Essay.06-WeinanQian-2100017831
Keywords: goal representation, agent, reinforcement learning
Abstract: Humans possess inherent capabilities to act in alignment with their intentions and showcase an early ability to discern the intentions of others through observation. Harnessing inspiration from this innate quality, the realm of Reinforcement Learning (RL) places significant importance on delving into meaningful goal representations within a computational framework. In the intricate task of designing an agent responsible for observing the environment and making decisions to achieve specific objectives, it becomes imperative to engage in the discourse regarding the apt representations of goals and the potential pathways to actualize these objectives. We introduce two conceivable approaches for representing goals in goal-oriented agents: reward-based goal representations and subtask-based goal representations. Each method of representations comes with its own set of merits and limitations.
Submission Number: 130
Loading