Evaluation of Procedural Content Generators for Two-Dimensional Top-Down Dungeon Levels

Abstract: This research evaluates two-dimensional top-down dungeon generated levels regarding fundamental and micro dungeon design patterns. Additionally, it investigates the meaningfulness of the evaluation results in terms of accessibility to level designers and similar. The research method concentrates on two dungeon-generation techniques – Cellular Automata and Drunkard Walk. Each generated level gets evaluated based on three evaluation stages that build on top of each other: the passability of each tile; categorization of each collection of tiles with the same attributes; and player-centric gameplay data. The results show key differences between Cellular Automata and Drunkard Walk as the risk of using Cellular Automata to generate up to 90% unreachable space, while drunkard walk always has a playable relative space size of 100%. The evaluation also shows results that depend on the requirements of a game or constraints of a level designer. Cellular Automata generates more rooms, while Drunkard Walk provides more decisions per room. In conclusion, the evaluation results show differences between the two algorithms, presented using a vocabulary that is familiar to a level designer.