Simulation Study

Read this page as an oracle-object simulation for the transition-density-free estimation step. The transition model is withheld from theta estimation but used afterward to score policies, values, Q functions, and counterfactuals.

The TD-CCP simulation study uses the shapeshifter_encoded_state_locally_robust synthetic cell: encoded state features, a finite linear reward, and known oracle objects for every comparison. The simulation asks whether the cross-fitted, locally robust semigradient path recovers the reward parameters, the implied dynamic objects, counterfactual behavior, and valid standard errors.

The result generator is tdccp_run.py. It writes the results file tdccp_results.json.

PYTHONPATH=src:. python validation/estimators/tdccp/run.py --quiet-progress

This is not a raw neural reward-recovery test. It is a structural parameter test with encoded state features and stochastic transitions.

Known transition tensors were not used to estimate theta. They were supplied after fitting so the simulation harness could compare recovered policies, values, Q functions, and counterfactual decisions with oracle solutions.

The standard-error check re-simulates the same encoded-state design 25 times with 300 individuals and 35 periods per replication. Each replication uses cross-fitted, locally robust TD-CCP settings and individual-clustered covariances. The 25-replication run is the current simulation receipt because it is small enough to regenerate routinely. A paper-final audit should rerun with --mc-replications 100 under a CPU budget before claiming final Monte Carlo precision.

The canonical_low_action cell is a simple sanity check. The canonical_high_action cell is a diagnostic stress test that currently fails all checks. The raw neural-reward diagnostic passes 5 of 8 checks and fails reward, value, and Q recovery. It has no finite true reward parameter vector, so it is not part of the primary finite-parameter study.

Evidence

TD-CCP appears on the bus engine page. See the simulation studies index for what each study shows.