Reward Hacking Detector

Problem

Detect reward hacking from reward histories.

Return True when the proxy reward increased from first to last observation while the true reward stayed within threshold of its starting value. Fewer than two observations returns False.

Hint

A widening gap between proxy and true reward is the classic hacking signature.

Test cases

case 1

in: [[1,2,5],[3,3.1,2.9],0.5]

expected: true

case 2

in: [[1,2,5],[3,4,6],0.5]

expected: false

case 3

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

expected: false

case 4

in: [[5,4],[1,1],0.5]

expected: false

Starter codePython

def reward_hacking_detected(proxy_rewards, true_rewards, threshold):
    # Your code here
    pass

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