CS 497: Planning and Decision Making

**Information on Projects:**

Each student will be asked to submit a project. A substantial amount of flexibility exists in the kind of project that can be done. The main criteria for grading will be: 1) level of originality, creativity, cleverness, etc., 2) relevance to the theme of the course, 3) general difficulty, amount of effort, etc., 4) quality of the written account of the project (which could exist as a web page, ps/pdf file, or written document). Here are some examples of the kinds of projects, to help give you ideas:

- Carefully study frequentist decision theory, and use this knowledge to make a frequentist version of something covered in the first part of the class or in one of the papers. For example, can define a sequential decision making scenario, and apply dynamic programming to find optimal solutions? Another possibility would be to implement frequentist decision rules and design experiments that illustrate the tradeoffs and issues between Bayesian and frequentist philosophies.
- Implement and experimentally evaluate an algorithm from one of the papers. Can you identify strengths and weaknesses of the approach?
- Attempt to extend or improve an algorithm given in the first part of the class or in a paper.
- Define a variant of a model considered in one of the papers, and attempt to adapt the tools and techniques to apply to your variant.
- For a more-specific idea, Pareto optimality was covered, and also sequential decision making for problems other than multiobjective optimization. Can you take dynamic programming principles that were designed for single objectives and extend them to multiple objectives? For example, can you design an algorithm that behaves in a way similar to Dijkstra's algorithm for graph search, but instead yields all minimal solutions with respect to the partial order defined in class for multiobjective optimization problems? Can you analyze the running time of such an algorithm?
- For a given scenario, possibly taken from a paper, attempt to perform sensitivity analysis, either experimentally, or by proving bounds on the input parameters (e.g., prior distributions).