Quick Start

This page shows the anchored heterogeneous AIRL path. The exit action and absorbing state are not optional bookkeeping; they are the normalizations that make the segment rewards interpretable.

AIRL-Het is accessed through AIRLHetConfig and AIRLHetEstimator. Both the exit action index and the absorbing state index must be supplied; the estimator raises an error if either is missing.

from econirl.estimation.adversarial.airl_het import AIRLHetConfig, AIRLHetEstimator

config = AIRLHetConfig(
    num_segments=2,
    exit_action=2,       # index of the anchor action (reward = 0)
    absorbing_state=20,  # index of the absorbing state (V = 0)
    reward_type="linear",
    initialization="behavioral_anchor",
    max_em_iterations=30,
    verbose=True,
)

estimator = AIRLHetEstimator(config=config)

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

Key outputs from the returned EstimationSummary:

Attribute	Meaning
summary.parameters	Concatenated segment reward parameters (all segments, flattened).
summary.policy	Mixture policy (prior-weighted average over segments).
summary.value_function	Mixture value function.
summary.converged	Whether the EM loop converged by the tolerance criterion.
summary.num_iterations	Number of EM iterations completed.
summary.log_likelihood	Mixture log-likelihood at the final EM iterate.
summary.metadata["segment_priors"]	Estimated prior probability per segment.
summary.metadata["segment_posteriors"]	Posterior assignment probability per trajectory.
summary.metadata["segment_policies"]	Segment-specific policy arrays.
summary.metadata["segment_reward_matrices"]	Segment-specific reward matrices.
summary.metadata["segment_assignments"]	Hard segment assignment (argmax of posteriors).

Initialization

Two initialization schemes are available:

config = AIRLHetConfig(..., initialization="random")             # random reward start
config = AIRLHetConfig(..., initialization="behavioral_anchor")  # cluster by behavior, invert anchor

behavioral_anchor clusters trajectories by observed action shares, inverts the anchored soft policy into a starting reward for each cluster, and sets the initial posteriors accordingly. It typically converges faster than a random start for the serialized-content setting.

Reward Type

config = AIRLHetConfig(..., reward_type="tabular")   # state-action table (S x A parameters per segment)
config = AIRLHetConfig(..., reward_type="linear")     # linear features (one theta per segment)

The linear mode uses the feature matrix from the supplied utility object. The tabular mode directly learns an entry per state-action pair and is more flexible but requires a larger sample.