Quick Start

This page shows the public GLADIUS wrapper. Use the projected parameters and projection diagnostics for structural reading; the neural Q objects alone are not enough to claim reward recovery.

Sklearn-Style API

GLADIUS accepts a pandas DataFrame directly.

from econirl import GLADIUS
from econirl.core.reward_spec import RewardSpec

model = GLADIUS(
    n_actions=3,
    discount=0.95,
    anchor_action=2,
    q_hidden_dim=128,
    q_num_layers=3,
    ev_hidden_dim=128,
    ev_num_layers=3,
    max_epochs=300,
    patience=50,
    bellman_weight=1.0,
    verbose=False,
)
model.fit(
    df,
    state="state_bin",
    action="action",
    id="individual_id",
    features=reward_spec,      # RewardSpec or feature matrix; omit for policy only
)

print(model.params_)           # projected structural parameters (dict)
print(model.se_)               # projection standard errors
print(model.policy_)           # imitation policy, shape (n_states, n_actions)
print(model.value_)            # soft value function, shape (n_states,)
print(model.projection_r2_)   # R-squared of the action-difference projection

Fitted attributes:

Attribute	Meaning
`params_`	Projected structural parameters by name.
`se_`	Standard errors from the projection OLS.
`coef_`	Parameter vector as a numpy array.
`policy_`	Imitation policy P(a given s), shape (n_states, n_actions).
`value_`	Soft value function V(s), shape (n_states,).
`projection_r2_`	R-squared of the action-difference regression.
`converged_`	Whether early stopping triggered before `max_epochs`.
`n_epochs_`	Number of training epochs completed.

Full Estimator API

GLADIUSEstimator accepts a Panel and an explicit utility, problem, and transition tensor, mirroring the other full estimator APIs.

from econirl.estimation import GLADIUSConfig, GLADIUSEstimator

config = GLADIUSConfig(
    anchor_action=2,
    anchor_rewards=anchor_rewards,     # known reward for anchor action, shape (n_states,)
    anchor_bellman_loss=True,
    anchor_bellman_mode="anchor_moment",
    q_hidden_dim=128,
    q_num_layers=3,
    v_hidden_dim=128,
    v_num_layers=3,
    max_epochs=300,
    patience=50,
    bellman_penalty_weight=1.0,
)
estimator = GLADIUSEstimator(config=config)
summary = estimator.estimate(panel, utility, problem, transitions)

print(summary.parameters)      # projected structural vector
print(summary.policy)          # imitation policy
print(summary.value_function)  # soft value function

Anchor Bellman Modes

# Default: anchor moment pins Q to the continuation-value estimate
config = GLADIUSConfig(anchor_bellman_mode="anchor_moment")

# Paper's bi-conjugate minimax objective
config = GLADIUSConfig(anchor_bellman_mode="paper_minimax")

Use anchor_moment for normal estimation. The paper_minimax mode is kept for comparing against the original paper objective and is not recommended for production.