Simulation Study

Read this page as an oracle-object simulation for anchored latent heterogeneity. The simulation checks both reward recovery within segment and assignment of trajectories to segments.

AIRL-Het runs on a synthetic serialized-content heterogeneous cell with two latent segments, repeated books per user, and three actions (read/wait/exit). The cell has known segment-level rewards, transitions, policies, values, Q functions, and Type A, Type B, and Type C counterfactual oracle objects, so every recovery metric is checked against the truth.

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

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

The Type B regret is close to its threshold. This is expected for the heterogeneous transition-change counterfactual: each segment’s reward must adapt to a different new dynamic, and the recovered reward carries more approximation error under that intervention than under reward shifts or action removals.

Evidence

Cross-estimator simulation-study coverage for AIRL-Het is planned. See the simulation studies index for the current study roster.