Satisficing Models of Bayesian Theory of Mind for Explaining Behavior of Differently Uncertain Agents: Socially Interactive Agents Track

The Bayesian Theory of Mind (ToM) framework has become a common approach to model reasoning about other agents’ desires and beliefs based on their actions. Such models can get very complex when being used to explain the behavior of agents with different uncertainties, giving rise to the question if simpler models can also be satisficing, i.e. sufficing and satisfying, in different uncertainty conditions. In this paper we present a method to simplify inference in complex ToM models by switching between discrete assumptions about certain belief states (corresponding to different ToM models) based on the resulting surprisal. We report on a study to evaluate a complex full model, simplified versions, and a switching model on human behavioral data in a navigation task under specific uncertainties. Results show that the switching model achieves inference results better than the full Bayesian ToM model but with higher efficiency, providing a basis for attaining the ability for "satisficing mentalizing" in social agents.