Settling the Reward Hypothesis

TL;DR

This paper rigorously analyzes the reward hypothesis, clarifying the conditions under which goals and purposes can be represented as reward maximization in reinforcement learning.

Contribution

It provides a comprehensive framework specifying the implicit requirements for the reward hypothesis to hold, moving beyond simple affirmation or refutation.

Findings

Defines explicit conditions for the reward hypothesis validity

Clarifies the relationship between goals and reward signals

Establishes a formal framework for analyzing goal representation

Abstract

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.