A Game Architecture for Emergent Story-Puzzles in a Persistent World

This paper presents a double-layer game architecture for generating short, linked story-puzzles in a persistent virtual world, using a simplified detective game set in a Victorian mansion modeled on the Cluedo characters as its test-bed. The central aim is to reconcile two competing approaches to automatic plot generation: centralized systems that control plot development tightly but lack truly emergent behavior, and distributed multi-agent systems that produce emergence but struggle with narrative consistency. The proposed solution splits the system into a narrative layer, which uses case-based reasoning (CBR) and a narrative ontology to generate a crime and a recommended initial state of character relationships, and an agent layer, which runs a sense-think-act cycle for autonomous characters whose behaviors reveal clues through dialogue and emotional display.

The narrative layer's generation process follows retrieval and adaptation stages: a randomly generated crime is matched against a case base using ontology-based semantic distance, and the retrieved world state is adapted to the current persistent world using heuristic rules inspired by the Minstrel system's transform-recall-adapt methods. Character relationships are encoded as three-component emotional vectors (attraction/disgust, trust/distrust, love/hate) with values from -10 to 10, categorized as balanced, weak, moderated, or strong. The paper illustrates this with worked tables showing a previous world state, a randomly generated theft crime, a retrieved case, and the resulting adapted state.

The authors conclude that the architecture can produce a new, coherent story-puzzle in each session, with emotional relationships evolving according to player performance and character interaction, which they argue would increase replayability while keeping designers in indirect control. The paper is explicitly a design proposal rather than an empirical study: the authors state that implementation in a modern game engine and evaluation with regular gamers remain future work, so no user-testing data or performance measurements are reported.