Pre-Estimation Checks

Read this page before fitting f-IRL. The primary question is whether the expert state marginal is well covered enough for behavioral matching to be meaningful.

f-IRL estimates a tabular reward by gradient ascent on a divergence objective, so its failure modes differ from structural estimators. Check these before fitting:

Check	Why it matters for f-IRL
State coverage	The expert state marginal is empirical; unobserved states carry zero expert density and will drive the reward toward negative infinity under forward KL.
Reward range after fitting	A near-zero reward range signals a flat-reward solution; stop and investigate before reading other metrics.
Action support per state	A state where only one action is ever taken may still be usable, but sparse action support degrades log-likelihood tracking.
Transition row sums	Transition tensors must be row-stochastic in the (n_actions, n_states, n_states) orientation.
Horizon length	The forward propagation runs for horizon steps; a horizon that is too short underestimates long-run state visitation.
Learning rate and clip bound	Too large a learning rate causes oscillation; too small a clip bound truncates the reward signal.

Canonical Simulation Checks

Values from the primary source-paper benchmark (see Simulation Study):

Check	Value	Status
Observed states	8 / 8	pass
State-action coverage	1.000	pass
Feature rank	3 / 3	pass
Feature condition number	18.471	pass
Minimum action share	0.320	pass
Anchor valid	true	pass
Reward range (post-fit)	1.292	pass

Common Risk Patterns

Flat reward. Forward KL is the most common divergence to collapse to a flat solution when the expert marginal has near-zero mass at some states, because the gradient log-ratio becomes numerically unstable. Check reward_range first; if it is near zero, try reducing the learning rate or switching to Jensen-Shannon.

State coverage gaps. If the expert panel does not visit all states, those states have zero expert marginal. Forward KL assigns infinite cost to the model visiting them; in practice this is clipped, but the resulting reward at unvisited states is not trustworthy. If coverage is thin, prefer a feature-based IRL method such as MaxEnt-IRL.

Marginal scope mismatch. Using marginal_space="state_action" with reward_scope="state" is not supported and raises a configuration error. Action-dependent marginal matching paired with an action-dependent reward (reward_scope="state_action") is supported but is a diagnostic exercise under current evidence; see the Identification and Applicability sections of the main f-IRL page.