Lyra: Simulating believable opinionated virtual characters

Abstract

Creating believable simulations of large populations of characters in virtual worlds represents a grand challenge for interactive artificial intelligence, requiring reasoning about social intelligence. In this paper, we focus on one aspect of this challenge: the dynamics of opinion change for virtual characters and its relationship with social affinity. We developed a simulated population of characters that debate politically-charged topics, called Lyra. Characters' knowledge, opinions, and biases spread through this society based on existing cognitive models and social science theories. Our simulation generates outlines of group conversations that portray the system's evolution, and clusters characters into affinity groups based on the outcome of the debates. We conducted a human-subjects study to evaluate these generated conversations and affinity groups for their believability and to inform future iterations of the simulation. We believe successful simulation of opinion change in social dynamics provides a foundation for computational recognition, prediction, and interfacing with humans.