Simulation Study

Read this page as an oracle-object simulation for feature-matching IRL. The simulation can score reward and counterfactual recovery because the transition law and reward-feature basis are fixed before the panel is generated.

MCE-IRL runs on two synthetic cells. The primary cell is mce_low_high_reward, a compact state-space problem with eight action-dependent reward features. The sanity cell is canonical_low_action. The simulation checks whether the feature-matching route recovers reward and counterfactual behavior when the transition law and reward-feature basis are supplied. Real data cannot answer that question because the reward, policy, value function, Q function, occupancy measure, and counterfactual oracles are not observed.

The numbers come from the simulation harness. In that harness, the transition law, reward features, and reward weights are fixed before generating the panel. The estimator sees the generated demonstrations, the transition law, and the supplied reward features. The reward, policy, value function, Q function, occupancy measure, and counterfactual oracles are held back for evaluation.

The full result generator is mce_irl_run.py. It writes the results file mce_irl_results.json. To rerun it from the repository root:

PYTHONPATH=src:. python validation/estimators/mce_irl/run.py --quiet-progress --enforce-gates

Evidence

MCE-IRL is compared against the full structural and IRL rosters in six simulation studies: bus engine, taxi gridworld, route choice, stockpiling, fleet maintenance, and vehicle scrappage. See the simulation studies index for what each study shows.