Simulation Study

Read this page as an oracle-object simulation. The simulation can test reward, policy, value, Q, and counterfactual recovery because the data-generating objects are known before estimation.

NFXP runs on the canonical_low_action synthetic cell. The cell has known rewards, transitions, policies, values, Q functions, and Type A, Type B, and Type C counterfactual oracles, so every recovery claim is checked against the truth. Real bus data cannot answer that question because the true reward, value function, policy, and counterfactual oracles are not observed.

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

PYTHONPATH=src:. python validation/estimators/nfxp/run.py

The synthetic cell has action-dependent reward features and an exit action that anchors the reward level. The estimates differ from truth because the panel is finite.

Evidence

NFXP 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.