Settling the Reward Hypothesis
This work addresses a foundational theoretical problem in AI and reinforcement learning, with potential implications for how goals are modeled across the field.
The paper tackles the reward hypothesis in reinforcement learning, aiming to fully specify the implicit requirements on goals and purposes under which the hypothesis holds, rather than simply affirming or refuting it.
The reward hypothesis posits that, "all of what we mean by goals and purposes can be well thought of as maximization of the expected value of the cumulative sum of a received scalar signal (reward)." We aim to fully settle this hypothesis. This will not conclude with a simple affirmation or refutation, but rather specify completely the implicit requirements on goals and purposes under which the hypothesis holds.