Copy-when-uncertain: bumblebees rely on social information when rewards are highly variable

To understand the relative benefits of social and personal information use in foraging decisions, we developed an agent-based model of social learning that predicts social information should be more adaptive where resources are highly variable and personal information where resources vary little. We tested our predictions with bumblebees and found that foragers relied more on social information when resources were variable than when they were not. We then investigated whether socially salient cues are used preferentially over non-social ones in variable environments. Although bees clearly used social cues in highly variable environments, under the same conditions they did not use non-social cues. These results suggest that bumblebees use a ‘copy-when-uncertain’ strategy.

distribution with k = 0.183 and θ = 45.45; for comparison with our earlier model [1] this equates to a Gini index of G = 0.81). The average amount of resources was the same in both conditions. In high-variance simulations patch resources π change with a constant probability τ = 10 −4 in each time-step. In every time-step all patches are replenished to their according value of π.
Individuals forage for resources in patches. Individuals can exploit patches they know about, either through personal exploration (individual learning) or through observation of other individuals exploiting a patch (social learning). Each individual is either an individual learners or a social learner. Individual learners learn about a randomly selected new patch, while social learners learn about a patch of a randomly selected individual that exploited a patch in the last round. In both cases individuals learn the amount of resources a patch currently provides, i.e. its anticipated reward.
Individuals exploit with probability β = 0.8 (thus, they learn with probability 1− β ). If an individual chooses to exploit, it selects the patch with the highest anticipated reward from its memory. The amount of resources the individual collects is p = π/n, where n is the number of individuals that chose the same patch in this time-step. Thus, individuals engage in exploitative competition and equally share resources.
Information about anticipated rewards persists in an individual's memory until it is updated either when the individual learns the same patch again or gains resources different from the anticipated reward, or until the individual dies. learner with fitness 2, then the new individual will be an individual learner with a 75% chance. We use the relative amount of resources collected by an individual as a performance measure that is commonly used as fitness proxy in foraging theory [5].
This way, the initial proportion of individual and social learners (1:1) evolves to an equilibrium point where the fitness of individual and social learners does not differ.
We ran simulations with 100 patches and 33 individuals for 10,000 time-steps. In each of the last 2,500 rounds of each simulation, we recorded the proportion of individual and social learners. We conducted 100 simulations for each resource variance. Resource variance (high/no-variance) and the ratio between cues and flowers (1:3) of our empirical experiments matched those of our simulation.

S2 Animals and treatment
We used three bumblebee colonies (Bombus terrestris), provided by Koppert

S3 Additional observations beyond the first landing
For all test flights we observed all landings on flowers. In Fig. S1a and S2a, additional to the proportion of bees landing on flowers with a cue ('first', which is identical to the results in Fig. 1b), we report the average proportion of landings on flowers with social or non-social cues for the 'first four' and 'first ten' landings.
Again for bees trained with social cues we find significant differences between bees from the high-and no-variance treatment (Pearson's Chi-squared test for first four landings: P < 0.01; Pearson's Chi-squared test for first four landings: P < 0.001, Fig.   S1a). However, bees trained with non-social cues did not differ from each other, neither for the first four nor for the first ten landings (Pearson's Chi-squared test for both: P > 0.05, Fig. S2a). Only bees trained to social cues and the high-variance treatment landed more often flowers with a social cue in their first four and first ten landings than would be expected by chance alone (binomial test for both: N = 16, P < 0.001, Fig. S1a). Bees from all the other treatments did not land more on flowers with a cue than would be expected by chance alone (binomial test for social cue, novariance, and non-social cue and high-and no-variance: P > 0.05, Fig. S1a and S2a).
In Fig. 1b of the main text we present only results for the first landing while earlier studies have shown that the decision to use social information can strongly differ between first and subsequent landings [7]. Based on the effect a rewarded flower without a cue or a non-rewarded flower with a cue could have on the subsequent choice, we decided to only look at the first landing because at this stage all foragers have the same knowledge about yellow flowers and cues. Especially for bees from the no-variance treatment this is the first encounter of yellow, water-filled.
The results for the subsequent landings show a slight increase (no-variance, Fisher's Exact test P = 0.16) or slightly decreased (high-variance treatment, Fisher's Exact test P = 0.39) in the use of cues for bees from the social cue training between the first and the second landing (Fig. S1b). Interestingly, for bees from the non-social cue training this effect appears to be reversed between the two different treatments, where the use of non-social cues was slightly decreased (high-variance, Fisher's Exact test P = 0.09) or showed no change (no-variance, Fisher's Exact test P = 1) (Fig. S2b). However, none of these are statistically significant.