Reinforcement Learning - An Introduction

Reinforcement Learning is a sub-discipline of machine learning, in which an agent learns from interactions with an environment that provides sparse feedback in form of rewards. The reward encodes what the agent should do, but not how the task should be solved. An example is a dog, which cannot be told directly what it should do. Instead, its behaviour needs to be reinforced through positive and negative feedback. From the perspective of mathematics, reinforcement learning is the problem of finding optimal policies in the context of Markov decision processes.

This lecture introduces the fundamental concepts of reinforcement learning. Programming examples are given, and provided online, whenever they are illustrative. The target audience are students and post-graduates with little or no knowledge about reinforcement learning.

References

• Reinforcement Learning: An Introduction, Sutton & Barto, 1998
• Markov decision processes, Puterman, 2005
• Reinforcement Learning | State-of-the-art, Wiering & van Otterlo (Eds.), 2012

Date and time info
Thursday 14.00 - 15.30

Keywords
MDPs, Dynamic Programming, Bellman Equation, Temporal Difference Learning, Monte Carlo Methods, Q-Learning, Bandit Problem, POMDPs

Audience
MSc students, PhD students, Postdocs