Contribution
It showed in finite simulated communities that less communication can make a group more reliable. It located the cause in connectivity rather than in a central hub, and named a speed–reliability trade-off.
Key insight
When you learn only about what you try, a run of bad luck on the better option can lead everyone to stop trying it. After that, no evidence can bring them back. Dense communication makes everyone see the same bad run at once; sparse communication leaves pockets of people still trying it.
Assumptions
- Two states, two actions: A1 with the same known expected payoff in both states (uninformative), A2 better in one state and worse in the other.
- Agents choose myopically the action with highest expected utility, and update by simple Bayes on their own and neighbours’ outcomes, not on the fact that a neighbour chose an action (n. 4).
- Initial beliefs uniform on (0, 1); the world set to the state in which A2 is better.
- A run ends when everyone plays A1 (failure) or everyone believes the truth above 0.9999 (success).
Key results
- Cycle > wheel > complete in probability of success, for every size studied; complete fastest (Figures 2–3, 10,000 runs per size).
- Six-agent census. Density is the strongest single predictor of failure; clustering adds a little; degree variance is uncorrelated with success (§3.2).
- Mechanism. In connected networks four times as many agents switch strategy after a disappointing round, and with one A2 player left a connected network is almost three times as likely to have none next round (§3.2).
- Trade-off. Four of the five most reliable six-agent graphs are minimally connected; the five fastest are complete or nearly so (Figure 5).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | In this model sparser networks are more reliable and slower | strong | simulation over all networks of up to six agents |
| C2 | Connectivity, not hub centrality, causes the loss | moderate | regression on graph statistics, six-agent networks |
| C3 | The ordering holds for any parameter setting | weak | footnote 7; refuted by LIT-733 |
| C4 | Limiting information can maintain the division of cognitive labour without epistemically impure motives | moderate | argued in §5 from the model |
Method
Agent-based simulation of the Bala–Goyal model on fixed graphs: cycle, wheel and complete graphs of 3–12 agents, and all graphs up to isomorphism of 3–6 agents, 10,000 runs each.
Concepts
- royal family — Bala & Goyal’s agents connected to everyone; the wheel’s hub in the finite case.
- density — the proportion of possible links present.
Connections
Takes its model from Bala & Goyal (1998) and contrasts its solution with Kitcher’s and Strevens’s reward-based accounts of the division of cognitive labour, which require scientists to pursue theories they think less likely for the sake of credit.
Bearing on the record
Source of THEORY-143, with Rosenstock et al. as the second source fixing its scope. It supports the owner’s essay (§8.2) in the weak form the essay states: more communication does not necessarily improve collective judgment. The paper’s own scope is narrow, as the Limitations below say.
Limitations
Small networks; one payoff gap (the paper says absolute probabilities can be “manipulated by altering the expected payoffs”); myopic agents. The model needs an uninformative action whose payoff is known to all, and outcome evidence that depends on choices. Section 4 grants that its second and third assumptions are “less innocuous”.
Open questions
Which real communities are in the parameter region where this holds. The paper does not say, and LIT-733 suggests it is a small one.