Fig 1.
Network structures used in the two experimental conditions.
The efficient network was a fully connected network (left), and the inefficient network was a one-dimensional lattice where each node was connected to its immediate four closest neighbors (right).
Fig 2.
Best solution discovered in 8 experimental trials.
In each of the eight experimental trials, groups with inefficient networks (orange) found better solutions than groups with efficient networks (blue). Within each trial (1–8), both groups were given the same data set and prediction problem, and began with identical initial solution distributions across the population. Across each trial, different data sets and prediction problems were used. Figures for each trial are scaled based on the best possible solution to the problem (Performance = 1) compared to both groups’ average starting performance within a trial (Performance = 0). In trials 5, 6, 7, and 8, the groups in inefficient networks found the best possible solution, which was never found in any group with an efficient network.
Fig 3.
Temporal sequence of solution discovery for all members of both groups in a single trial.
Solution quality in rounds t = 1, 2, 5, 10, and 15 (A to E) in a single trial (i.e., Trial 6). In the initial rounds, (A) t = 1, and (B) t = 2, individuals in the efficient network (blue) converged on a small set of solutions, whereas individuals in the inefficient network (orange) explored a greater diversity of solutions. (C) t = 5, the individuals in the efficient network showed little improvement, whereas individuals in the inefficient network had greater solution diversity and better solution quality. (E) t = 15, nearly every member of the inefficient network had converged on the best solution, whereas those in the efficient network showed little improvement from their initial solutions. Performance is scaled based on the best possible solution (Performance = 1) compared to both group’s starting performance (Performance = 0).
Fig 4.
Time series showing the tradeoff between diffusion and diversity.
The dynamics of solution discovery in groups with efficient (blue dashed) and inefficient (orange solid) networks shows that inefficient groups performed better in terms of solution diversity (A and B) and solution quality (C and D). All panels plot the average values for each experimental condition over all eight trials. In groups with efficient networks, good solutions rapidly spread to other group members, whereas diffusion was slower in inefficient groups (A). Diffusion is measured by the fraction of individuals who adopted the best available solution in the group over time. Due to this slower rate of diffusion, groups with inefficient networks discovered more distinct solutions (B), and the quality of the best solutions in these groups was much higher, both in terms of the value of the best solution found (C) and the fraction of the population that adopted a solution that was better than the best available solution from the other network (D).