2015 Informs Annual Meeting

SD15

INFORMS Philadelphia – 2015

SD16 16-Franklin 6, Marriott New Optimization Modeling and Effective Techniques Sponsor: Optimization/Linear and Conic Optimization Sponsored Session Chair: Jiming Peng, Associate Professor, University of Houston, UH, Dept of Industrial engineering, Engineering Bldg 2 221A., Houston, TX, 77204, United States of America, jopeng@Central.UH.EDU 1 - Dropconnect in Deep Learning via Lagrangian Diego Klabjan, Professor, d-klabjan@northwestern.edu, Mark Harmon Dropconnect is a regularization technique for deep neural nets when random weights are enforced to be zero. We present a Lagrangian-based algorithm for restricting weight values. 2 - New Global Algorithm for Linearly Constrained Quadratic Programming Jiming Peng, Associate Professor, University of Houston, UH, Dept of Industrial engineering, Engineering Bldg 2 221A., Houston, TX, 77204, United States of America, jopeng@Central.uh.edu How to find the global optimal solution to LCQP has been a long-standing challenge in optimization. In this talk, we introduce a new design framework for LCQP that integrates several simple effective optimization techniques such as Lagrangian methods, Alternate update method, convex relaxation, initialization and partitioning. We establish the global convergence of the algorithm and estimate its complexity. Promising numerical results for large-scale LCQPs will be reported. 3 - An Optimization Perspective on Systemic Risk Aein Khabazian, PhD Student/research Assistant, Univirsity of Houston, UH, Dept of Industrial engineering, Engineering Bldg 2, We consider the issue of assessing the systemic risk under uncertainty in a financial system based on the model proposed by Eisenberg and Noe, in which the interbank liabilities are assumed to be known, and the non-interbank assets are assumed to be constant. However, in real world application this information is typically unknown or subject to market fluctuation. In this regard, we develop robust optimization and worst case optimization to account for the uncertainties in the constraints. SD17 17-Franklin 7, Marriott Transportation Network Modeling and Optimization Sponsor: Optimization/Network Optimization Sponsored Session Chair: Vladimir Stozhkov, University of Florida, 2330 SW Williston Rd Apt 2826, Gainesville, FL, 32608, United States of America, vstozhkov@ufl.edu 1 - A Distributed Hierarchal Shortest Path Algorithm for Large-Scale Transportation Networks Ala Alnawaiseh, Postdoctoral Researcher, Southern Methodist University, 3101 Dyer St.,, #219, Dallas, TX, 75205, United States of America, aalnawai@smu.edu, Khaled Abdelghany, Hossein Hashemi This paper presents a distributed hierarchical shortest path algorithm for large- scale transportation networks. The algorithm integrates a network augmentation as well as a divide-and-conquer techniques to solve the all-to-all shortest path problem in a distributed fashion. Preliminary results that illustrate the superiority of the algorithm are presented. 2 - How to Design an Effective Off-hour Delivery (OHD) Program: A Network Design Perspective Sevgi Erdogan, Faculty Research Associate, University of Maryland-NCSG, 1112 J Preinkert Field House, College Park, MD, 20742, United States of America, serdogan@umd.edu, Jiangtao Liu, Wenbo Fan, Xuesong Zhou The OHD has been considered as an effective policy to overcome the negative externalities like congestion due to freight traffic in urban areas during business hours. This study proposes a network design approach to determine links to be restricted for truck traffic to shift truck demand to off-hours and routes to less congested areas so that the maximum benefit from an OHD program can be achieved. The approach will guide cities in implementing effective OHD programs. Houston, TX, 77204, United States of America, akhabazian@uh.edu, Jiming Peng, Aida Khayatian

2 - Generalized Sequential Assignment Problem Arash Khatibi, University of Illinois, 201 North Goodwin Avenue, Urbana, IL, United States of America, khatibi2@illinois.edu, Sheldon Jacobson The Sequential Stochastic Assignment Problem (SSAP) assigns sequentially arriving tasks with stochastic parameters to workers with fixed success rates so as to maximize the total expected reward. This paper uses the Secretary Problem to propose assignment policies for the SSAP, when there is no prior information on task values. This paper also discusses the Doubly Stochastic Sequential Assignment Problem (DSSAP), which is an extension of SSAP with the workers’ success rates assumed to be random. 3 - Stochastic Versus Dynamic Programming for a Transportation Procurement Problem Francesca Maggioni, Assistant Professor, University of Bergamo, We consider the problem of a producer which has to ship a load to a customer at discrete times and fixed horizon. Different companies offer a transportation price with realization available at the end of the time period. A penalty is paid for the quantity that remains to be sent. We consider two variants of the problem: the minimum expected cost and the min-max total cost. We compare the performance of stochastic and dynamic programming approaches. Theoretical worst-case results are provided. 4 - Data-driven Schemes for Resolving Misspecified MDPs: Asymptotics and Error Analysis Hao Jiang, UIUC, 104 S. Mathews Ave., Urbana, IL, United States of America, jiang23@illinois.edu, Uday Shanbhag We consider the solution of a finite-state infinite horizon Markov Decision Process (MDP) in which both the transition matrix and the cost function are misspecified, the latter in a parametric sense. Learning-enhanced value and policy iteration schemes are proposed and shown to be convergent almost surely. Finally, we present a constant steplength misspecified Q-learning scheme and provide an error analysis. Via dei Caniana n 2, Bergamo, 24127, Italy, francesca.maggioni@unibg.it, Luca Bertazzi Nonlinear Optimization Algorithms Sponsor: Optimization/Nonlinear Programming Sponsored Session Chair: Frank E. Curtis, Lehigh University, 200 W Packer Ave, Bethlehem, PA, 18015, United States of America, frank.e.curtis@gmail.com 1 - A Stochastic Programming Model for Nurse Staffing in Post- Anesthesia Recovery Units Yueling Loh, Johns Hopkins University, 3400 North Charles Street, Baltimore, MD, 21218, United States of America, yueling.loh@gmail.com, Sauleh Siddiqui, Daniel Robinson We present a stochastic programming model to determine nurse staffing requirements in Post-Anesthesia Recovery Units under high variability in patient flow and length-of-stay. We will formulate the problem as a two-stage stochastic mixed integer program and provide some numerical results. 2 - Distributed Parallel Coordinate Descent Methods for Sparse Inverse Covariance Problem Seyedalireza Yektamaram, Lehigh University, Mohler Laboratory, 200 West Packer Ave, Bethlehem, PA, 18015, United States of America, sey212@lehigh.edu, Katya Scheinberg In graphical models, recovering the structure of underlying graph corresponding to conditional dependencies of random variables is of great importance,which is obtained by recovering the corresponding Sparse Inverse Covariance matrix. However, as the problem size grows larger, efficiency of most solution approaches reduce significantly, thus using distributed parallel techniques becomes essential. Here we explore distributed parallel coordinate descent methods to solve this problem efficiently. 3 - A Nonconvex Nonsmooth Optimization Algorithm with Global Convergence Guarantees Frank E. Curtis, Lehigh University, 200 W Packer Ave, Bethlehem, PA, 18015, United States of America, frank.e.curtis@gmail.com, Xiaocun Que An algorithm for minimizing nonconvex nonsmooth objective functions is presented. The algorithm is based on a BFGS strategy, enhanced with gradient sampling mechanisms to ensure convergence to a stationary point with probability one. An open source C++ implementation of the algorithm is also described along with results for a set of test problems. SD15 15-Franklin 5, Marriott

124