A Brief Tutorial on Reinforcement Learning

par/by Doina Precup
School of Computer Science
University McGill

Reinforcement learning is an approach for learning from interaction with an environment. Instead of being explicitly taught what they should do, reinforcement learning agents learn how to act by observing the consequences of their actions. In this tutorial, I will briefly review the mathematical framework of Markov Decision Processes (MDPs), on which most reinforcement learning methods rely. I will then show how temporal-difference learning methods can be derived from standard MDP solution methods. I will discuss some of the current key problems in reinforcement learning: the use of function approximation, dealing with environments which are partially observable, and approaches for scaling up reinforcement learning for large tasks.