2MMS50 - Stochastic Decision Theory

Aims / Summary

The goal of the course is to familiarize students with the mathematical concepts and computational techniques for stochastic decision and optimization problems, and illustrate the application of these methods in various scenarios. The methodological framework of Markov decision processes and stochastic dynamic programming models will play a central role, and the students are expected to obtain knowledge of the main problem formulations and be able to apply the the main computational approaches in that domain to stylized problems.

You will learn of state-of-the-art approaches in:

Markov Decision Theory
Multi-Armed Bandits
Reinforcement Learning
Stochastic / Distributed / Robust Optimization

Lecturers

This course will be taught by prof.dr. Bert Zwart (TU/e, CWI) and dr. Jaron Sanders (TU/e, personal).

Course material

We will be using the following material throughout the course:

Lecture notes Advanced Stochastic Operations Research: Stochastic Decision Theory, 2009
Sebastian Bubeck and Nicolo Cesa-Bianchi’s lecture notes Regret Analysis of Stochastic and Nonstochastic Multi-armed Bandit Problems, 2012 (Links to an external site.)

If you want additional background, we recommend you these books:

Richard Sutton and Andrew Barto, Reinforcement Learning: An introduction
Dimitri Bertsekas and John Tsitsiklis book Neuro-dynamic programming, 1996

We encourage you to examine these seminal papers, whose contributions will come by:

William Thompson, On the likelihood that one unknown probability exceeds another in view of the evidence of two samples, Biometrika 25(3-4):285-294, 1933
Robbins, Some aspects of the sequential design of experiments, Bulletin of the American Mathematical Society. 58(5):527-535, 1952
Gittins, Bandit Processes and Dynamic Allocation Indices, Journal of the Royal Statistical Society, 1979

Exam / Grading

In order to pass this course, you must hand in four homework assignments. We ask you to partner up and deliver your work in student groups of size two.Each homework assignment will cover one of the topics of the course, and we will hand one out after every 3rd lecture. You will then have two weeks to complete each assignment.

We will grade each of the four assignments. Each will count for 25% towards your final grade.

Online lectures via Canvas Conferences

All lectures and instructions will be taught digitally, using Canvas Conferences. We will be using Powerpoint slides predominantly, and at times, may use Microsoft Whiteboard. Be on time and present, digitally, and participate. We encourage you strongly to enable your camera and have a functional microphone for interactivity.

Canvas Conferences can be found in the menu to your left on our course page on canvas.tue.nl.

Course Overview

Here is a week-by-week breakdown of the course:

Week	Topic	Professor	Activity	Date
17	Markov Decision Theory	Bert Zwart	two lectures	April 22, 24
18	Markov Decision Theory New assignment: MDT	Bert Zwart	one lecture, one instruction	April 29, May 1
19	Multi-Armed Bandits	Jaron Sanders	two lectures	May 4, 8
20	Multi-Armed Bandits *Deadline: MDT, May 14th* New assignment: MABs, May 15th	Jaron Sanders	one lecture, one instruction	May 13, 15
21			no class activities
22	Reinforcement Learning	Jaron Sanders	two lectures	May 27, 29
23	Reinforcement Learning *Deadline: MABs, June 4th* New assignment: RL	Jaron Sanders	one lecture, one instruction	June 3, 5
24	Stochastic / Distributed / Robust Optimization	Bert Zwart	two lectures	June 10, 12
25	Stochastic / Distributed / Robust Optimization Deadline: RL, June 18th New assignment: SDRO	Bert Zwart	one lecture, one instruction	June 17, 19
26			no class activities
27	Deadline: SDRO, July 2nd