Reward Z-Score Normalization

Problem

Normalize a group of rewards by z-score: (r - mean) / std using the population standard deviation. If the standard deviation is 0, return zeros.

The signature is reward_normalization(rewards).

Hint

Group-relative normalization is what makes GRPO advantages scale-free.

case 1

in: [[1,3]]

expected: [-1,1]

case 2

in: [[5,5]]

expected: [0,0]

case 3

in: [[1,2,3]]

expected: [-1.2247448714,0,1.2247448714]

case 4

in: [[0,10,20]]

expected: [-1.2247448714,0,1.2247448714]

Starter codePython

import math
def reward_normalization(rewards):
    # Your code here
    pass

4 test cases · run your code against them in the editor