Pre-Estimation Checks

Read this page before fitting AIRL-Het. The anchors and segment separation are not secondary diagnostics; they are what make heterogeneous reward recovery interpretable.

AIRL-Het has a richer failure surface than single-segment estimators because segment identification depends on both the anchor design and the behavioral signal that separates segments. Check these before fitting:

Check	Why it matters for AIRL-Het
Anchor validity	If the exit-action index or absorbing-state index is wrong, the anchor constraints enforce the wrong normalization and reward recovery fails.
Feature rank	A rank-deficient design matrix leaves reward directions unidentified even under correct anchors.
Feature condition number	Ill-conditioning inflates the adversarial gradient noise and slows convergence.
State coverage	Reward recovery depends on the discriminator seeing transitions from most states. The three unobserved states in the primary cell are a known boundary.
State-action coverage	Rare action-state pairs are weakly identified. The minimum action share measures exposure to rare actions.
Segment behavioral separation	EM identifies segments only if the segments choose differently across states. Very similar segments require more data to separate.
Within-individual trajectory count	The consistency constraint requires at least two trajectories per individual to be useful; single-trajectory users contribute to the prior but not to within-user smoothing.
Transition row sums	Transition tensors must be row-stochastic in the \((A, S, S)\) orientation.

Primary Simulation Checks

Values from the primary airl_het_paper_identification run recorded in aairl.json:

Check	Value	Status
Feature rank	20 / 20	pass
Feature condition number	20.411	pass
Observed states	58 / 61	context
State-action coverage	0.934	pass
Minimum action share	0.204	pass
Max transition row error	0.0	pass
Anchor valid	true	pass

The three unobserved states are outside the simulation support by design and do not indicate a data problem. They remain visible here because segment heterogeneity claims are support-sensitive.

Common Risk Patterns

Wrong anchor indices. exit_action and absorbing_state must be the correct integer indices for the specific environment. There are no defaults; the estimator raises a ValueError if either is missing.

Segment collapse. One segment can absorb most of the prior mass if the true behavioral difference is small or the initialization is unlucky. Monitor segment_priors during EM; a prior near zero for any segment signals collapse. The prior_min and prior_smoothing settings resist collapse.

EM stopping too early. The primary cell converges in 2 EM iterations, but more complex DGPs may require more. Inspect em_log_likelihoods in the result metadata to confirm the LL has stabilized before treating the output as final.