Streaming Reward Normalization

Problem

Normalize a list of rewards by their own mean and population standard deviation: (r - mean) / (std + epsilon). If the standard deviation is zero return zeros, and an empty list returns [].

Hint

Reward normalization keeps the advantage scale stable as training progresses.

Test cases

case 1

in: [[1,2,3],0]

expected: [-1.224744871391589,0,1.224744871391589]

case 2

in: [[5],0]

expected: [0]

case 3

in: [[],0]

expected: []

case 4

in: [[2,2,2],1e-8]

expected: [0,0,0]

Starter codePython

import math
def streaming_reward_normalize(rewards, epsilon):
    # Your code here
    pass

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