econirl 0.0.2

Behavioral inference. IRL and DDC with standard errors.

econirl

Benchmarking dynamic discrete choice and inverse RL algorithms on a variety of MDPs — comparing reward recovery, imitation, and generalization.

Install

pip install econirl

Try It

Load a bundled dataset and fit one estimator in under a second.

from econirl.datasets import load_rust_bus
from econirl.estimation import BehavioralCloningEstimator

df = load_rust_bus()
print(df.head())
print(f"{len(df)} observations, {df['bus_id'].nunique()} buses")

Benchmark

Compare 10 estimators on a simulated 5-state bus engine replacement MDP. This runs each estimator sequentially and takes a few minutes.

from econirl.evaluation.benchmark import BenchmarkDGP, run_single, get_default_estimator_specs

dgp = BenchmarkDGP(n_states=5, discount_factor=0.95)
specs = get_default_estimator_specs()

for spec in specs:
    result = run_single(dgp, spec, n_agents=100, n_periods=50, seed=42)
    print(f"{result.estimator:12s}  {result.pct_optimal:6.1f}%  {result.time_seconds:5.1f}s")

Results

The table below shows all 18 algorithms (10 default from get_default_estimator_specs() plus 8 additional from econirl.contrib).

Estimator	Category	Recovers Params	Recovers Reward	% Optimal	% Transfer	Time
Structural Estimators
NFXP	Structural	Yes	Yes	99.7%	99.8%	13.9s
CCP	Structural	Yes	Yes	99.7%	99.8%	18.6s
SEES	Structural	Yes	Yes	99.6%	99.6%	28.6s
NNES	Structural	Yes	Yes	99.6%	99.1%	13.7s
Entropy-Based IRL
MCE IRL	IRL	Yes	Yes	99.7%	99.7%	20.6s
MaxEnt IRL	IRL	No	Yes	98.2%	97.8%	9.1s
Deep MaxEnt	IRL	No	Yes	98.3%	98.2%	52.3s
BIRL	IRL	No	Yes	99.5%	99.5%	237.8s
Margin-Based IRL
Max Margin	IRL	Yes	Yes	99.3%	99.3%	64.8s
Max Margin IRL	IRL	No	Yes	31.1%	34.2%	0.3s
Distribution Matching
f-IRL	IRL	No	Yes	99.1%	99.1%	44.9s
Neural Estimators
TD-CCP	Neural	Yes	Yes	99.8%	99.7%	16.3s
GLADIUS	Neural	Yes	Yes	99.6%	88.7%	4.2s
Adversarial Methods
GAIL	Adversarial	No	No	54.3%	50.9%	112.9s
AIRL	Adversarial	No	Yes	99.4%	99.5%	123.0s
GCL	Adversarial	No	Yes	92.7%	95.3%	166.5s
Inverse Q-Learning
IQ-Learn	IRL	No	Yes	99.5%	99.1%	0.0s
Baseline
BC	Baseline	No	No	99.5%	99.5%	0.1s

5-state MDP, 100 agents x 50 periods, seed=42. % Optimal = value achieved vs true optimal on training dynamics (baseline-normalized). % Transfer = same metric on held-out transition dynamics (same rewards, different wear rates). Recovers Params = recovers interpretable structural parameters. Recovers Reward = recovers a reward function (enables transfer to new dynamics).

Algorithms

Structural Estimators

Assume the econometrician knows the model and recover flow utility parameters by maximum likelihood.

Algorithm	Paper	Method
NFXP	Rust (1987)	Full-solution MLE via nested fixed point
CCP	Hotz & Miller (1993)	Two-step conditional choice probability with NPL iterations
SEES	Luo & Sang (2024)	Sieve basis V(s) approximation + penalized joint MLE
NNES	Nguyen (2025)	Neural V(s) network (Bellman residual) + structural MLE

Entropy-Based IRL

Recover reward functions from demonstrations using maximum entropy or Bayesian principles.

Algorithm	Paper	Method
MCE IRL	Ziebart (2010)	Maximum causal entropy IRL with soft value iteration
MaxEnt IRL	Ziebart et al. (2008)	Maximum entropy IRL with state visitation frequencies
Deep MaxEnt	Wulfmeier et al. (2016)	Neural reward network + MaxEnt feature matching
BIRL	Ramachandran & Amir (2007)	Bayesian MCMC (Metropolis-Hastings) over reward parameters

Margin-Based IRL

Recover rewards by maximizing the margin between expert and non-expert behavior.

Algorithm	Paper	Method
Max Margin	Ratliff et al. (2006)	Structured max-margin planning
Max Margin IRL	Abbeel & Ng (2004)	Apprenticeship learning via margin maximization

Distribution Matching

Match state-marginal distributions rather than feature expectations.

Algorithm	Paper	Method
f-IRL	Ni et al. (2022)	State-marginal matching via f-divergences (KL, chi-squared, TV)

Neural Estimators

Approximate value functions with neural networks for scalability to large state spaces.

Algorithm	Paper	Method
TD-CCP	Adusumilli & Eckardt (2022)	TD-learning + CCP with neural approximate value iteration
GLADIUS	Kang, Yoganarasimhan & Jain (2025)	Dual Q + EV networks with Bellman consistency penalty

Inverse Q-Learning

Recover reward and policy by learning a single soft Q-function, avoiding adversarial training.

Algorithm	Paper	Method
IQ-Learn	Garg et al. (2021)	Inverse soft-Q learning with chi-squared divergence

Adversarial Methods

Learn reward or policy via a discriminator that distinguishes expert from generated behavior.

Algorithm	Paper	Method
GAIL	Ho & Ermon (2016)	Generative adversarial imitation learning
AIRL	Fu et al. (2018)	Adversarial inverse RL with disentangled reward
GCL	Finn et al. (2016)	Guided cost learning with importance sampling

Baseline

Algorithm	Paper	Method
BC	—	Supervised: empirical P(a|s) from demonstrations

Pseudocode

License

MIT