Quick Start

This page shows the wrapper path for NNES. The network shown here approximates the continuation value; the reported reward parameters remain finite structural parameters.

The public wrapper accepts dataframe-style panel data. This quick start is the package smoke path: create a model, call fit, and read the fitted sklearn-style attributes. The full NNES simulation study lives in the simulation results file linked from Simulation Study.

from econirl.datasets import load_rust_bus
from econirl import NNES

df = load_rust_bus()

model = NNES(
    n_states=90,
    n_actions=2,
    discount=0.9999,
    utility="linear_cost",
    bellman="npl",
    hidden_dim=32,
    num_layers=2,
    v_epochs=500,
    n_outer_iterations=3,
)
model.fit(df, state="mileage_bin", action="replaced", id="bus_id")

print(model.params_)
print(model.se_)
print(model.policy_.shape)
print(model.v_network_.shape)

Output from the package smoke run:

{'theta_c': 0.001034, 'RC': 3.073617}
{'theta_c': 0.000383, 'RC': 0.077527}
(90, 2)
(90,)

The fitted wrapper exposes structural parameter estimates, standard errors, the estimated policy, the value function, the first-stage transition estimate, and the trained value-network predictions over states.

Attribute	Meaning
`params_`	Estimated structural reward parameters.
`se_`	Standard errors for reward parameters.
`policy_`	Estimated action probabilities by state.
`value_`	Estimated value function by state.
`v_network_`	Neural value-network predictions by state.
`transitions_`	First-stage transition matrix used by the wrapper.

Use bellman="npl" for the reported NNES path. The wrapper also accepts bellman="nfxp" as a neural soft-Bellman diagnostic variant.

Full Estimator API

Use econirl.estimation.nnes.NNESEstimator when the model already has package-native objects or when the simulation study needs exact control over the DDC problem, transition tensor, reward specification, and NNES training configuration.

from econirl.estimation.nnes import NNESConfig, NNESEstimator

estimator = NNESEstimator(
    config=NNESConfig(
        hidden_dim=32,
        num_layers=2,
        v_epochs=500,
        n_outer_iterations=3,
        compute_se=True,
    )
)
summary = estimator.estimate(panel, utility, problem, transitions)

print(summary.parameters)
print(summary.metadata["v_loss_per_outer"][-1])

No fixed output is shown for the estimator-API snippet because panel, utility, problem, and transitions are caller-supplied objects.

The synthetic data simulation uses the full estimator API because it supplies the true DDC problem, transition tensor, and reward specification directly. That is also the route to use when reproducing paper-style designs with custom state encodings or supplied transition laws.