PPO Clip Fraction Diagnostics

Problem

Compute the fraction of probability ratios outside [1 - clip_eps, 1 + clip_eps], a diagnostic for how often PPO's clipping is active. An empty batch returns 0.0.

Hint

A large clip fraction signals that the policy is moving too fast per update.

Test cases

case 1

in: [[1,1.05,0.9,1.3],0.2]

expected: 0.25

case 2

in: [[1,1.1],0.2]

expected: 0

case 3

in: [[],0.2]

expected: 0

case 4

in: [[0.5,1.5,1],0.5]

expected: 0

Starter codePython

def ppo_clip_fraction(ratios, clip_eps):
    # Your code here
    pass

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