Agent-Based Models of Scientific Interaction—Daniel Frey and Dunja Šešelja
Brit. J. Phil. Sci. 71 (2020), 1411–1437
This is a good illustration of the virtues and vices of toy models.
Kevin Zollman's agent-based models of scientific networks indicate, counterintuitively, that better-connected groups of scientists can do worse at converging on true theories. The explanation appears to be that when initial results favour a false theory, the news spreads quickly around the well-connected network, and creates an inertia which the true theory has to struggle a bit to overcome.
F&S add features to Zollman's model which change that outcome. In particular they add a “rational inertia” term whereby scientists don’t shift their preferred theory until the other one has demonstrated superiority over an extended time period. (They also add “criticism”, whereby reports of successful results in the opposing camp prompt re-examination of one’s own views, and dynamic success, whereby theory-testing experiments improve their own reliability over time; these improve performance but don’t change any signs.)
With rational inertia added (along with the other features), the better-connected group does better. The interpretation seems to be that rational inertia slows the spread of any initial misleading reports—just as a lower number of connections would.
So: ABMs in a microcosm. Zollman demonstrates a suprising emergent phenomenon. F&S show that (and what) tweaks make it go away, and so give an idea of what causes it. Both have (on reflection) intuitive real-world parallels/interpretations.
Does any of this accurately reflect what scientists do? Quantitatively no, it is all bizarre nonsense. There are no “rounds” of “pulls” from “bandits”; a prediction that scientific communities will improve performance if rational inertia lasts 10 rounds but not 5, say, is utterly empty. Nonetheless these toy models can provide valuable proofs of concept: that fewer or slower lines of communication in a network can give the truth time to get its boots on.