Analogical Theory of Mind: Computational Model and Applications

Rabkina, Irina Rabkina

doi:https://doi.org/10.21985/n2-w8cm-rr04

Work

Analogical Theory of Mind: Computational Model and Applications

Public

Download PDF

Download All Files (.zip)

Theory of mind (ToM) reasoning is defined as the ability to reason about another’s internal states, such as beliefs, goals, and desires. It is a major aspect of human social interaction and is mastered by most typically developing children by age five. On the other hand, simulated agents generally lack such reasoning abilities—even in situations when they must interact with others. This dissertation shows that human-like ToM reasoning improves simulated agents’ decision-making in complex multi-agent environments. The major contribution of this work is the Analogical Theory of Mind (AToM) model, an implemented computational cognitive model of human ToM reasoning and development. AToM claims that human ToM reasoning and development occur via analogical processes (i.e., Gentner’s Structure-mapping Theory). This claim is tested through simulations of three related phenomena: (1) children learning ToM from structured stories; (2) children learning ToM as a side effect of learning a complex grammatical structure; and (3) children’s failures in pretend play. AToM successfully models the children’s performance in each and makes testable predictions. The model is then applied to simulated agents reasoning in two complex multi-agent environments. First, it is used for goal recognition in the Minecraft game. AToM slightly underperforms a state-of-the-art goal recognition system under standard goal recognition conditions (i.e., when reasoning from planner outputs), and significantly outperforms it when reasoning from only external observations. Next, AToM is used to recognize intent to cooperate among simulated players in stag-hunt, a prisoner’s dilemma-style game. AToM’s performance does not differ from a Bayesian model or human performance on a limited dataset from the literature and performs well when the dataset is extended in size and complexity. These results suggest that using AToM during reasoning can improve multi-agent interaction.

Creator