---

name: usfa_policy_conditioning description: Represent USFA policies with reward-weight encodings z and compute greedy policy actions under those encodings.

USFA Policy Conditioning

Use this skill when a USFA recovery needs to condition successor-feature predictions on a policy encoding z. Do not use it to evaluate a target task over multiple candidates; the GPI skill owns that aggregation.

Inputs

• Candidate policy encodings z.
• A successor-feature table keyed by state, action, and encoding.
• Action set and state.

Outputs

• Canonical candidate encodings.
• psi(s,a,z) lookups.
• Greedy action under an encoded policy: argmax_a psi(s,a,z)^T z.

Workflow

1. Build a deterministic candidate set from training, target, or sampled encodings.
2. Preserve each encoding as the policy identifier.
3. For Bellman bootstrapping, score each action with the same encoding z.
4. Return the greedy action and score table for audit logs.

Validation

Run:

python tests/test_policy_conditioning.py

Limitations

This skill assumes tabular or already-computed successor features. It does not train neural USFA approximators.