CS394R: Reinforcement Learning: Theory and Practice -- Fall 2016

CS394R: Reinforcement Learning: Theory and Practice -- Fall 2016

Instructor: Peter Stone
Department of Computer Science

Tuesday, Thursday 9:30-11:00am
GDC 2.410


Jump to the assignments page.
Jump to the resources page.
Jump to the textbook page.

Please complete the midterm course evaluation survey.

Instructor Contact Information

office hours: TBA and by appointment
office: GDC 3.508
phone: 471-9796
fax: 471-8885
email: pstone@cs.utexas.edu


Teaching Assistant

Sanmit Narvekar
office hours: TBA and by appointment
office: GDC 3.424E
email: sanmit@cs.utexas.edu


Course Description

"The idea that we learn by interacting with our environment is probably the first to occur to us when we think about the nature of learning. When an infant plays, waves its arms, or looks about, it has no explicit teacher, but it does have a direct sensori-motor connection to its environment. Exercising this connection produces a wealth of information about cause and effect, about the consequences of actions, and about what to do in order to achieve goals. Throughout our lives, such interactions are undoubtedly a major source of knowledge about our environment and ourselves. Whether we are learning to drive a car or to hold a conversation, we are all acutely aware of how our environment responds to what we do, and we seek to influence what happens through our behavior. Learning from interaction is a foundational idea underlying nearly all theories of learning and intelligence."

"Reinforcement learning problems involve learning what to do --- how to map situations to actions --- so as to maximize a numerical reward signal. In an essential way these are closed-loop problems because the learning system's actions in uence its later inputs. Moreover, the learner is not told which actions to take, as in many forms of machine learning, but instead must discover which actions yield the most reward by trying them out. In the most interesting and challenging cases, actions may affect not only the immediate reward but also the next situation and, through that, all subsequent rewards. These three characteristics --- being closed-loop in an essential way, not having direct instructions as to what actions to take, and where the consequences of actions, including reward signals, play out over extended time periods --- are the three most important distinguishing features of the reinforcement learning problem."

These two paragraphs from chapter 1 of the course textbook describe the topic of this course. The course is a graduate seminar. There will be some assigned readings and discussions. The exact content of the course will be guided in part by the interests of the students. It will cover at least the first 9 chapters of the (2nd edition of the) course textbook. Beyond that, we will either continue with the text or move to more advanced and/or recent readings from the field with an aim towards focussing on the practical successes and challenges relating to reinforcement learning.

There will be a programming component to the course in the form of a final project. Students will be expected to be proficient programmers.


Prerequisites

Some background in artificial intelligence and strong programming skills are recommended.


Text

The course textbook is:
Reinforcement Learning: An Introduction.
By Richard S. Sutton and Andrew G. Barto.
MIT Press, Cambridge, MA, 1998.
Note that the book is available on-line
As much as possible, we will be using the 2nd edition of the book, which is available in draft from from that webpage.


Assignments

Reading, written, and programming assignments will be updated on the assignments page. A tentative schedule for the entire semester is posted. But the readings and exercises may change up until the Wednesday of the week before they are due (1 week in advance).


Resources

Slides from class and other relevant links and information are on the resources page. If you find something that should be added there, please email it to me.


Discussion Forum

While the Professor and the TA would be glad to answer any questions you have, you would frequently find your peers to be an equally important resource in this class.

Please subscribe to our class piazza page.


Course Requirements

Grades will be based on:

Written responses to the readings (10%):
By 5pm on the afternoon before a class with a new reading assignment due, everyone must submit a brief question or comment about the readings as an email in plain ascii text. Please send it in the body of the email, rather than as an attachment. Please use the subject line "class readings for [due date]". In some cases, specific questions may be posted along with the readings. But in general, it is free form. Credit will be based on evidence that you have done the readings carefully. Acceptable responses include (but are not limited to):
  • Insightful questions;
  • Clarification questions about ambiguities;
  • Comments about the relation of the reading to previous readings;
  • Solutions to problems or exercises posed in the readings;
  • Critiques;
  • Thoughts on what you would like to learn about in more detail;
  • Possible extensions or related studies;
  • Thoughts on the paper's importance; and
  • Summaries of the most important things you learned.
  • Example successful responses from a previous class are available on the sample responses page.

