Recent advances in leveraging human guidance for sequential decision-making tasks

A longstanding goal of artificial intelligence is to create agents capable learning perform tasks that require sequential decision making. Importantly, while it the agent learns and acts, still up humans specify particular task be performed. Classical task-specification approaches typically involve providing stationary reward functions or explicit demonstrations desired tasks. However, there has recently been a great deal research energy invested in exploring alternative ways which may guide may, e.g., more suitable for certain less human effort. This survey provides high-level overview five recent machine frameworks primarily rely on guidance apart from pre-specified conventional, step-by-step action demonstrations. We review motivation, assumptions, implementation each framework, we discuss possible future directions.