Projects
Goal & deliverables
The deliverables consists in a single .zip file posted on bspace containing:
A 1015 page project report detailing the case study. The report should be written preferably with LaTeX.
All the figures (in .pdf format) and matlab files needed to reproduce the results.
A README file detailing
The name of the .zip file should be of the form <XXX>_EE227BTF14.zip, where XXX is a tag for the project. Example: GRAPHMOD_EE227BTF14.zip.
Logistics
Each project involves a team of 34 students.
Groups and topics should be decided no later than October 28.
There will be a single project grade for each group.
Timeline
By Sep. 16: teams are formed.
By October 28: topics are chosen.
By November 13: project midpoint review. You are to present an outline of your report and a progress status.
By December 5: final reports due, 1011am, in 421 SDH.
December 512: project presentations, in class.
Types of projects
Projects can have the following different formats:
Literature review: The project describes in detail a set of 45 papers. This should include reproducing experiments. Example: learning sparse graphical models.
Design methodology: the project describes a (possibly new) methodology for addressing a particular design problem. Here the focus is not on algorithms or past work. Example: geometric programming for water distribution networks.
Theory: the project examines a theoretical issue related to convex optimization. Example: quality of SDP relaxation for sparse PCA.
Algorithms: the project tests various algorithms for solving a particular type of problem. Example: Stochastic vs. accelerated gradient for largescale logistic regression.
Past project examples
Project ideas
See also the resources page on bCourses.
SPARSEPROB: this project deals with statistical estimation of sparse probabilities, in the context of lowrank data. Contact: Mert Pilanci (mert@berkeley.edu).
ROBPRICE: this project involves the realtime pricing of millions of products sold online, based on uncertain demand. Contact: Arnau Puig (apuig@walmartlabs.com).
UNBAL: This project involves the solution of large binary classification problems where the number of examples in one class is much smaller than in the other. Contact: Edouard Grave (edouard.grave@gmail.com).
COVEST: This project involves the estimation of large covariance matrices with sparse inverses, using market data. Contact: Laurent El Ghaoui.
MATCOMP: This project involves matrix completion problems, more specifically an extension of the robust PCA paradigm (which aims at decomposing a data matrix as a sum of a sparse and a lowrank matrix), to the case when the data matrix has missing entries. Contact: Laurent El Ghaoui.
HYDRO: this project is concerned with the management of a hydroelectric production line, which has been already formulated as a linear program with uncertain coefficients. Here the uncertainty is related to the demand, which has to be predicted. The focus of the project is on socalled affine recourse policies, as a method to handle the uncertain components. Contact: Laurent El Ghaoui.
DCOPT: this project addresses a class of nonconvex problems known as ‘‘difference of convex functions’’, which arise in a host of problems, including those involving the estimation of mixtures of probability distributions based on data. The project will examine: a) a review of the literature; b) a connection with existing algorithms, such as blockcoordinate descent, ExpectationMaximization; and c) the development of innovative approaches, based for example on semidefinite programming relaxations. Contact: Jerome Thai (jerome.thai@berkeley.edu).
