Quick Start

This page shows the tabular AIRL estimator path. Keep reward_arg="state" when you want the original AIRL reward-recovery guarantee; state-action variants need the identification-boundary caveat.

The low-level AIRLEstimator accepts a Panel, a utility specification, a DDCProblem, and a transition tensor. The high-level econirl.estimators.AIRL alias points to the neural wrapper; use AIRLEstimator for tabular work.

from econirl.estimation import AIRLConfig
from econirl.estimation.adversarial import AIRLEstimator

config = AIRLConfig(
    reward_type="linear",
    reward_arg="state",
    use_shaping=True,
    max_rounds=150,
    reward_lr=0.01,
    discriminator_steps=10,
)
estimator = AIRLEstimator(config=config)

summary = estimator.estimate(
    panel=panel,
    utility=utility,
    problem=problem,
    transitions=transitions,
)

print(summary.policy)       # shape (n_states, n_actions)
print(summary.converged)    # True if policy converged before max_rounds

Key fitted objects:

Attribute	Meaning
`summary.policy`	Recovered choice probabilities, shape (n_states, n_actions).
`summary.parameters`	Learned reward parameters (linear) or reward matrix entries (tabular).
`summary.converged`	Whether policy change fell below `convergence_tol` before `max_rounds`.
`summary.num_iterations`	Training rounds used.
`summary.metadata["final_disc_loss"]`	Final discriminator loss; near log(2) at the adversarial optimum.

Reward Argument

# State-only: matches Fu et al. (2018) identification guarantee
config = AIRLConfig(reward_arg="state", ...)

# State-action: needed when the DGP has action-dependent payoffs
# Longer schedule required; increase max_rounds and reward_lr
config = AIRLConfig(
    reward_arg="state_action",
    max_rounds=300,
    reward_lr=0.01,
    discriminator_steps=10,
)

The state-only mode projects the learned reward matrix onto the state-only subspace by averaging across actions before computing the discriminator logit. The state-action mode lets the reward differ by action, which is needed for discrete choice problems with action-dependent payoffs but loses the disentanglement guarantee.

Reproduce the Simulation

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