    These responses will be graded on a 10-point scale with a grade of 9 being a typical full-credit grade. Responses will be due by 5pm on Monday. Responses received between then and 8:00a.m. on Tuesday will be deducted 1 point (for a maximum score of 9). Responses received between then and 8:00a.m. Thursday will be deducted 2 points (for a maximum score of 8). Responses received after that will be deducted 4 points (for a maximum score of 6).

    These deadlines are designed both to encourage you to do the readings before class and also to allow us to incorporate some of your responses into the class discussions.

    Class participation (10%):
    Students are expected to be present in class having completed the readings and participate actively in the discussions.

    Oral presentation/discussion moderation (10%):
    Each student will be expected to lead a discussion on one of the readings. The discussion can begin with a brief summary/overview of the important points in the readings, but the assumption is to be that everyone has already completed the readings. The student may either present material related to the readings (perhaps from an outside source) or moderate a class discussion about the readings. In the latter case, the student must be prepared to keep the conversation flowing. Here are some tips on leading a discussion. It is required that you present your plan for the discussion, including any slides you intend to show, to the Professor and TA at least two nights prior to your discussion (Sunday or Tuesday night, depending on what day you're presenting). Sign ups for discussion slots can be found here. Sign up for one slot -- no day should get 2 people unless necessary.

    Preliminary programming exercises (4) (30%):
    Each student will be required to complete four minor programming assignments of his/her own choosing. In most cases these will come from the exercises, though other options are possible upon consultation with the instructor. These exercises need not involve extensive or elaborate programs. The emphasis is to be on empirically analyzing various learning algorithms and reporting on the results. The reports should be emailed to the instructor and TA and all relevant code and data should be submitted on canvas. Each student may choose when to complete these exercises and on what topic. However at least three must be completed during the first half of the semester, at least one of which will be presented in class. It is recommended that the other be completed in conjunction with the student's oral presentation/discussion moderation.
    Grading criteria for programming assignments (out of 10):
    7 and 7.5 - Adequate, but really didn't go beyond the minimal analysis Example
    8 and 8.5 - Good job, but there is room for improvement Example
    9 and 9.5 - Good analysis, results well presented Example
    10 - Excellent, with interesting research issues identified. Doing more than what has been asked. Example

    From the TA: My general rubric is based on 3 things (keep in mind this is only a rough guide of what I'm looking for):

    (1) Clarity - How clearly was the experiment, hypothesis, and results motivated and explained?
    This is usually the "first cut," and reports that weren't clear usually won't pass the 8.5 mark (and it drops from there depending how unclear). If I'm confused about what you did, then the rest of the report usually doesn't go well. While we're not evaluating your English, it does still make a difference.

    (2) Insightfulness - How interesting was the question, and how much "digging" was involved?
    For example, a lot of people did "parameter sweep" experiments using some algorithm. This is fine, but it doesn't really "transfer." Knowing that some setting of parameters worked in some grid world doesn't necessarily help you on a new problem. There was one report that looked at how the structure of the domain (e.g. connectedness, etc.) suggested parameter choices, and that justified bumping it up. As another example, sometimes, experiments would give strange results (for example, the most common one was the Gambler's problem). The degree to which they investigated the reason for those results also factored into the score. This is the second cut, and is what mostly separates the 9.5 and up from below.

    (3) Effort - This is pretty subjective, as there usually isn't a "right" or "wrong" answer to the assignments. See the example reports above.

    Final programming project (40%):
    A more extensive final programming project, along with written report, will be due one week after the last day of class. Students will be expected to agree with the instructor on the topic of the project by about halfway through the semester. The report should be roughly equivalent to a conference paper in format, length, and style. If it's an application-oriented project, empirical results with statistical significance analysis should be included to evaluate the approach. Please upload a copy of your source code, your final report, and any other relevant data to canvas by one week after the last day of class. Please also send an email with a pdf version of the report attached to both the Professor and TA at that time.


    Related Courses


    UTCS Reinforcement Learning Reading Group


    [Back to Department Homepage]

    Page maintained by Peter Stone
    Questions? Send me mail