The Arrow cluster and our 22 CPLEX licenses were funded under a CFI grant. This grant requires us to complete annual reports explaining how this infrastructure is being used. If you use the cluster or CPLEX licenses for a project, large or small, please enter a bullet item here that gives a short description of your project and the role the cluster played. If you used these resources for multiple projects and/or for a period of time that spans the time periods given here, please create multiple bullets.

May 2007 -- May 2008

• Sample project: description (year(s); students/faculty involved)

May 2006 -- May 2007

• Sample project: description (year(s); students/faculty involved)
• Learning network structure using L1 regularization: many machine learning algorithms are embarassingly parallel, since they involve running the same code on different subsets of the data for purposes of cross validation, estimating error bars, etc; we used many Arrow nodes to speed up the experiments described in "Learning Graphical Model Structure using L1-Regularization Paths", M Schmidt, A Niculescu-Mizil, K Murphy. AAAI'07
• Bayesian learning of network structure features: in a related project, we looked at new MCMC algorithms for estimating probabilities of network features. We used Arrow to speed up the experiments described in "Bayesian structure learning using dynamic programming and MCMC", D Eaton, K Murphy, UAI'07 to appear.
• Action Graph Games: Using the arrow cluster we carried out computational experiments evaluating the performance of our proposed algorithms for action graph games, which is a compact representation of game-theoretic models. (May 2006 - ongoing; Albert Xin Jiang and Kevin Leyton-Brown, CS)
• CS 540 project: obstacle navigation project which needed to solve a MILP problem to get the trajectory. Using CPLEX(and the cluster) allowed me to construct more interesting cases and get the solutions faster than the free solver GLPK. (April 2007, Alex Gukov, CS)
• TAC SCM: Used CPLEX's MILP solver to schedule production and deliveries for a SCM scenario (Sept 2006 - on going; Erik Zawadzki, CS)
• Exact regularization of convex programs: I used CPLEX (thanks, Kevin!) for the numerical results of a paper (see reference below). This paper gives the most general conditions possible under which a general convex program (eg, linear, quadratic, semidefinite, second-order cone, etc) can be regularized, and a solution of the original problem still be obtained.
  • Michael P. Friedlander and Paul Tseng, Exact regularization of convex programs, To appear in SIAM Journal on Optimization. Also listed as Dept. of Computer Science Tech. Rept. TR-2006-26, November 2006 (revised April 2007).
• Empirical hardness model for SAT: Using arrow cluster, we collect SAT solvers’ runtimes data and problem instances’ features data. Empirical hardness models are trained to predict a solver’s runtime for a given problem instance based on instance’s features. We can also predict the satisfiability for an instance based on those features.(May 2006 - ongoing; Lin Xu, Holger H. Hoos and Kevin Leyton-Brown, CS)

-- KevinLeytonBrown - 05 May 2007