---

name: maximum_entropy_objective description: Compute SAC maximum entropy objective diagnostics for rewards, log probabilities, discounts, and temperature-scaled entropy bonuses.

Maximum Entropy Objective

Use this skill when a recovery or implementation must verify the SAC maximum-entropy objective from Haarnoja et al. The skill is appropriate for bounded checks, synthetic replay batches, and cross-checks of entropy bonus signs. Do not use it to claim full MuJoCo performance.

Inputs

• A reward sequence.
• Matching policy log probabilities.
• Discount factor gamma.
• Temperature or reward-scale coefficient alpha.

Outputs

• Discounted reward-only return.
• Discounted soft return using reward - alpha * log_prob.
• Per-step entropy bonuses.

Workflow

1. Validate reward and log-probability lengths.
2. Convert entropy to the sign convention -log_prob.
3. Add alpha * entropy to reward at each step.
4. Compute discounted totals for reward-only and soft objectives.
5. Use the difference as a mechanism diagnostic, not as a task score.

Validation

Run python tests/test_entropy_objective.py or validate the whole tree with the Distiller skill validator.

Limitations

This module computes objective diagnostics only. It does not implement replay sampling, neural critics, or environment interaction.