Simulation Study

Read this page as an oracle-object simulation for the anchored neural reward map. There is no unique parameter vector to compare, so the scorecard is the reward matrix and induced behavior.

Deep MCE-IRL runs on a synthetic cell with a fixed nonlinear neural reward, known stochastic transitions, linear state features, and an anchor action that normalizes the reward. The cell has 32 states, 3 actions, and full state-action coverage, so every recovery claim is checked against the oracle reward matrix, policy, value function, Q function, and counterfactual objects.

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

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

The primary cell results contain no parameter vector. That is by design: a neural reward map does not have a unique structural parameter vector that can be compared across networks, so reward-map recovery and behavioral metrics are the right scorecard. The support cells (deep_mce_neural_features and deep_mce_neural_reward_features) exercise the projected theta path, but projection quality is contingent on the projection being identified.

Evidence

Deep MCE-IRL is compared against the full structural and IRL rosters on the bus engine and taxi gridworld pages. See the simulation studies index for what each study shows.