2013 Poster Sessions : Tools and Methods for Large-Scale Convex Optimization

Student Name : Eric Chu
Advisor : Stephen Boyd
Research Areas: Information Systems
We present a set of tools for large-scale convex optimization. These tools consist of a domain-specific language for modeling optimization problems and solvers for prototyping and deployment. We also present a first-order solver amenable to parallelization and capable of solving large, convex optimization problems. This solver is based off the observation that a solution of a cone program and its dual is a point in the intersection of an affine set and a cone. We apply an operator splitting method, the alternating direction method of multipliers, to either find a point in the intersection of the two sets or return a hyperplane that separates them. If the intersection is nonempty, we return a solution to the cone program; if the intersection is empty, we attempt to return a certificate of infeasibility. The resulting algorithm is simple and yields solutions of modest accuracy in competitive times. Several versions of the algorithm are amenable to parallelization.

(Joint work with Brendan O'Donoghue, Alex Domahidi [ETHZ], and Neal Parikh)

Eric Chu is a fifth-year PhD candidate in the Electrical Engineering department at Stanford. He is co-advised by professors Stephen Boyd and Dimitry Gorinevsky. His research interests are in distributed optimization and large-scale data analysis and is currently developing domain-specific languages and solvers for distributed optimization.