---

name: rnd_normalization_streams description: Maintain RND observation whitening/clipping and intrinsic reward scaling statistics.

RND Normalization Streams

Use this skill when preparing observations or intrinsic rewards for RND. Do not use it to normalize policy-network observations unless the experiment explicitly chooses that; the paper normalizes predictor/target inputs.

Inputs

• Numeric observation batches.
• Intrinsic reward or intrinsic-return batches.
• Epsilon and clip bounds.

Outputs

• Running mean, variance, and count.
• Whitened and clipped observations.
• Reward values scaled by running return standard deviation.

Workflow

1. Initialize running statistics from early random-agent or synthetic preflight observations.
2. Update statistics with each new batch using a numerically stable batch merge.
3. Normalize observations with (x - mean) / std and clip to [-5, 5] by default.
4. Scale intrinsic rewards by reward/return standard deviation with epsilon protection.

Validation

Run python tests/test_normalization.py from this skill directory.

Limitations

This skill provides deterministic numeric helpers, not a distributed RL statistics service.