Toward Complex and Structured Goals in Reinforcement Learning

Goals play a central role in the study of agentic behavior. But what is a goal, and how should we best represent them? The traditional reinforcement learning answer is that all goals are expressible as the maximization of future rewards. While parsimonious, such a definition seems insufficient when viewed from both the perspective of humans specifying goals to machines and autotelic agents that self-propose tasks to explore and learn. We offer a critical perspective on the distillation of all goals directly into reward functions. We identify key features we believe goal representations ought to have, and then offer a proposal we believe meets those considerations