econirl 0.0.6

Behavioral inference. IRL and DDC with standard errors.

econirl

Structural dynamic discrete choice and inverse reinforcement learning in Python.

EconIRL helps estimate forward-looking choice models, recover reward functions, and evaluate counterfactual policies from panel data.

Documentation: https://econirl.readthedocs.io/

Install

pip install econirl

Quick Start

from econirl.datasets import load_rust_bus
from econirl import NFXP

df = load_rust_bus()

model = NFXP(n_states=90, discount=0.9999, utility="linear_cost")
model.fit(df, state="mileage_bin", action="replaced", id="bus_id")

print(model.params_)

cf = model.counterfactual(RC=4.0)
print(cf.policy[50, 1])

Example output:

{'theta_c': 0.0010028828858836278, 'RC': 3.0722093435989524}
0.05519477716656161

Public Estimator Guides

The public docs describe the supported estimator surface and the current validation scope for each method.

Estimator	Best for	Public evidence scope
NFXP	Exact tabular dynamic discrete choice.	Synthetic tabular validation.
CCP / NPL	Hotz-Miller and NPL-style tabular DDC.	Synthetic tabular validation with support conditions.
MPEC	Constrained likelihood formulation.	Synthetic constrained-likelihood validation.
NNES	Neural value approximation inside NPL.	Low- and high-dimensional synthetic validation.
TD-CCP	Transition-density-free CCP estimation.	Encoded-state finite-theta validation.
MCE-IRL	Maximum causal entropy reward-feature matching.	Supplied-feature simulation validation.
Deep MCE-IRL	Neural reward-map recovery with known transitions.	Anchored neural reward-map validation.
AIRL	Adversarial state-reward recovery under AIRL assumptions.	State-only AIRL validation.
AIRL-Het	Anchored adversarial recovery with latent segments.	Serialized-content simulation validation.
f-IRL	f-divergence state-marginal matching.	State-marginal simulation validation.
GLADIUS	Neural Q and continuation modeling.	Preview diagnostics.
IQ-Learn	Inverse soft-Q learning.	Preview diagnostics.

Package Surface

The recommended API is sklearn-style:

from econirl import NFXP, CCP, MPEC, NNES, TDCCP, MCEIRL

Additional estimators and lower-level configuration objects are available under econirl.estimation, econirl.estimators, and econirl.contrib for advanced workflows.

Repository Layout

• src/econirl/: package source.
• tests/: unit, integration, and validation-evidence tests.
• docs/: public Read the Docs source.
• validation/: reproducible validation runners and machine-readable results.
• examples/: public examples and notebooks.

Manuscripts, PDFs, local research workspaces, and assistant-specific notes are not tracked in this public package repository.

License

MIT