The problem
Agent-based models (ABMs) let researchers specify simple, local behavioral rules and
observe the aggregate patterns those rules produce. But this creates an inverse
problem: given only the aggregate pattern, many different underlying rule sets can
often reproduce it equally well. This property — equifinality — means that fitting
a single “best” mechanism to observed data can be misleading, since it hides the range
of mechanisms that would have looked just as plausible.
Why equifinality matters
For fields that rely on agent-based modeling to explain social phenomena — from
segregation dynamics to information diffusion — treating a fitted rule set as the
explanation risks overstating what the data can actually tell us. A more honest
scientific stance is to ask: given what we observed, what is the full range of
mechanisms that remain plausible?
From observation to mechanism
This project treats mechanism identification as an inference problem: start from
simulated or empirical observations, and work backward toward the set of behavioral
rules consistent with them, rather than a single point estimate.
Our approach
The approach combines agent-based simulation with machine learning models trained to
map observed outcomes back to candidate generative rules, then uses conformal
prediction to produce calibrated, uncertainty-aware sets of plausible rules rather
than a single prediction.
Conformal prediction provides a statistically principled way to output a set of
plausible answers with a guaranteed coverage rate, rather than a single answer with an
unstated (and often overconfident) level of confidence. Applied to rule identification
in ABMs, this produces certified sets of behavioral rules that remain consistent with
the observed data.
My contribution
Chathura is developing this methodology as lead author, including the framing of rule
identification as a conformal-prediction problem, the experimental design for testing
it on agent-based segregation models, and the resulting conference presentation and
in-preparation manuscripts.
This project extends directly from Chathura’s earlier work on information diffusion
and social simulation — both areas where the same underlying question recurs: what
process could plausibly have generated what we observe?