Towards Improving Reward Design in RL: A Reward Alignment Metric for RL Practitioners

Calarina Muslimani, Kerrick Johnstonbaugh, Suyog Chandramouli +3

Reinforcement learning agents are fundamentally limited by the quality of the reward functions they learn from, yet reward design is often overlooked under the assumption that a we…