On The Fragility of Learned Reward Functions

Lev McKinney, Yawen Duan, David Krueger +1

Reward functions are notoriously difficult to specify, especially for tasks with complex goals. Reward learning approaches attempt to infer reward functions from human feedback and…