Order, Confidence, and Division of Labor in Groups of Rational Agents

Humans and other animals accumulate information to make decisions under uncertainty. Mathematical models show that even groups of rational agents make counterintuitive choices. For instance, the order of decisions can strongly impact accuracy: in large groups, the first agents to decide almost always hold the strongest initial bias and decide accordingly, while slow agents decide as if they were unbiased. Yet all agents are equally confident, even when their accuracy differs dramatically. In the second part of the talk, I will discuss a normative model of collective foraging in which agents choose whether to commit to costly exploration or remain idle, with rewards and information shared across the group. An optimal policy can be readily found using dynamic programming and studied using asymptotic analysis, but how it can be implemented without coordination is unclear. I will show that optimal performance can be achieved without coordination when the community is heterogeneous, and that such communities evolve in a simple replicator model. These communities feature a small, diverse subset of risk-taking scouts and a majority of synchronized deliberators who commit only when conditions are likely to be rewarding. Together, these results show how groups can make accurate collective decisions without central coordination.