Multi-agent Naïve Utility Calculus: Intent Recognition in the Stag-Hunt Game

Lux Miranda, Özlem Özmen Garibay

Abstract: The human ability to utilize social and behavioral cues to infer each other’s intents, infer motivations, and predict future actions is a central process to human social life. This ability represents a facet of human cognition that artificial intelligence has yet to fully mimic and master. Artificial agents with greater social intelligence have wide-ranging applications from enabling the collaboration of human-AI teams to more accurately modelling human behavior in complex systems. Here, we show that the Naïve Utility Calculus generative model is capable of competing with leading models in intent recognition and action prediction when observing stag-hunt, a simple multiplayer game where agents must infer each other’s intentions to maximize rewards. Moreover, we show the model is the first with the capacity to out-compete human observers in intent recognition after the first round of observation.