A Formal Architecture of Shared Mental Models for Computational Improvisational Agents

This paper proposes a formal approach of constructing shared mental models between computational improvisational agents (improv agents) and human interactors based on our socio-cognitive studies of human improvisers. Creating shared mental models helps improv agents co-create stories with each other and interactors in real-time interactive narrative experiences. The approach described here allows flexible modeling of non-Boolean (i.e. fuzzy) knowledge about scene and background concepts through the use of fuzzy rules and confidence factors in order to allow reasoning under uncertainty. It also allows improv agents to infer new knowledge about a scene from existing knowledge, recognize when new knowledge may be divergent from the other actor's mental model, and attempt to resolve this divergence to reach cognitive consensus despite the absence of explicit goals in the story environment.