Generalizing and Categorizing Skills in Reinforcement Learning Agents Using Partial Policy Homomorphisms

摘要

A life long learning agent has to continuously learn, adapt. and perform tasks. However, making real time decisions without control knowledge becomes easily intractable.rnBiological systems encounter similar problems and still learn to perform complex tasks in a dynamic environment. Developmental theories (Lakoff 1987; Mandler 1992) suggest that they do this by learning to abstract information and to form reusable skills and concepts, allowing them to apply this knowledge to more complex tasks.