CS 6789: Foundations of Reinforcement Learning

Modern Artificial Intelligent (AI) systems often need the ability to make sequential decisions in an unknown, uncertain, possibly hostile environment, by actively interacting with the environment to collect relevant data. Reinforcement Learning (RL) is a general framework that can capture the interactive learning setting and has been used to design intelligent agents that achieve super-human level performances on challenging tasks such as Go, computer games, and robotics manipulation.

This graduate level course focuses on theoretical and algorithmic foundations of Reinforcement Learning. The four main themes of the course are (1) provably efficient exploration, (2) policy optimization (especially policy gradient), (3) control, and (4) imitation learning.

After taking this course, students will be able to understand both classic and state-of-art provably correct RL algorithms and their analysis. Students will be able to conduct research on RL related topics.


Instructors: Wen Sun (Cornell) and Sham Kakade (University of Washington)

TAs: Jonathan Chang

Lecture time: Tuesday/Thursday 3-4:15pm ET

Office hours: By Appointment

Contact: cornellcs6789@gmail.com.

Please communicate to the instructors and TA only through this account. Emails not sent to this list, with regards to the course, will not be responded to in a timely manner.

Zoom Information

Zoom information has been posted on Piazza. If you are not enrolled/wait listed (or you are not from Cornell), but want to have access, please email cornellcs6789@gmail.com to ask for permission. We will make a decision based on the capacity of the class and your research background (please in email briefly describe your research interestes and your background on machine learning theory. Thanks).


This is an advanced and theory-heavy course: there is no programming assignment and students are required to work on a theory-focused course project.

Students need strong grasp on Machine Learning (e.g., CS 4780), Probability and Statistics (e.g., BTRY 3080 or ECON 3130, or MATH 4710), Optimization (e.g., ORIE 3300), and Linear Algebra (e.g., MATH 2940). For undergraduate students enrollment: permission of instructor with minimum grade A in CS 4780.

Grading Policies

Assignments 55% (HW0:10%, HW1-HW3: 15% each) and Project 45%

All homework will be mathematical in nature, focussing on the theory of RL and bandits; there will not be a programming component. The entire HW must be submitted in one single typed pdf document (not handwritten). HW0 is MANDATORY to pass to satisfactory level; it is to check your knowledge of the prerequisites in probability, statistics, and linear algebra.

Homework Rules: Homework must be done individually: each student must understand, write, and hand in their own answers. It is acceptable for students to discuss problems with each other; it is not acceptable for students to look at another students written answers. You must also indicate on each homework with whom you collaborated with and what online resources you used.

Late days: Homeworks must be submitted by the posted due date. You are allowed up to 5 total LATE DAYs for the homeworks throughout the entire semester. These will be automatically deducted if your assignment is late. For example, any day in which an assignment is late by up to 24 hours, then one late day will be used (up to two late days). After your late days are used up, late penalties will be applied: any assignment turned in late will incur a reduction in score by 33% for each late day, so if an assignment is up to 24 hours late, it incurs a penalty of 33%. Else if it is up to 48 hours late, it incurs a penalty of 66%. And any longer, it will receive no credit. We will track all your late days and any deductions will be applied in computing the final grades. If you are unable to turn in HWs on time, aside from permitted days, then do not enroll in the course.

Course Project

Please see the course project page.

Diversity in STEM

While many academic disciplines have historically been dominated by one cross section of society, the study of and participation in STEM disciplines is a joy that the instructors hope that everyone can pursue, regardless of their socio-economic background, race, gender, etc. The instructors encourage students to both be mindful of these issues, and, in good faith, try to take steps to fix them. You are the next generation here.

Course Notes: RL Theory and Algorithms

The course will be largely based of the working draft of the monograph "Reinforcement Learning Theory and Algorithms", available here. We will be updating these notes in V2 to use unnormalized rewards. Please be aware of this difference if you read ahead and look at V1 of the draft. If you find typos or errors, please let us know. We would appreciate it!

Honor Code

Cornell University Code of Academic Integrity, CS Department Code of Academic Integrity.

Schedule (tentative)

09/3/20 Fundamentals: Markov Decision Processes Reading: Ch.1 Slides, Annotated slides, HW0
09/08/20 Fundamentals: Policy Iteration and Value Iteration Reading: Ch.1 Slides, Annotated slides
09/10/20 Fundamentals: Computational Complexity & The LP-Formulation Reading: Ch.1 Slides, Annotated slides
09/15/20 Fundamentals: Statistical Limits Reading: Ch.2 Slides, Annotated slides
09/17/20 Exploration: Multi-Armed Bandit Reading: Ch.1 Slides, Annotated slides , Guest Lecturer: Thodoris Lykouris, HW1
09/22/20 Exploration: Efficient Exploration in Tabular MDPs Note Slides, Annotated slides
09/24/20 Fundamentals: Generalization in RL Reading: Ch.3 Slides, Annotated slides
09/29/20 Exploration: Linear Bandit Paper
10/1/20 Exploration: Efficient Exploration in Linear MDPs Paper
10/6/20 Exploration: Learning in Large Scale MDPs (Bellman Rank/Witness Rank) Bellman Rank, Witness Rank HW1 Due (10/2/20)
10/8/20 Policy Optimization: Policy Gradient (REINFORCE, Actor-Critic, Variance Reduction)
10/13/20 Policy Optimization: Global Convergence of PG
10/15/20 Policy Optimization: Natural Policy Gradient (NPG) and its Global Convergence
10/20/20 Policy Optimization: NPG (Continued) HW2 Due
10/22/20 Policy Optimization: Conservative PI and Trust Region Methods CPI, Covariant Policy Search HW3
10/27/20 Control: LQR/LQG
10/29/20 Control: LQR/LQG
11/3/20 Control: Kernelized Nonlinear Regulator (KNR) LC^3 for KNR
11/5/20 Batch RL: Fitted Q Iteration and Recent Advances

HW3 Due
Guest Lecturer: Akshay Krishnamurthy

11/10/20 Imitation Learning: Behavior Cloning
11/12/20 Imitation Learning: Learning from Observations Alone FAIL
11/17/20 Imitation Learning: Interactive Setting (AggreVaTe)
11/19/20 Tentative
11/24/20 Student Project Presentations
12/01/20 Student Project Presentations
12/03/20 Student Project Presentations
12/08/20 No Class: NeurIPS
12/10/20 No Class: NeurIPS Project Report Due
12/15/20 No Class