Projects can be done by you as an individual, or in teams of two students.   Each project will also be assigned a 725 instructor as a project consultant/mentor.   They will consult with you on your ideas, but of course the final responsibility to define and execute an interesting piece of work is yours.  Your project will be worth 30% of your final class grade, and will have two final deliverables:

1. a writeup in the format of a NIPS paper (8 pages maximum in NIPS format, including references; this page limit is strict), due May 5th by 3pm by email to the instructors list, worth 60% of the project grade, and

2. a poster presenting your work for a special class poster session on May 1st, 3-6pm in the NSH Atrium, worth 20% of the project grade.

In addition, you must turn in a midway progress report (5 pages maximum in NIPS format, including references) describing the results of your first experiments by April 9th in class, worth 20% of the project grade. Note that, as with any conference, the page limits are strict! Papers over the limit will not be considered.

Project Proposal

You must turn in a brief project proposal (1-page maximum) by March 5th in class.  Read the list of potential project ideas below.  You are encouraged to use one of these ideas.   If you prefer to do a different project and you are proposing your own data set you must have access to this data already, and present a clear proposal for what you would do with it.

Project proposal format:  Proposals should be one page maximum.  Include the following information:

• Project title

• Project idea.  This should be approximately two paragraphs.

• Data set you will use.

• Software you will need to write.

• Papers to read.  Include 1-3 relevant papers.  You will probably want to read at least one of them before submitting your proposal.

• Teammate: will you have a teammate?  If so, whom?  Maximum team size is two students.

• April 9 milestone: What will you complete by April 9?  Experimental results of some kind are expected here.

Project suggestions:  Below are descriptions of some suggested project ideas.  You are encouraged to select and flesh out one of these projects, or make up you own well-specified project.  If you have other project ideas you would like to work on, we would consider that as well, provided you get approved by the course instructors.

# Suggested Projects

Project 1: Implement a no-regret learner for some interesting problem -- e.g., a structured classification problem, or the problem of finding a tour in a graph, or the problem of finding an extensive-form correlated equilibrium in a game tree with incomplete information.  Some interesting no-regret algorithms include follow the perturbed leader (Kalai & Vempala) or Lagrangian hedging (http://www.cs.cmu.edu/~ggordon/ggordon.CMU-CALD-05-112.no-regret.pdf).  And, on request, We can provide a draft related to extensive-form correlated equilibrium

Project 2: Implement differential dynamic programming (DDP) to find a controller that is optimal (perhaps subject to some constraints) in an interesting problem.  The world would be modeled as x(t+1) = f(x(t), action(t), noise(t)), where f() is differentiable but nonlinear.  There would be a reward function r(x(t), action(t)) which defines optimality.  The controller would look like action(t) = A(t) (x(t) - x0(t)), where A(t) and x0(t) are parameters to be learned. DDP works by starting from some nominal trajectory x(t), action(t), and Taylor expanding f() and r() around the nominal trajectory.  Then it finds the optimal controller in this linear-quadratic problem using linear algebra, and updates the states and actions to be closer to what the optimal controller says.

Project 3: Implement max-margin planning (Ratliff & Bagnell) and use it to learn how to solve an interesting class of planning problems.

Project 4:  Implement Maximum margin Markov networks, and one of the solution methods: SMO, exponentiated gradient or subgradient.

Project 5:
Implement one of the generalizations of Max margin Markov Nets
Learning Structured Prediction Models: A Large Margin Approach; Ben Taskar, Vassil Chatalbashev, Daphne Koller and Carlos Guestrin;
In the 22nd International Conference on Machine Learning (ICML 2005), Bonn, August 2005. [PS version]

Project 6: Use the Convex-Concave Procedure to solve a non-convex optimization problem and compare to local search techniques such as simulated annealing.  A.L. Yuille and Anand Rangarajan, “The Concave-Convex Procedure (CCCP),” Neural Computation, Vol. 15, No. 4, April 2003, pp 915-936.pubs\ucla\A177_ayuille_NC2003.pdf

Project 7: use SATURATE to perform robust experimental design in Gaussian processes:  A. Krause, B. McMahan, C. Guestrin, and A. Gupta. Selecting observations against adversarial objectives. In NIPS, 2007.

