Optimality principles guide how animals adapt to changing environments. During foraging for nonsocial resources such as food and water, species across taxa obey a strategy that maximizes resource harvest rate. However, it remains unknown whether foraging for social resources also obeys such a strategic principle. We investigated how primates forage for social information conveyed by conspecific facial expressions using the framework of optimal foraging theory. We found that the canonical principle of Marginal Value Theorem (MVT) also applies to social resources. Consistent with MVT, rhesus macaques (Macaca mulatta) spent more time foraging for social information when alternative sources of information were farther away compared to when they were closer by. A comparison of four models of patch-leaving behavior confirmed that the MVT framework provided the best fit to the observed foraging behavior. This analysis further demonstrated that patch-leaving decisions were not driven simply by the declining value of the images in the patch, but instead were dependent upon both the instantaneous social value intake rate and current time in the patch.
Optimal foraging theory describes the behavior of animals seeking out resources in a patchy environment according to an energy-maximizing strategy. As animals use the supply of resources within a patch, patch value naturally declines and animals must decide when to leave the current patch in search of a new one. The Marginal Value Theorem (MVT) is a broadly applied optimality model that predicts foraging behavior in a variety of taxa1,2,3. MVT predicts that animals should leave the current patch when the energy intake rate within the patch diminishes to the average energy-harvesting rate in the environment4,5. Thus, the time that animals spend within a patch (i.e., patch-residence time) depends upon a variety of factors, including the value of the current patch (in terms of the resource being consumed), the value of other patches in the environment, and the time it would take to travel to the next closest patch (i.e., travel time).
We hypothesized that (1) residence time within patches would increase with travel delay time, as predicted by MVT; and (2) an MVT-based model including both t and r(t) terms would outperform other models in predicting patch-leaving behavior.
In our Social Information Foraging Task, monkeys foraged in three environment types (1): color-cued emotional valence (2), fractal-cued emotional valence, and (3) color-cued social vs nonsocial (Fig. 1b). In both emotional valence environments, negative patches contained images of negative-valence expressions (threat, fear) and non-negative patches contained images of non-negative expressions (coo, lipsmack, neutral). In the first environment, red and blue targets cued negative and non-negative facial expressions, respectively, whereas in the second environment black-and-white fractal patterns were used to cue negative and non-negative facial expressions. This was done to ensure that preferences were based on image valence rather than target color. In the color-cued social vs nonsocial environment, we assessed the baseline social preference in the absence of explicit emotional valence. Social patches contained images of conspecifics with neutral facial expression, and nonsocial patches contained scrambled versions of these images.
Behavioral patterns of social information foraging. (a) Daily contrast ratios (CR) indicating preference in color-cued emotional valence, fractal-cued emotional valence, and color-cued social vs nonsocial environments (see Methods). Narrow horizontal black lines indicate mean daily CRs (within-subject) and wide horizontal lines indicate standard errors. (b) Patch choice order within the color-cued emotional valence environment presented as the cumulative proportion of negative-valence (red) or non-negative (blue) patches. Black lines indicate the means (within-subject) and shading indicates standard errors.
We compared four models of patch-leaving probability to assess the validity of using an MVT-based approach. Comparison of the models (Table 1) revealed that the MVT-based simulation that included both r(t) and t terms consistently provided the best fit to the observed data (Fig. 5). Across all travel times, AIC values from this model were lower than those from all other models (Table 1), and the relative likelihood of this model, calculated as the ratio of model weights, was greater than zero across all pairwise model comparisons (Fig. 6). These results indicate that social information foraging in this task is better explained by the MVT framework than a simple decline in subjective valuation over time, and depends upon both the instantaneous value intake and current time in patch.
Fit comparison of simulations based on four patch-leaving models to observed social information foraging behavior. Probability of leaving the patch over time, comparing observed behavior (black) with simulations using (1) the adapted MVT model which included terms for social value r and time t, (2) a model with only a social value term (i.e., model 1 with time terms excluded), a model with only time terms (i.e., model 1 with value term excluded), and hyperbolic delay-discounting model for each travel time (see Methods).
