Essay.06-FeiyangXie-2100017837

Feiyang Xie

Keywords: hierarchical reinforcement learning; goal representation; decision transformer

Abstract: Goals can be expressed in many ways. For humans, in daily life, goals can be expressed as short sentences or specific future states. As external manifestations of beliefs and desires, goals can guide action selection and are thus critical for planning problems. In reinforcement learning (RL), goal representation has been an active research area. Effective goal representations can enable agents to efficiently solve tasks. However, current goal representation research often focuses on particular tasks and environments, lacking generalizability and versatility. This essay will first explain goal usage in hierarchical RL. Then, it will introduce goal representation methods from simple to complex, discussing their advantages and disadvantages. Finally, it will consider more general goal representations given modern pretrained models.

Submission Number: 140