Final project assignments
Write an annotated bibliography listing papers and games relevant to applications of reinforcement learning. Briefly discuss each paper/game with respect to four categories of learning problems: 1) Policy 2) World model 3) Measurement model 4) Reward model. (See figure below, and Paul's lecture.)
You should also provide a rating on 5-star scale in which you assess overall quality in the context of whether you'd recommend this paper to the rest of the seminar group. 5 stars means "yes, I'd highly recommend this paper", and 1 star means "no, don't read this".
The bibliography should include references to published literature, but can also include citations to games (e.g. you could expand on Shawn's list). When discussing a game, provide a paragraph that briefly describes the game, and explains how some aspect of the game illustrates learning at least one of the four levels.
Pay particular attention to the question of credit assignment and implications for learning transfer to new tasks. Specifically, provide support and/or criticism of two hypotheses:
1) Learning occurs at more specific levels first, because less information is required
2) If policy learning is sufficient to perform up to a given aspiration level, learning will not occur at the more general levels
Aim for about 20 papers and/or games, with a paragraph on each. The main goal is to relate theory as discussed in class to applications, such as empirical work drawn from human and animal behavioral studies. Theoretical/computational papers that we have not discussed in class may also be included. The idea is to provide a broad sampling of the literature without necessarily going into great detail or depth.
Your paper is due: Saturday, May 16th, 2009. Feel free to discuss among yourselves, but the final paper should be your own writing and thoughts.
Email your bibliographies to: kersten@umn.edu.
In your email, let me know if it is OK to share your bibliography with your identity, with the rest of the class via the class web page.