Quick Start

This page shows the feature-based MCE-IRL workflow. The reward that comes out is only as interpretable as the supplied feature matrix, transition tensor, and normalization.

The wrapper follows the sklearn-style pattern: build an estimator, call fit, then read fitted attributes. For multi-action MCE-IRL, provide reward features explicitly.

import numpy as np

from econirl.datasets import load_rust_bus
from econirl.estimators import MCEIRL

n_states = 90
n_actions = 2
features = np.zeros((n_states, n_actions, 2))
features[:, 0, 0] = -np.arange(n_states) / 100.0
features[:, 1, 1] = -1.0

df = load_rust_bus()

model = MCEIRL(
    n_states=n_states,
    n_actions=n_actions,
    discount=0.99,
    feature_matrix=features,
    feature_names=["keep_mileage_cost", "replace_cost"],
)
model.fit(df, state="mileage_bin", action="replaced", id="bus_id")

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

For a problem other than the Rust bus, pass the dynamics explicitly. Supply a transition tensor of shape (n_actions, n_states, n_states) and the observed next-state column.

model.fit(
    df,
    state="state",
    action="action",
    id="id",
    next_state="next_state",
    transitions=transitions,
)

transitions=None estimates only the two-action Rust-bus keep/replace kernel. A model with more than two actions requires an explicit tensor. Build one from observed transitions with estimate_empirical_transitions(panel, n_actions, n_states) from econirl.estimators.

Estimates depend on the supplied reward features, transition specification, and inference settings, so no canonical output is shown here. The fitted estimator exposes reward parameters, standard errors when requested, the recovered reward, the policy, the value function, and feature-matching diagnostics.

Attribute	Meaning
params_	Estimated reward parameters.
se_	Standard errors for the reward parameters when available.
reward_matrix_	Structural reward matrix by state and action.
policy_	Estimated action probabilities by state.
value_	Estimated value function by state.
log_likelihood_	Log likelihood of the demonstrations under the recovered policy.

Simulation Rerun

To reproduce the simulation, run the validation script:

PYTHONPATH=src:. python validation/estimators/mce_irl/run.py --quiet-progress --enforce-gates

The command writes the results file and reports the pass/fail summary for the two simulation cells.

Use econirl.estimation.mce_irl.MCEIRLEstimator when you need direct control over Panel objects, utility objects, DDCProblem, transition tensors, the root feature-matching optimizer, or standard-error computation.