Social learning solves the problem of narrow-peaked search landscapes: experimental evidence in humans

Abstract

The extensive use of social learning is considered a major reason for the ecological success of humans. Theoretical considerations, models and experiments have explored the evolutionary basis of social learning, showing the conditions under which learning from others is more adaptive than individual learning. Here we present an extension of a previous experimental set-up, in which individuals go on simulated 'hunts' and their success depends on the features of a 'virtual arrowhead' they design. Individuals can modify their arrowhead either by individual trial and error or by copying others. We study how, in a multimodal adaptive landscape, the smoothness of the peaks influences learning. We compare narrow peaks, in which solutions close to optima do not provide useful feedback to individuals, to wide peaks, where smooth landscapes allow an effective hill-climbing individual learning strategy. We show that individual learning is more difficult in narrow-peaked landscapes, but that social learners perform almost equally well in both narrow- and wide-peaked search spaces. There was a weak trend for more copying in the narrow than wide condition, although as in previous experiments social information was generally underutilized. Our results highlight the importance of tasks' design space when studying the adaptiveness of high-fidelity social learning.

Keywords: cultural evolution; cultural transmission; social learning.

Figure 1.

Examples of a wide- (blue line) and a narrow-peaked (red line) search landscape. The two peaks correspond to two arbitrary values (30 and 70) of the attribute. The search landscape shown here was used for the continuous attributes of the arrowhead, i.e. height, width and thickness, but with different optimal values for each. The red line has an enforced minimum of 560 calories to ensure that there is equal area under both lines.

Figure 2.

Performance (score in calories per hunt) over time (i.e. hunt) across the conditions and seasons. Scores started out at similar values, but diverged in the different conditions: individual learners performed better in the wide condition, while social learners performed similarly well in wide and narrow conditions. Error bars show 95% confidence intervals.

Figure 3.

Difference in final hunt score between wide and narrow conditions in (a) individual learners, (b) social learners’ non-normalized raw scores and (c) social learners’ normalized scores to account for differences in demonstrator scores between the two conditions. Each point represents one participant’s mean score across all three seasons. Boxplots show medians and interquartile ranges, with whiskers extending to 1.5 IQR.

Figure 4.

Comparison of copying frequency in the narrow and wide conditions, across the three seasons. The value shown is the proportion of hunts on which participants chose to copy, from 0 (never copied) to 1 (always copied). The size of the circles are proportional to the number of participants at that frequency. Boxplots show medians and interquartile ranges, with whiskers extending to 1.5 IQR.

Figure 5.

Relationship between copying frequency and final normalized cumulative score across social learners in the wide and narrow conditions. Lines are best-fit multilevel regression lines with season as a random factor. Shaded areas show 80% prediction intervals calculated using the predictInterval function from package merTools [39].