Computational Models of Mentalizing

Abstract

Humans have a remarkable ability to infer and represent others' mental states such as their beliefs, goals, desires, intentions, and feelings. In this chapter, we review some of the innovations that have developed in economics, computer science, and cognitive neuroscience in modeling the computations underlying several mentalizing operations. Broadly, this involves building models of how agents infer the mental states of other agents within constrained environments. These models include modules for: representing the goals and desires of an agent (e.g., maximize reward, or minimize embarrassment), inferring the mental states of other agents (e.g., beliefs, goals, desires, intentions, and feelings), and integrating these goals and mentalizing computations to produce optimal behavioral policies to navigate the environment. The mathematical operationalization of these constructs provides a general framework that can be validated by behavior and neural recording techniques and extended in the future by contributions from multiple scientific disciplines.

Theory of Mind and Preference Learning at the Interface of Cognitive Science, Neuroscience, and AI: A Review

“…We also focus more on preference ToM (algorithms which focus on inferring preferences, rather than more generally on inferring mental states), as there has arguably been more recent work on preference-ToM than on belief-ToM (e.g., in the fields of Inverse Reinforcement Learning—IRL—and Preference Learning), and it has been shown that ToM can indeed be cast as an IRL problem (Jara-Ettinger, 2019 ). For broader surveys with less of a focus on very recent IRL and preference learning algorithms, see Rusch et al ( 2020 ) and Gonzalez and Chang ( 2021 ). We also focus on first-order ToM models, since first-order ToM is already challenging enough for current AI models.…”

“…We also focus on first-order ToM models, since first-order ToM is already challenging enough for current AI models. See Gonzalez and Chang ( 2021 ) for a discussion about computational models of higher-order ToM and higher-order ToM in humans and Arslan et al ( 2017b ) for an example of a model of children's development of second-order ToM. Some examples of recent proposals of recursive reasoning models, relevant for higher-order ToM (in a Reinforcement Learning framework), include Wen et al ( 2019 ) and Moreno et al ( 2021 ).…”

“…Gonzalez and Chang ( 2021 ) split computational models of ToM into several broad categories, these include Game ToM, Observational Reinforcement Learning, Inverse Reinforcement Learning and Bayesian ToM. We cover these concepts to varying degrees in this review, but detailed reviews for each are provided in the following papers (Game ToM in Yoshida et al, 2008 ; Observational Reinforcement Learning in Albrecht and Stone, 2018 ; Inverse Reinforcement Learning in Arora and Doshi, 2021 and Bayesian ToM in Baker et al, 2011 ).…”

“…In Game ToM, players' beliefs are represented using probability distributions over actions, states, or other players' beliefs. Psychological Game Theory, a subcategory of Game ToM, models motivations that depend on beliefs (either one's own or another's) and has been used to build formal operationalisations of emotions, as well as to model how players perceive other players' intentions behind their actions [see Gonzalez and Chang ( 2021 ) for more details]. For more examples of multi-agent models of ToM (including Game ToM) see de Weerd et al ( 2017 , 2018 ); Veltman et al ( 2019 ).…”

“…Bayesian ToM (Baker et al, 2011 ) models both beliefs and rewards in a Partially Observed Markov Decision Processes (POMDP) setting, using Bayesian inference. Currently, though, we still have little mechanistic knowledge of how humans perform ToM inferences and the current modeling oversimplifies the mental operations and the environments, leading to uncertainty about how well the current models will generalize (Gonzalez and Chang, 2021 ). Rusch et al ( 2020 ) propose a typology of ToM tasks and computational models across the two dimensions of interaction and uncertainty, suggesting that it is necessary to take into account relatively high levels of both uncertainty and interaction in order to effectively model the more sophisticated ToM processes.…”