Our results extend the explanatory power of MVT in intangible information foraging to the social domain, providing novel evidence that animals use a similar strategy to enhance social resource intake as they use to maximize primary resource intake2. The demonstration that social information altered pure juice maximization suggests that individuals may balance competing strategies for optimizing values of social and nonsocial resources. The degree of symmetry in balancing foraging strategies across multiple resources likely depends on the value distributions across the social and primary resources available. Individuals may adaptively adjust the relative weight of primary versus social resources depending upon a variety of factors, including environmental conditions and availability of potential mates and social partners. Future work can help clarify this contingency by systematically manipulating the resource ratios between primary and social resources. Finally, phenotypic differences among individuals may also contribute to differences in social information use30 and predictably relate to social competency.
When a monkey selected a patch, one image was randomly drawn from the bank of task stimuli for the corresponding category (i.e., negative or non-negative; social or nonsocial) with randomized replacement. In both emotional valence environments, open-mouth threat and fear grimace were grouped as negative-valence expressions23, and lip smack24,25, coo23, and neutral faces were grouped as non-negative expressions (Fig. 1b). Open-mouth threat is typically given by dominants to subordinates as an aggressive approach signal23. Fear grimace is a negative withdraw signal typically given by subordinates to dominants23. Coo is an affiliative approach signal typically given during grooming interactions, eating, or affiliative approaches23, and lip smack is an approach signal typically given by subordinates to dominants or by individuals engaged in grooming23, which is often followed by affiliation25.
where E n is net energy intake rate (in this case, the net social value intake rate), P i is the proportion of patches of a given type, g i is the amount of energy gained corrected for the energetic cost of searching (the social value intake function for a given image over time in the patch), T i is time spent hunting (time in patch), t is travel time (time from selecting the travel bar to returning to fixation), and E T is the energy cost per unit time during traveling6.
We made the same assumptions made by a previous study investigating nonsocial resource foraging in rhesus macaques2: that the (1) cost of searching was zero, (2) energy cost per unit time during travel was zero, (3) patches of a given type were identical, and (4) handling time was constant across patches (400 ms). We made the simplifying assumption that the information capture rate function g(t) is the same for all patches in a given environment, i.e., there is effectively one patch type in each environment in our task. Thus, the cost of search, travel, handling time, and rate of information capture remained constant, but energy (i.e., social value) gain per encounter varied among patches and declined as a patch was exploited. Under these assumptions, net energy, or social value, intake rate of a patch was defined as
The value of the patch decreases over time spent in the patch because the social value of the image diminishes with time as social information is harvested and novelty declines. Maximal value intake rate is achieved when the intake rate within a patch is equal to the average intake rate for the environment4.
With g(T), the social value intake rate E n was quantified as a function of time in patch across travel times (Eq. (3); Fig. 4c). We then identified the maximum social value intake rate for each travel time, representing optimal time in patch as a function of travel time, a prediction of how monkeys should behave to optimize social information foraging according to MVT. We also simulated expected behavioral results if monkeys optimize juice - a primary, nonsocial resource. In our task, monkeys could maximize juice intake rate by leaving each patch immediately, regardless of travel time and without viewing the image. Thus, patch-residence time under optimal juice foraging should ignore the travel time and therefore is theoretically the minimum fixation required to select the travel bar, 200 ms (when not considering any nonspecific time required for eye movement) across all travel time values (i.e., slope of zero). Because the animals could be looking at any position inside or outside the screen (outside the limit of eye tracking) when the image was first revealed, calculating the nonspecific eye movement related times becomes problematic. For the purpose of displaying the theoretical optimal juice and social foraging curves, we added an additional 250 ms to both curves as a minimum estimate of saccade reaction time and movement time. 781b155fdc