Bayesian Inference of Other Minds Explains Human Choices in Group Decision Making

Abstract

To make decisions in a social context, humans have to predict the behavior of others, an ability that is thought to rely on having a model of other minds known as theory of mind. Such a model becomes especially complex when the number of people one simultaneously interacts is large and the actions are anonymous. Here, we show that in order to make decisions within a large group, humans employ Bayesian inference to model the "mind of the group," making predictions of others' decisions while also considering the effects of their own actions on the group as a whole. We present results from a group decision making task known as the Volunteers Dilemma and demonstrate that a Bayesian model based on partially observable Markov decision processes outperforms existing models in quantitatively explaining human behavior. Our results suggest that in group decision making, rather than acting based solely on the rewards received thus far, humans maintain a model of the group and simulate the group's dynamics into the future in order to choose an action as a member of the group.

Introducing tomsup: Theory of Mind Simulations Using Python

“…Several other computational implementations of Theory of Mind have been proposed which do not rely on the Free Energy Principle, and do not use variational Bayes to estimate the opponent's parameter values. One such model was designed for multiplayer games, where the agent utilize a ToM-like estimation of the group as a whole and the influence of its own choices on the group (Khalvati et al, 2018). This model outperformed all previous models in predicting participants' behaviour in a Public Goods game (Khalvati et al, 2018).…”

“…One such model was designed for multiplayer games, where the agent utilize a ToM-like estimation of the group as a whole and the influence of its own choices on the group (Khalvati et al, 2018). This model outperformed all previous models in predicting participants' behaviour in a Public Goods game (Khalvati et al, 2018). More technologically oriented approaches also include neural network based ToM models (Rabinowitz et al, 2018;Foerster et al, 2018).…”

“…A more dynamic approach to prior beliefs might also be appropriate in environments with repeated games and multiple opponents. Other non-variational Bayes implementations of Theory of Mind models have utilized having a group-wide representation of intentions (Khalvati et al, 2018), or building their initial beliefs about opponents based on earlier encountered opponents (Rabinowitz et al, 2018). In an evolutionary setting, prior estimates of new ToM agents could be sampled from a distribution with the mean sampled from the estimates made by successful agents, which should allow the prior estimate to drift from the current arbitrary default priors.…”

Classification of Theory of Mind tasks and their computational models

“…The public goods game ( Figure 2B) deploys anonymous group decisions to study fairness and reciprocity in a situation where individual interests conflicts with the group's welfare (Houser & McCabe, 2014;Khalvati et al, 2018): All members of an anonymous group are endowed with the same initial amount. They may choose how much of their individual money to secretly invest in a public fund.…”

“…Capturing a comparable level of sophistication as the influence model, a recent paper by Khalvati et al (2018) utilized a sub-intentional social influence variant of the POMDP framework to model behavior in a public goods game ( Figure 2B). In this version of the task, all other players in the group were displayed as identical avatars, thus rendering the social interaction anonymous (i.e.…”