Adaptação automática de histórias para RPGs não-lineares: uma abordagem com inteligência artificial generativa e otimização evolutiva

Abstract

Role-playing games (RPGs) have a consistent presence in both the analog and digital markets, exerting a significant influence on culture in Brazil and worldwide. RPGs are strongly tied to the narrative elements of the adventure constructed throughout game sessions, which are governed by mechanics derived from representations of their universe, plot, characters, and other elements of that adventure. This means that the intricate structure of the game must be tangible to game designers, serve the narrative framework, and remain viable within the constraints imposed by an RPG system. As with other artistic productions, games enable the revisiting and reinterpretation of classic works playfully, whether through reinterpretation of those works or as references to material familiar to gaming audiences. However, adapting existing narratives to the structure of RPGs is a challenging and time-consuming task. This work aims to propose a new automatic method for: (1) generating non-linear narrative structures from a generative stochastic graph grammar, which defines reusable unfolding patterns for RPG adventures; (2) weighting the generative graph grammar via optimization using Genetic Algorithms; (3) extracting the narrative structure from public domain texts; and (4) adapting this structure to interactive RPG adventure models using Generative Artificial Intelligence techniques. This narrative structure is stored in a directed graph, which enables the exploration of the aesthetic qualities of the adventure, such as its length, number of choices, and the number of distinct endings. The proposed method uses Graph Grammars to generate the game structure, ensuring narrative cohesion through the chaining of causes and consequences based on the player’s choices. The process is optimized using Genetic Algorithms to adapt the grammar to the creation of games based on the textual content of literary works extracted from a base text and adapted for the adventure through generative neural networks. The models obtained positive results in the evaluations of the fitness functions applied in the genetic algorithm, in addition to a reasonable distribution in the qualitative evaluation established by a set of LLM when evaluating aesthetic qualities.