Counterfactuals

Read this page as behavioral counterfactual evidence. The reported regret is about the policy induced by the learned reward, not a structural parameter being re-solved under new primitives.

f-IRL recovers a tabular state reward rather than structural utility parameters, so counterfactual analysis works differently from structural estimators. The recovered reward is not in the same parameterization as the data-generating utility. It cannot be re-solved under a reward shift in the structural sense. Instead, the post-fit policy is used directly for behavioral prediction.

The validation checks whether the recovered policy still performs well in each counterfactual world, using welfare regret as the measure: the loss relative to the policy that would be optimal in that world.

Counterfactual Families

The primary cell checks recovered counterfactual behavior against exact oracle objects for three intervention families:

Family	Intervention	Checked against
Type A	Reward shift (a payoff component changes).	Oracle policy, value, and welfare regret.
Type B	Transition change (the dynamics change).	Oracle policy, value, and welfare regret.
Type C	Action removal (one action is penalized away).	Oracle policy, value, and welfare regret.

Reported Results (Primary Cell)

Counterfactual	Policy TV	Regret	Value RMSE
Type A	0.0102	0.00708	0.00710
Type B	0.0151	0.01236	0.01237
Type C	0.00770	0.00273	0.00277

These results hold in the primary source-paper cell (state marginal matching, state-only reward). They reflect the quality of the recovered policy rather than the ability to re-solve a structural reward under each intervention.

The full record is in f_irl.json. The diagnostic action-dependent cell is not included in counterfactual evidence, because that cell fails the reward-range check and the resulting policy is not a reliable starting point.

Interpretation

Low regret in the primary cell means the soft-optimal policy recovered under the state-marginal objective happens to behave well under the three intervention types tested. This is not the same as structural re-solving: the reward was not updated to reflect the new world, and the policy was not re-derived from first principles. For interventions that substantially change the optimal action distribution, a structural estimator (NFXP, CCP, UFXP) with a linear utility specification will produce more reliable counterfactual predictions.