificial agents, as measured by Monte-Carlo tree search simulations.See it on page 6
05
Procedural persona agents modeled on specific playstyles (runner, monster killer, treasure collector) reveal that specific creator-furnisher pairings create inherent biases toward certain gameplay outcomes.See it on page 6
06
The modularity of this two-step approach provides a computationally efficient solution for generating diverse, playable content within the constraints of mobile hardware.See it on page 1

Summary

This research explores a two-step constructive approach to procedural level generation for the rogue-like puzzle game MiniDungeons 2. The primary objective is to decouple the creation of architectural layouts from the placement of game elements, such as monsters, treasures, and traps. By separating these phases, the study evaluates how different combinations of layout creators and object furnishers impact level structure, gameplay difficulty, and player experience. The research aims to provide a systematic understanding of how various constructive algorithms perform when applied to specific game design constraints.

The methodology involves testing three distinct layout creators—constraint-based, cellular automata, and agent-based—paired with three corresponding furnishers. This results in nine unique creator-furnisher combinations, totaling 9,000 generated levels. To assess the quality and playability of these levels, the study employs procedural persona agents—modeled on runner, monster killer, and treasure collector playstyles—using Monte-Carlo tree search simulations. Key metrics include the number of floor tiles, the length of the longest path between the entrance and exit, the number of wall chunks, and the completion rates of the persona agents.

Findings indicate that layout creators significantly influence the physical architecture, while furnishers dictate the distribution and density of game objects. The constraint-based creator consistently produces levels with longer paths and fewer isolated wall segments. Playtesting reveals that the combination of constraint-based creators and furnishers creates the most challenging levels for artificial agents. Furthermore, specific combinations demonstrate clear biases toward certain playstyles, such as increased monster kills or treasure collection. The study concludes that this modular, two-step constructive framework is computationally efficient for mobile environments while offering a versatile, controllable method for generating diverse and playable game content.