Simulation Study

Read this page as an oracle-object simulation for the anchored projected-reward route. Small regret is reported, but the reward checks determine whether the learned objects support structural interpretation.

GLADIUS is evaluated on two related high-dimensional-state synthetic cells. Each cell has 21 discrete states paired with a high-dimensional encoded state-feature vector, a low-dimensional action-dependent reward basis, and an anchor action whose rewards are known. All recovery claims are compared against the truth from the data-generating process.

The full result generator is run.py. It writes two results files: gladius.json (primary cell) and gladius_scaled.json (scaled cell).

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

The primary and scaled cells share the same structural support: full feature rank, full state-action coverage, and a valid anchor. The raw Bellman reward and value NRMSE fail their checks on both cells. Projected reward, policy, and Q recover reasonably. The absolute reward level implied by the raw Bellman equation does not. Counterfactual regret is small, but low regret alone is not sufficient for structural validity when the upstream reward checks fail.

Evidence

GLADIUS is compared against the full structural and IRL rosters on the bus engine, taxi gridworld, and direct optimization pages. GLADIUS also appears on the route choice, stockpiling, and fleet maintenance pages. See the simulation studies index for what each study shows.