Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

WB11

6 - Extreme-point Search Heuristics for Solving Fixed-charge Generalized Network Problems Using Parametric Ghost Image Process Approach Angelika Leskovskaya, PhD Candidate, Southern Methodist University, Caruth Hall 3145 Dyer Street, Suite 372, Dallas, TX, 75275, United States, Richard S. Barr While researchers have studied generalized network flow problems extensively, the powerful addition of fixed charges on arcs has received scant attention. This work describes network-simplex-based algorithms that efficiently exploit the quasi-tree basis structure of the problem relaxations, proposes heuristics that utilize a tabu search and a progressive modification of a parameterized objective function for fixed-charge transportation problems, extends the parametric GIP approach to fixed-charge transshipment problems, and presents computational comparisons with commercial solution alternatives. n WB08 North Bldg 124A Stochastic Optimization Methods and Approximation Theory in Machine Learning I Sponsored: Optimization/Nonlinear Programming Sponsored Session Chair: El Houcine Bergou, INRA-KAUST, Jeddah, Saudi Arabiaö Co-Chair: Aritra Dutta, INRA-KAUST, Thuwal, Saudi Arabia 1 - Robust PCA by Manifold Optimization Teng Zhang, University of Central Florida, Orlando, FL, United States Robust PCA is a widely used statistical procedure to recover an underlying low- rank matrix with grossly corrupted observations. This work considers the problem of robust PCA as a nonconvex optimization problem on the manifold of low-rank matrices and proposes two algorithms based on manifold optimization. It is shown that, with a properly designed initialization, the proposed algorithms are guaranteed to converge to the underlying low-rank matrix linearly. Compared with a previous work based on the factorization of low-rank matrices the proposed algorithms reduce the dependence on the condition number of the underlying low-rank matrix theoretically. Simulations and real data examples confirm the competitive performance of our method. 2 - Kolmogorov Representation and Deep Learning Xin Li, University of Central Florida, 4000 Central Florida Blvd, Orlando, FL, 32816, United States In essence, Kolmogorov’s representation theorem states that every continuous function of high dimensional space is completely determined by a function of a single variable. The drawback is that the single variable function used in the representation can be highly non-smooth and hard to compute. On the other hand, Barron proved an approximation theorem with a universal convergence guarantee. This talk will explore the connection between the ideas of Kolmogorov and Barron with the aim at a possible understanding of deep learning through a form or structure of function approximation that is not as rigid as Kolmogorov’s representation but ``deeper’’ and more inclusive than Barron’s approximation. 3 - The Multiple Shades of Dropout for both Discriminative and Generative Deep Neural Networks Boqing Gong, Principal Researcher, Tencent AI Lab, Seattle, WA, United States Dropout, which independently zeros out the outputs of neurons at random, has become one of the most popular techniques in training deep neural networks due to its simplicity and remarkable effectiveness. This talk reveals multiple shades of dropout for both discriminative and generative deep neural networks. The first half of the talk focuses on the discriminative models and presents an improved version of the dropout. In the second half of the talk, I will provide a new perspective for understanding dropout under the context of deep generative neural networks. Despite being impactful on a variety of problems and applications, the generative adversarial nets (GANs) are remarkably difficult to train. In particular, our approach gives rise to the inception score of more than 5.0 with only 1,000 CIFAR-10 images and is the first that exceeds the accuracy of 90% on the CIFAR-10 dataset using only 4,000 labelled images, to the best of our knowledge. n WB10 North Bldg 125A Joint Session MSOM/APS: Design and Analysis of Emerging Service Systems Sponsored: Manufacturing & Service Oper Mgmt Sponsored Session Chair: Mohammad Delasay, Stony Brook University, Stony Brook, NY, 11794, United States

Co-Chair: Sherwin Doroudi, University of Minnesota, Minneapolis, MN, 55455, United States 1 - Signaling in Online Retail: Efficacy of Public Signals David Lingenbrink, Cornell University, Ithaca, NY, 14850, United States, Krishnamurthy Iyer We study an online retail setting, where strategic customers seek to purchase an item with a finite inventory in one of two time periods. Only the firm knows the inventory and demand, and seeks to persuade customers to buy in the more expensive first time period through signaling. We analyze both public and private signaling, and find that, with homogeneous customers, it is optimal to signal publicly: the firm cannot raise its revenue by sending a different private signal to each customer. However, when customers are heterogeneous, we find examples where private signaling outperforms public. 2 - Coffee Shop Operations with Mobile Ordering Kang Kang, University of Minnesota, Minneapolis, MN, 55455, United States, Mohammad Delasay, Sherwin Doroudi Mobile ordering is becoming popular in the world of coffee shop, fast food or restaurants. Our work examines the operations of a coffee shop where some customers can use a mobile app to skip the ordering and payment queue. Using queueing theory, we explore the impact of mobile customers on walk-in customers and vice versa across a variety of service policies. 3 - A Queueing Model and Analysis for Autonomous Vehicles on Highways Neda Mirzaeian, Carnegie Mellon University, Pittsburgh, PA, United States, Soo-Haeng Cho, Alan Scheller-Wolf Autonomous vehicles (AVs) have a potential to significantly improve highway congestion, since these vehicles are able to maintain smaller inter-vehicle gaps, and travel together in larger platoons (or batches) than human-driven vehicles (HVs). We model a highway segment as a queueing system, and analyze two policies: the designated-lane policy (designating one lane to AVs) and the integrated policy (allowing both AVs and HVs in any lane). After calibrating our model to data, we show that, depending on the proportion of AVs and the highway load, either of these polices can outperform the other. 4 - Assignment of Jobs to Servers under Interference Scott Votke, Stony Brook University, Stony Brook, NY, United States, Jazeem Abdul Jaleel, Mohammad Delasay, Sherwin Doroudi, Anshul Gandhi In cloud computing, interference is a major drawback to latency sensitivity and performance analysis. In this paper, we study an M/M/c queuing system where the individual servers are subject to periods of interference, which result in lower than normal service rates. A controller decides whether to send incoming jobs to an available server or to the queue. We assume that the length of interference for all servers are i.i.d. exponentially distributed random variables. By modeling such a system as a Markov decision process, we investigate the optimal policy on how to route incoming jobs. n WB11 North Bldg 125B Joint Session MSOM/Practice Curated: Ride-sharing Services Research at Didi Chuxing Sponsored: Manufacturing & Service Oper Mgmt Sponsored Session Chair: Zhixi Wan, DiDi Chuxing, Beijing, China 1 - Modelling of Driver Supply Behaviour in On-demand Shared Transportation Platforms Hai Wang, Singapore Management University, Room 5023, School of Information Systems, 178902, Singapore, Hao Sun, Zhixi Wan, Guobin Wu, Qun Li With the popularization of ride-sharing services, drivers working as freelancers on ride-sharing platforms can design their schedules flexibly. They make daily decisions regarding whether to participate in work, and if so, how many hours to work. Driver’s extra income, participation cost, and maximum allowed working time affect these decisions. We incorporate these features into classical theory of labour economics and propose a theoretical model to describe how drivers make working decisions on the shared platform. We characterize the labour supply pattern of participation, working hours, extensive and intensive margin elasticity on the platform and compare with the traditional taxi.

443