Project 8: Implement Queyranne's algorithm for submodular function minimization and apply to finding independent sets of variables, e.g., for structure learning in graphical models.   Maurice Queyranne. A combinatorial algorithm for minimizing symmetric submodular functions. In SODA, 1995.

Project 9: Use the submodular-supermodular procedure to solve a non-submodular combinatorial optimization problem: M. Narasimhan and J. Bilmes. A submodular-supermodular procedure with applications to discriminative structure learning. In Advances in Neural Information Processing Systems (NIPS) 19, 2006.

Project 10:  Learn about LP decomposition techniques using ideas from graphical models by implementing a solution method for factored MDPs: Planning Under Uncertainty in Complex Structured Environments; Carlos Guestrin;  Ph.D. Dissertation, Computer Science Department, Stanford University, August 2003.

Project 11: Learn distance metrics for learning problems using semi-definite programming: E.P. Xing, A.Y. Ng, M.I. Jordan and S. Russell, Distance Metric Learning, with application to Clustering with side-information, Advances in Neural Information Processing Systems 16 (NIPS2002), (eds. Becker et al.) MIT Press, 521-528, 2002. ps

Project 12: $\ell_1\text{-}\ell_q$ Regularized Regression and It's Applications.  In this project,  we consider the problem of grouped variable selection in high-dimensional regression using $\ell_1\text{-}\ell_q$ regularization  ($1\leq q \leq \infty$), which can be viewed as a natural generalization of the $\ell_1\text{-}\ell_2$ regularization (the group Lasso). We have analyzed some theoretical properties of such an estimator and implement the algorithm in Matlab using blockwise coordinate descent algorithms and nonlinear thresholding. The main purpose of this project is apply these programs on some simultation data to illustrate the corresponding theoretical results. We will also use it to analyze some real-world data, like text data and microarry data.  For the reference, see http://arxiv.org/abs/0802.1517

Project 13:   Simultaneous Sparse Additive Models for Multi-task Learning.   Sparse additive models (SpAM) is a new class of methods for high-dimensional nonparametric regression and classification. It combine ideas from sparse linear modeling and additive nonparametric regression. Efficient algorithms has been developed to fit the mdoel even when the number of covariates is larger than the sample size and has been show to be effective on both simulation data and real-world datasets.  Current SpAM mdoels are implmented in R using a backfitting type algorithm.  We also developed an algorithm which extend the SpAM from single-task setting to multi-task setting for joint nonparametric variable selection. In this project, we will try to apply the multi-task SpAM models to both simulation data and real-world data. For the reference, see http://arxiv.org/abs/0711.4555

Project 14 : (Implement an algorithm for optimization).  This type of project should involve a significant amount of implementation; it should not be e.g. implementing subgradient descent for SVMs or anything else which might fit in a regular homework.  Some possibilities include:

• Randomized simplex algorithm: A randomized version of the simplex algorithm was recently developed, and unlike the original simplex method, it is polynomial (in RP).  See "A randomized polynomial-time simplex algorithm for linear programming" by Kelner and Spielman (2006).
• Linear program solver which deals with degenerate cases.
• Interior point algorithm for linear programming. Originally, implementations were mainly based on Karmarkar's algorithm (Karmarkar, N. "A New Polynomial-Time Algorithm for Linear Programming. " Combinatorica 4, 373-395, 1984.), but modern implementations are based upon Mehrotra's predictor-corrector technique (Mehrotra, S. "On the Implementation of a Primal-Dual Interior Point Method." SIAM J. Optimization 2, 575-601, 1992.).
• Some variant of ellipsoid algorithm, to gauge how it does in practice. Would have to carefully think about the comparison though. Do we want them to implement a simplex based algorithm as well, or just compare against cplex / matlab?
• Boyd presents a modern interior-point method, which is also a good project idea, or affine scaling is another possible choice.