Pre-Estimation Checks

Read this page before fitting Neural MCE-IRL. The main risks are anchor normalization, state encoding, and thin occupancy support for the neural reward map.

Deep MCE-IRL inherits all MCE-IRL data requirements and adds the neural reward normalization problem. Check these before fitting:

Check	Why it matters for Deep MCE-IRL
Transition orientation	Tensors must be row-stochastic in the (n_actions, n_states, n_states) orientation. A wrongly oriented tensor produces silent wrong results.
State encoding alignment	The state encoder output dimension must match the network input dimension. If no encoder is supplied, the default normalizes the state index to [0, 1].
State coverage	The soft Bellman solve covers all states, but occupancy matching uses only observed state-action pairs. Very thin coverage starves the gradient in those regions.
Action support per state	A state where one action is never taken contributes no gradient for that action. The anchor action is exempt - it has no trainable reward - so its absence from a state is safe.
Anchor action validity	The anchor_action index must fall within [0, n_actions) and should correspond to an action with reasonable support. An action that is never taken is a poor anchor.
Anchor consistency	The anchor action or absorbing state must be set consistently between the training run and any downstream comparison. Changing the anchor invalidates reward-map comparisons.
Feature rank (projection only)	If features= is supplied, the feature matrix for the projected cell should be full rank with a low condition number. A rank-deficient projection cannot identify the projected parameters.

Primary Cell Checks

Values from the primary synthetic run (see Simulation Study):

Check	Value	Status
Feature rank	4 / 4	pass
Feature condition number	39.86	pass
Observed states	32 / 32	pass
State-action coverage	1.000	pass
Minimum action share	0.189	pass
Anchor action valid	true	pass

Common Risk Patterns

Anchor mismatch is the most common mistake. If two runs use different anchor actions, or one uses an anchor action and the other uses an absorbing state, the reward matrices use different normalizations and cannot be compared directly. Set the same anchor before any cross-run comparison.

Projection overconfidence is the next most common mistake. The neural reward is projected onto a linear feature basis as a convenience. When the projection R-squared is low or the projected condition number is high (as in the deep_mce_neural_features support cell, where the condition number reaches 478), the projected theta is not a reliable substitute for the reward matrix. Report the reward matrix and behavioral metrics first.

Thin tail coverage compounds with early stopping. If demonstrations cluster in a narrow part of the state space, the occupancy gradient is nearly zero at unvisited states, the network receives no signal there, and early stopping halts before any signal arrives. Use max_epochs generously when coverage is thin.