LANS Informal Seminar: Kibaek Kim

Seminar Title: Scalable Mixed-Integer Programming Methods and Applications
Speaker: Kibaek Kim, Postdoctoral Appointee, MCS, Argonne National Laboratory

Date/Time: 2016-05-11 15:00
Location: Bldg 240 rms 1404-1405

Description:
Mixed-integer programming (MIP) is a mathematical optimization in which some of the variables are restricted to be integers. MIP has provided great practical modeling power with the capability of representing logical and countable decisions. However, due to computational difficulties by the integrality restriction, finding a good solution for a MIP problem can take days and hours of computing time.

We present scalable MIP methods that exploit decomposable structures embedded in certain classes of MIP problems. The first part of this talk focuses on decomposition methods and algorithmic innovations for solving stochastic MIP problems in parallel. Then we introduce our parallel MIP solver (DSP) that implements the decomposition methods capable of running on HPC cluster using MPI library. To benchmark DSP with existing methods and solvers, computational experiments were performed by using stochastic unit commitment and standard test instances. The second part presents an application that considers better use of data centers to harness stranded power in electricity grid system. The numerical analysis was performed on the Western Electricity Coordinating Council (WECC) system with thousands of wind scenarios, for which we used DSP and Swift together on the Blues cluster. The remaining talk discusses how our decomposition idea can be applied for other MIP problems, including 1) decomposition methods for stochastic mixed-integer nonlinear programming; 2) temporal decomposition for production cost modeling; and 3) network decomposition for multiple interdependent infrastructure systems.