Contribution
A short review of economics and philosophy results on how a group’s internal communication structure affects what it comes to know, organised by type of learning problem.
Key insight
There is no epistemically best network. A group cannot be criticised for communicating too little, or too hierarchically, until one knows what kind of learning problem it faces.
Assumptions
Varies by model, and the survey is explicit about it. The main switches are whether information can be passed on second-hand, whether links cost anything, whether evidence arrives independently of what agents do, and whether agents learn new facts during deliberation.
Key results
- Transmission (§2): with lossless second-hand transmission and costly links, minimally connected networks are optimal; with distance loss, stars. Optimal networks are usually stable but may not be reached.
- Pooling (§3.1): under DeGroot averaging, regular networks are optimal in the long run; 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 truth-pull and connectivity, structure is irrelevant to the limit; in bandit problems minimally connected networks did best in small simulated groups.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Which network is best depends on the learning problem | strong | the surveyed results, each proved or simulated in its source |
| C2 | Unequal centrality is not by itself evidence of an epistemic pathology | moderate | follows from star and specialised-knowledge results under their assumptions |
| C3 | In bandit problems less communication is better even when free | moderate | Zollman’s simulations; later narrowed by LIT-733 |
Concepts
- one way / two way, one pays / both pay — three cost models for a communication link.
- truth pull — an idealised small attraction of some agents’ beliefs towards the truth during averaging.
Connections
Surveys Bala & Goyal, Jackson & Wolinsky, DeGroot, Hegselmann & Krause, Anderson, and the author’s own work (LIT-724, LIT-735).
Bearing on the record
Supports THEORY-143 by placing it in a wider landscape where denser is sometimes better. It is the best single citation for the owner’s sentence that research on epistemic networks “complicates any simple claim that more communication or tighter integration necessarily improves collective judgment”. It also bears on the essay’s corporate setting: §1 names corporations as group believers.
Limitations
A survey of idealised models. Utilitarian optimality only; fairness and robustness to link failure are noted as outside most models.
Open questions
How groups find their way to optimal structures, and models of agents sharing reasons rather than attitudes (Betz 2012 is the one example cited).