Theory of mind: did evolution fool us?

Abstract

Theory of Mind (ToM) is the ability to attribute mental states (e.g., beliefs and desires) to other people in order to understand and predict their behaviour. If others are rewarded to compete or cooperate with you, then what they will do depends upon what they believe about you. This is the reason why social interaction induces recursive ToM, of the sort "I think that you think that I think, etc.". Critically, recursion is the common notion behind the definition of sophistication of human language, strategic thinking in games, and, arguably, ToM. Although sophisticated ToM is believed to have high adaptive fitness, broad experimental evidence from behavioural economics, experimental psychology and linguistics point towards limited recursivity in representing other's beliefs. In this work, we test whether such apparent limitation may not in fact be proven to be adaptive, i.e. optimal in an evolutionary sense. First, we propose a meta-Bayesian approach that can predict the behaviour of ToM sophistication phenotypes who engage in social interactions. Second, we measure their adaptive fitness using evolutionary game theory. Our main contribution is to show that one does not have to appeal to biological costs to explain our limited ToM sophistication. In fact, the evolutionary cost/benefit ratio of ToM sophistication is non trivial. This is partly because an informational cost prevents highly sophisticated ToM phenotypes to fully exploit less sophisticated ones (in a competitive context). In addition, cooperation surprisingly favours lower levels of ToM sophistication. Taken together, these quantitative corollaries of the "social Bayesian brain" hypothesis provide an evolutionary account for both the limitation of ToM sophistication in humans as well as the persistence of low ToM sophistication levels.

Figure 1. MCMC average payoffs of all pairs of ToM agents.

This figure depicts the MCMC average of the payoff matrices for both “hide and seek” (left) and “battle of the sexes” (right) after learning has occurred. The i^th line gives the accumulated payoff of the i^th type of agent, when playing against each and every other ToM phenotype. Note that the absolute payoff levels of both types of games cannot be compared.

Figure 2. Accuracy of behavioural predictions in competitive and cooperative contexts: example of 0-ToM playing against 1-ToM.

The behavioural prediction of ToM players (y-axis) is plotted against her opponent’s true behavioural tendency (x-axis) for each trial of a simulated repeated game with trials. The grey line indicates the best-fitting straight line in the data. Upper half: “Hide and Seek”. Lower half: “Battle of the Sexes”. Left: accuracy of 1-ToM predictions when playing against 0-ToM. Right: accuracy of 0-ToM predictions when playing against 1-ToM.

Figure 3. MCMC average prediction accuracy of all pairs of ToM agents.

This figure depicts the MCMC average of the linear trend between the behavioural prediction of ToM players and their opponent’s true behavioural tendency . In other words, this corresponds to the slope of the best-fitting straight line in Figure 2. The figure uses the same format as Figure 1.

Figure 4. MCMC empirical distribution of learned opponent’s sophistication level for “twin” pairs of ToM agents.

Each bar gives the number of MCMC simulations (z-axis) that led to each particular combination of belief , both agents had on each other’s ToM sophistication level (x/y-plane). Histograms are truncated to the upper-left triangle for visualization purposes (they are symmetrical by construction). Upper half: “Hide and Seek”. Lower half: “battle of the Sexes”. Left: “twin” pairs of 2-ToM agents, Middle: “twin” pairs of 3-ToM agents, right: “twin” pairs of 4-ToM agents.

Figure 5. Replicator dynamics for purely cooperative and competitive social interactions.

The frequency of each ToM phenotype (y-axis) is plotted against evolutionary time (x-axis), for 128 different simulations with different initial conditions. Different ToM traits correspond to different colours (see legend). Pie charts depict the evolutionary stale states, i.e. the equilibrium or fixed point, replicator dynamics converge to (the colour coding is the same). Upper half: “Hide and Seek”. Lower half: “battle of the Sexes”.

Figure 6. Phase diagram of ToM evolution.

Each pie chart depict the evolutionary stable state that is induced by a particular combination of amount of learning τ (x-axis) and proportion ω of cooperative interactions (y-axis).

Figure 7. Phase diagram of ToM evolution: Impact of RL and Nash phenotypes.

This figure uses the same format as Fig. 6.