Generative Personas That Behave and Experience Like Humans

The Go-Blend framework advances automated game testing and procedural content generation by integrating reinforcement learning with human-centric behavioral and emotional modeling. By utilizing the Go-Explore algorithm alongside a dataset of over 100 human playtraces, the framework creates AI agents that replicate both specific performance metrics and affective responses. This dual-objective approach moves beyond traditional procedural personas, which often prioritize mechanical efficiency over the nuanced, idiosyncratic nature of human play.

Through hierarchical clustering of the AGAIN dataset, four distinct player personas—Beginner, Intermediate, Advanced, and Expert—were identified and successfully modeled. The findings demonstrate that incorporating emotional arousal traces as a reward function alongside traditional performance scores allows agents to mirror the specific behavioral and psychological profiles of these human groups. This methodology proves particularly effective for simulating proficient player personas, offering a more comprehensive tool for developers to evaluate game balance and player engagement across diverse skill levels.

While the framework represents a significant step forward in experience-driven AI, it faces technical constraints regarding trajectory efficiency, largely due to the sparse nature of score-based signals in complex game environments. Despite these limitations, the research establishes a viable foundation for future developments in automated playtesting. By reducing reliance on extensive human demonstration data and refining reward functions, this approach provides a scalable solution for creating AI agents that behave and experience games with greater human-like fidelity, ultimately enhancing the quality and depth of interactive digital experiences.