nucleation

THEORY-422: In the standard formal models of group learning, no communication network is best for every epistemic problem: costly transmission favours minimally connected or star networks, pooling of estimates by averaging favours regular networks, and learning by trying in bandit problems can favour sparse ones, so a group's network cannot be judged until its learning problem is known

3 min read

Properties4
StatusActive
Tagsepistemology, network-science, philosophy-of-science
SourceLIT-716 — Network Epistemology: Communication in Epistemic Communities
ExtendsTHEORY-143 — When a group learns only about the options its members pursue and good evidence is scarce, sparser communication makes it more likely to settle on the better option, at a large cost in speed; when evidence is plentiful, sparsity only slows learning

Graph View

Source

Zollman (2013), LIT-716, “Network Epistemology: Communication in Epistemic Communities”, read in full in NOTE-567 from the author’s manuscript of 5 November 2012.

What was actually shown

The survey sorts results by kind of learning problem. Each is a result of its source, under that source’s assumptions.

  • Transmission (§2). With lossless second-hand transmission and costly links, minimally connected networks are optimal. With loss over distance, stars are. Optimal networks are usually stable but may not be reached.
  • Pooling (§3.1). Under DeGroot averaging, regular networks are optimal in the long run, and more connected ones converge faster.
  • Specialised knowledge (§3.2). A few heavily contacted individuals emerge naturally, so prominence is not evidence of sociological distortion.
  • Active learning (§4). With a pull toward the truth and a connected network, structure does not affect the limit. In bandit problems, minimally connected networks did best in small simulated groups.

The switches that change the answer are named: whether information passes on second-hand, whether links cost anything, whether evidence arrives independently of what agents do, and whether agents learn new facts while deliberating.

What could have come out otherwise. One structure, say the complete network, could have been optimal across all the problem types. In these models it is not.

What this does not say

  • That real communities behave like the models. All of it is idealised models, proved or simulated. Where real groups sit in parameter space is not addressed.
  • That less communication is generally better. That holds only in the bandit case, and LIT-733 later narrowed even that to hard problems (THEORY-143). For pooling, denser is faster.
  • That a network optimal in principle is reachable. How groups find their way to optimal structures is the survey’s own open question.
  • Anything beyond utilitarian optimality. Fairness and robustness to link failure are outside most of the models.
  • Anything about sharing reasons. The agents share attitudes, not arguments; Betz (2012) is the one exception cited.

Connections

  • It widens THEORY-143, Zollman’s own bandit result as narrowed by Rosenstock et al., into a landscape in which that result is one case. THEORY-148 (transient diversity) is the active-learning mechanism behind it.
  • It is the record’s best single source for the claim that research on epistemic networks complicates any simple claim that more communication improves collective judgment (NOTE-567). §1 names corporations as group believers.
  • The one-step theory-averaging result of THEORY-649 is a pooling model of a different kind, over logical theories rather than estimates.

Backlinks