David Krueger–Coordination, Alignment, Academia
The Inside View
00:00
Reward Hacking in MDPs
The paper is basically that's not really possible unless you restrict the set of policies you're optimizing over. And the only like non hackable reward functions are like the trivial one where everything is equal on or something that's equivalent to the original reward function. I guess if I were to talk as a mentioned learning researcher would say like, yes but you're only talking about MDPs or like grid world or something and I feel like if you're talking about like a more joke or environment like the space is so like large that's like it like this like infinite.
Transcript
Play full episode
