 |
Reinforcement Learning and
Artificial
Intelligence (RLAI)
|
Graduate
Project - Literature Review
|
The deadline for turing in your literature review has been set at Tuesday December 14th
The goal of this project is produce a document that provides an
overview of an aspect of reinforcement learning that is interesting to
you. The topic choice will be up to you, but you should discuss
it with Dr. Sutton before committing to it to ensure that it is
appropriate.
There will be time available on the
day of the midterm (Nov 4) for students to meet with Dr. Sutton and
discuss their ideas. You should have some idea about your
topic before this day. We'll only have 5-10 minutes per student,
so the idea is to help direct your ideas, not to give you an idea.
The final document should be between 10-20 pages, single spaced, 12 pt
font. This page limit will be enforced. We expect that it
is well written, with proper citations/references, etc. Your
research should reference at least
five papers that you have read and understood.
You will be marked primarily on three things: 1) the set of papers you
chose (can you search the literature on a topic and find the most
relevant papers), 2) the quality of the document (how well it is
written),
and 3) how
well you understood the topics discussed in the paper (as reflected in
your discussion of them and how they interrelate. For the
intended audience of the paper you should take your fellow RLAI
students. They have read the book and are familiar with what is
there. Now take them into a new part of the literature.
Below, we're going to try and provide a list of vague, brainstormed
ideas that you may find interesting. This list is by no means
exhaustive, bring you own if you have any!
Literature Review Ideas:
------------------------------
- Balancing exploration and exploitation (Alborz)
- Function approximation with reinforcement learning
- Reinforcement learning with real valued actions
- Practical successes and failures of reinforcement learning
- Policy-search and policy gradient methods
- Average-reward RL
- Least-squares methods (Cosmin)
- Planning with reinforcement Learning
- Reinforcement learning for games
- Reinforcement learning for control/robotics
- Customizing the reward function to improve agent performance
- Adding prior knowledge - teaching, coaching, showing, shaping, initializing
- Reinforcement learning with hierarchical and temporally abstract actions
- learning to achieve subgoals (David, Yusen)
- Discovery - finding good temporally abstract actions (Eddie)
- RL and psychology and neuroscience (Patrick)
- Reinforcement learning in partially observable or non-Markov
problems (Armita)
- RL in finance
- RL in telecommunications
- If you have more ideas and you are not planning on pursuing them,
please submit them so we can add to this list
If anything is unclear, or if there are questions - please ask as
soon as possible so that we can get the project details finalized.
-Brian
Back
to main page
Or does the review have to be a unique project just for this course?
Well, yes and no. You can't really do the same work and use it to satisfy two requirements. But you could do two related things, or do a larger, longer review that would be used for both purposes.