Automated Transfer for Reinforcement Learning Tasks

Abstract

Reinforcement learning applications are hampered by the tabula rasa approach taken by existing techniques. Transfer for reinforcement learning tackles this problem by enabling the reuse of previously learned behaviours. To be fully autonomous a transfer agent has to: (1) automatically choose a relevant source task(s) for a given target, (2) learn about the relation between the tasks, and (3) effectively and efficiently transfer between tasks. Currently, most transfer frameworks require substantial human intervention in at least one of the previous three steps. This discussion paper aims at: (1) positioning various knowledge re-use algorithms as forms of transfer, and (2) arguing the validity and possibility of autonomous transfer by detailing potential solutions to the above three steps.