Bayesian Framework Assesses Scenario Compatibility in Population Synthesis

In transportation analysis, future scenarios are often defined by aggregate population targets, which are then used to constrain generative models that produce detailed individual-level populations. Current methods typically rely on deterministic calibration, assuming these targets are inherently compatible with the generative model's underlying structure. This research highlights a critical gap: whether these scenario-level constraints actually fall within the model's learned distribution and how much they might distort the inherent structural uncertainty. To address this, the paper proposes an ensemble-based Bayesian updating framework designed to quantify scenario compatibility within conditional population synthesis. It employs a population-aware conditional variational autoencoder (CVAE) to learn a distribution of plausible population structures while maintaining aggregate accuracy. An ensemble of samples from this CVAE's prior distribution provides an empirical representation of structural uncertainty. Scenario targets are then treated as probabilistic evidence, and Bayesian updating is applied across the ensemble to derive posterior weights. The compatibility of a scenario is measured using the effective sample size (ESS), which indicates the posterior concentration and the degree to which structural uncertainty is compressed by the conditioning. Experiments demonstrate that a scenario's impact depends not only on the magnitude of its targets but also on how well they align with the learned joint structure, revealing potential structural failure modes when targets fall outside the model's learned support.