Counterfactuals

Read this page as policy-regret evidence, not structural re-solving. AIRL’s state-only recovered reward is not used here like an NFXP-style parameter vector.

AIRL does not recover a reward in the same parameterization as the data-generating process. The recovered reward component \(g_\theta\) is identified only up to potential-based shaping, so raw parameter values cannot be re-solved under an arbitrary intervention the way NFXP or MCE-IRL can. The package evaluates AIRL counterfactuals through the simulation harness, which fixes the intervention oracle-side and measures the welfare cost of deploying the recovered policy in the changed world.

Counterfactual Families

Type	Intervention	How AIRL is evaluated
Type A	Shift payoffs, hold transitions fixed.	Re-solve the oracle problem under the new payoffs; report regret from the old AIRL policy.
Type B	Change transitions, hold the recovered reward fixed.	Re-solve the oracle under new transitions; report regret from the old AIRL policy.
Type C	Disable one non-anchor action.	Re-solve the oracle with the penalized action; report regret from the old AIRL policy.

Because AIRL’s recovered reward is state-only and on its own scale, it is not re-solved under each intervention. The harness instead treats the recovered policy as fixed and measures how much welfare it loses in the counterfactual world relative to the oracle policy that was re-solved under the intervention.

Reported Results

From the primary airl_paper_identification simulation:

Counterfactual	Policy TV	Regret
Type A	0.0062	0.0029
Type B	0.0076	0.0038
Type C	0.0075	0.0050

Regret values below 0.01 are in the range of good IRL methods on this synthetic cell. They reflect a nearly correct policy in the base world; the recovered policy adapts reasonably to the counterfactual world even without re-solving, because the oracle policy change is also small.

API Boundary

The stable public objects after fitting are summary.policy, summary.parameters, and the value function on the result object. For controlled payoff, transition, or action-set interventions, use the package’s simulation and evaluation utilities with an explicit problem and transition environment. The AIRLEstimator does not expose a .counterfactual() method; the counterfactual evaluation in the results file comes from the simulation study runner.