Simulation Study

Read this page as a behavioral-validation page. The primary cell is the state-marginal matching setting; the action-dependent cell is a negative control, not structural evidence.

f-IRL is reported on two synthetic cells. The primary cell is the source-paper state-marginal configuration: a fully specified data-generating process with known reward, transitions, policy, value, Q function, and Type A, B, and C counterfactual oracle objects. The action-dependent cell is a diagnostic negative control that fails the reward-range check and is not treated as structural recovery evidence.

The full result generator is run.py. It writes the results file f_irl.json.

cd /path/to/econirl
PYTHONPATH=src:. python validation/estimators/f_irl/run.py

The action-dependent cell (canonical_low_action) fails the reward-range check with a reward range of 0.000, indicating a flat-reward solution. This cell is retained as a diagnostic to show that f-IRL with action-dependent marginal matching does not recover a non-trivial reward on the standard DDC benchmark. It is not structural recovery evidence.

Evidence

f-IRL appears on the taxi gridworld page. See the simulation studies index for what each study shows.