Reinforcement Learning: Theory and Algorithms

ECE59500

Credit Hours:

3

Learning Objective:

A student who successfully fulfills the course requirements will have demonstrated an ability to:

  • Explain different problem formulations for reinforcement learning
  • Apply various algorithmic solutions to a wide range of sequential decision-making problems
  • Analyze the performance capabilities and limitations of different algorithm for sequential decision making
  • An ability to conduct a research project by collaborating with one or more partners and write a scientific report of the research findings

Description:

This course introduces the foundations and he recent advances of reinforcement learning, an area of machine learning closely tied to optimal control that studies sequential decision-making under uncertainty. This course aims to create a deep understanding of the theoretical and algorithmic foundations of reinforcement learning while discussing the practical considerations and various extensions of reinforcement learning.  

Topics Covered:

Week Lecture Topics
1 Introduction, motivation, overview of relevant background
2 Dynamic programming and policy evaluation
3 Policy iteration and value iteration
4 Monte Carlo and temporal difference methods
5 Computational complexity and statistical limits
6 Linear quadratic regulators (LQR) and optimal control
7 Optimal control for nonlinear systems (Iterative LQR)
8 Prediction, estimation, and Kalman filtering
9 Model-based and model-free reinforcement learning
10 Approximate policy iteration and deep Q-learning
11 Conservative policy iteration and trust region methods
12 Stochastic gradient descent and policy gradient
13 Exploration in reinforcement learning and multi-armed bandits
14 Partially observable Markov decision processes and risk-averse reinforcement learning
15 Inverse reinforcement learning, meta-learning, transfer learning, and multi-agent reinforcement learning

 

Prerequisites:

Undergraduate understanding of linear algebra, probability, calculus 

Web Address:

https://purdue.brightspace.com

Textbooks:

Required:

None

Recommended:

  1. Bandit Algorithms, Lattimore, Tor; Szepesvari, Csaba, Cambridge University Press, 2020
  2. Dynamic Programming and Optimal Control, Bertsekas, Dimitri P., Athena Scientific, 2011
  3. Foundations of Deep Reinforcement Learning, Graesser, Laura; Keng, Wah Loon, Addision-Wesley Professional, 2019
  4. Markov Decision Processes: Discrete Stochastic Dynamic Programming, Puterman, Martin L., John Wiley & Sons, 2014
  5. Neuro-dynamic Programming, Bertsekas, Dimitri P.; Tsisiklis, John N., Athena Scientific, 1996
  6. Reinforcement Learning: An Introduction, Sutton, Richard S.; Barto, Andrew G., MIT Press, 2018
  7. Reinforcement Learning: Theory and Algorithms, Agarwal, Alekh; Jiang, Nan; Kakade, Sham M.; Sun, Wen, 2019 

ProEd Minimum Requirements:

view