2016 INFORMS Annual Meeting Program

MC01

INFORMS Nashville – 2016

Monday, 1:30PM - 3:00PM

Acceleration Of A Communication Efficient Distributed Dual Block Descent Algorithm Chenxin Ma, Lehigh University, 200 West Packer Avenue, Bethlehem, PA, 08801, United States, chm514@lehigh.edu Distributed optimization algorithms for very large-scale machine learning suffer from communication bottlenecks. Confronting this issue, a communication- efficient primal-dual coordinate ascent framework (CoCoA) and its improved variant CoCoA+ have been proposed, achieving a convergence rate of O(1/t) for solving empirical risk minimization problems with Lipschitz losses. In this paper, we propose an accelerated variant of CoCoA+ and show that it has a rate of O(1/t2) in terms of reducing dual suboptimality. Our analysis is also notable in that our convergence rate bounds involve constants that, except in extreme cases, are significantly reduced. Primal-dual Interior-point Methods With Domain-driven Barriers Mehdi Karimi, PhD Student, University of Waterloo, 200 University Avenue West, Department of Combinatorics and Optimization, University of Waterloo, Waterloo, ON, N2L 3G1, Canada, m7karimi@uwaterloo.ca, Levent Tuncel While primal-dual algorithms have yielded efficient solvers for convex optimization problems in conic form over symmetric cones, many other highly demanded convex optimization problems lack comparable solvers. To close this gap, we develop infeasible-start primal-dual interior-point algorithms for convex optimization problems in “domain-driven” formulation, which we show covers many interesting optimization problems including the conic ones. After presenting our techniques, we introduce our Matlab-based code that solves a large class of problems including LP, SOCP, SDP, QCQP, Geometric programming, and Entropy programming among others, and mention some numerical challenges. Upward Reward Perturbation For Reinforcement Learning Zhengyuan Zhou, Stanford University, 160 Comstock Circle, Unit 106002, Stanford, CA, 94305, United States, zyzhou@stanford.edu, Ling Zhu, Benjamin Van Roy, Nicholas Bambos Late Cancellation Applying Case Queries To Network Clusters: Identifying The Intussusception Signal In Rotashield Adverse Event Reports Matthew Foster, ORISE Fellow, FDA, 10903 New Hampshire Ave, Silver Spring, MD, 20993, United States, matthew.foster@fda.hhs.gov We constructed a network of Medical Dictionary for Regulatory Activities (MedDRA) Preferred Terms (PTs) from RotaShield adverse event reports and calculated the eigenvector, betweenness, and closeness centrality metrics. We used these metrics to cluster PTs and assessed the sensitivity of clusters using an intussusception (IS) case definition and a variety of Standardized MedDRA Queries. Clustering using eigenvector vs closeness centrality was superior to other combinations, although all identified the IS signal before the July 1999 CDC recommendation for discontinuation. The early detection of this safety signal supports the potential use of our methodology for safety surveillance. Neighborhood-based Reductions And Cuts For Signed Graphs Christopher Muir, University of Tennessee - Knoxville, 2044 Wilkerson Road, Knoxville, TN, 37922, United States, cmuir1@vols.utk.edu This research is concerned with improving solve times for instances of the Balanced Subgraph problem. We discuss various data reduction techniques based on the neighborhoods of vertices and on cut vertices. Additionally, we show that certain structures often present in signed graphs can be exploited to allow for faster solve times. Computational test results are also presented for previously explored problems, including the toll-like problem from the MIPLIB 2010 instance library.

MC01 101A-MCC Image and Shape Data Analysis Sponsored: Data Mining Sponsored Session Chair: Chiwoo Park, Florida State University, 2525 Pottsdamer Street, Tallahassee, FL, 32310-6046, United States, cpark5@fsu.edu Co-Chair: Kamran Paynabar, Assistant Professor, Georgia Institute of Technology, 755 Ferst Drive, NW, Atlanta, GA, 30332, United States, kamran.paynabar@isye.gatech.edu 1 - State Space Model For Time-varying Density Estimation Yanjun Qian, Texas A&M University, College Station, TX, 77840, United States, qianyanjun09@gmail.com, Yu Ding, Jianhua Huang In both scientific and industrial fields, estimating the time-varying density can provide an important tool for the monitoring or research purpose. In our work, we propose a new method for the time-varying density estimation based on the state space model. Our method can learn the system parameters off-line, and provides an on-line density curve updating from the observed histogram data. By adding a spatial penalty term, we can guarantee the smoothness of the estimated curves and prevent the over-fitting. At last, we show the application of our method on the study of the nanocrystal growth process. 2 - Dynamic Network Modeling Of In-situ Image Profiles For Statistical Process Control – Applications In Ultraprecision Machining And Biomanufacturing Process Chen Kan, Pennsylvania State University, PSU, University Park, PA, 16802, United States, CJK5654@psu.edu, Hui Yang Modern industries are investing in advanced imaging technology to increase information visibility, cope with system complexity, and improve the quality and integrity of system operations. Realizing the full potential of advanced imaging technology for process monitoring and control hinges on the development of new SPC methodologies. This paper presents a novel dynamic network methodology for monitoring and control of high-dimensional imaging streams. 3 - Structured Point Cloud Data Modeling Via Regularized Tensor Decomposition And Regression Hao Yan, Georgia Institute of Technology, yanhao@gatech.edu, Massimo Pacella, Kamran Paynabar Due to the easy accessibility of the 3D metrology tools such as Coordinate Measuring Machine or scanning tools, structured point cloud data is becoming more and more popular. Therefore, modeling the structured point cloud is an important task in many application domains. We model the structure point cloud as tensor and propose regularized tucker decomposition and regularized tensor regression to detect the variation patterns in the data and link these patterns to the process variables. Furthermore, the performance of the proposed method is evaluated through simulation and a real case study in the point cloud data in the turning process. 4 - Statistical Analysis Of Preferential Orientations Of Two Shapes In Their Aggregate Ali Esmaieeli, Florida State University, Tallahassee, FL, 32304, United States, ae13e@my.fsu.edu, Chiwoo Park, David Welch, Roland Faller, Taylor Woehl, James Evans, Nigel Browning Nanoscientists believe that adjacent nanoparticles aggregate with each other in specific preferential directions. This phenomenon is known as oriented attachment and can be studied by direct observations using dynamic electron microscopy. These studies relied on manual and qualitative analysis up to now; therefore, in this research we are proposing a statistical approach to study the oriented attachment believing that certain geometries have specific preferential orientation when they aggregate. We use multiple aggregation examples collected from dynamic microscope data in order to examine the performance of our approach.

182