Quick Start

This page shows the RHIP wrapper and the horizon knob. Treat horizon as part of the behavioral model: it changes the planning depth being matched.

The wrapper follows the sklearn-style pattern: build an estimator, call fit, then read fitted attributes. RHIP requires a pre-built Panel, an action-dependent feature array, and a transition tensor.

import numpy as np

from econirl import RHIP
from econirl.core.types import Panel

model = RHIP(
    horizon=3,
    discount=0.95,
    scale=1.0,
    feature_names=["edge_cost", "amenity", "goal"],
)
model.fit(panel, features=features, transitions=transitions)

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

features must be a 3D array of shape (n_states, n_actions, n_features) and transitions a 3D tensor of shape (n_actions, n_states, n_states). The fitted estimator exposes reward parameters, standard errors when requested, the recovered reward matrix, the policy, and the value function.

Attribute	Meaning
`params_`	Estimated reward parameters.
`se_`	Standard errors for the reward parameters when `compute_se=True`.
`reward_matrix_`	Reward matrix by state and action, shape `(n_states, n_actions)`.
`policy_`	Estimated action probabilities by state.
`value_`	Estimated value function by state.
`converged_`	Whether the gradient norm or occupancy moment crossed tolerance.

The Horizon Knob

The horizon parameter spans a family of behavioral models:

`horizon`	Behavioral model
`float("inf")` or `None`	Max Causal Entropy IRL (stochastic, the MCE-IRL endpoint)
`0`	Max-Margin-Planning-like (deterministic, cheapest)
`1`, `3`, …	Receding-horizon interpolation between the two ends

The H = inf endpoint is MCE-IRL, so the result is identical to MCE-IRL by construction. Finite horizons run H soft-Bellman backups on top of a deterministic continuation value.

from econirl import RHIP

mce_end = RHIP(horizon=float("inf"))          # the MCE-IRL endpoint
mce_end.fit(panel, features=features, transitions=transitions)

det_end = RHIP(horizon=0)                     # the deterministic end
det_end.fit(panel, features=features, transitions=transitions)

Advanced Estimator API

Use econirl.estimators.rhip.RHIPEstimator when you need direct control over Panel objects, DDCProblem, RHIPConfig, or transition tensors as JAX arrays. RHIPEstimator accepts a RHIPConfig that exposes Adam optimizer settings, deterministic tail parameters, and bootstrap SE controls.