Preliminary Poetics of Procedural Generation in Games

This paper proposes a poetics of procedural generation in games, arguing that existing taxonomies of procedural content generation (PCG) describe how generators are built but not why they produce particular aesthetic effects. Its central move is to introduce aesthetic terms, most notably the Apollonian (rule-governed coherence) and the Dionysian (chaotic, noise-driven variation), and to treat them as non-exclusive qualities. The most sophisticated generators can exhibit both at once, so the poetics is framed as a multidimensional space rather than a single continuum.

The argument develops by examining how randomness, such as Perlin noise, dice rolls, and shuffles, is balanced against structural order, as in cellular automata like Conway's Life. The paper holds that aesthetic effects depend on the tension among repetition, the gestalt, and individual artifacts, and that the aesthetic locus of a generator, whether in its structure, surface, or overall impression, shapes how players perceive it. Illustrative cases include the planetary biomes of No Man's Sky, the rare structures of Minecraft, and the generative novels produced through NaNoGenMo. Its conclusions are conceptual rather than quantitative: predictable distributions and deliberate "tells" allow players to infer hidden generative rules, and non-linear weightings tend to yield more interesting output than uniform probabilities.

A further claim is that aesthetic qualities such as multiplicity, style, and cohesion cross-cut the Apollonian/Dionysian distinction. The paper cautions that a generator's perceived variety can be far narrower than its theoretical output count, and that meaningful variation depends on tying results to visual priorities and surrounding game systems. It also argues that no generator is truly content-free, since all rely on designer-authored data and metadata whose selection strongly shapes outcomes, as seen in Markov chains, Deep Dream, Spelunky, and Dwarf Fortress.

The work draws on an extensive literature spanning PCG research, generative art theory, and interactive fiction, and it situates its framework against prior approaches including expressive range analysis, MDA-based models, and Galanter's effective complexity. Presented as a preliminary study, it establishes a conceptual vocabulary for discussing the aesthetics of generation rather than offering empirical validation of its claims.