When Exploration is Expensive -- Reducing and Bounding the Amount of Experience Needed to Learn to Make Good Decisions



Understanding the limits of how much experience is needed to learn to

make good decisions is both a foundational issue in reinforcement

learning, and has important applications. Indeed, the potential to

have artificial agents that help augment human capabilities, in the

form of automated coaches or teachers, is enormous. Such reinforcement

learning agents must explore in costly domains, since each experience

comes from interacting with a human. I will discuss some of our recent

theoretical results on sample efficient reinforcement learning.