Informs Annual Meeting 2017

WC18

INFORMS Houston – 2017

3 - Diffusion Approximations for Load Balancing Mechanisms in Cloud Storage Systems Eric Friedlander, University of North Carolina at Chapel Hill, Chapel Hill, NC, 27599, United States, ebf2@live.unc.edu Analysis of large-scale communication networks (e.g. ad hoc wireless networks, cloud computing systems, server networks etc.) is of great practical interest. The massive size of such networks frequently makes direct analysis intractable. Asymptotic approximations using hydrodynamic and diffusion scaling limits provide useful methods for approaching such problems. In this talk, we explore an application of these approximation techniques to a model for load balancing in large, cloud-based, storage systems using Maximum Distance Separable (MDS) codes to improve reliability and retrieval speed. This is joint work with Amarjit Budhiraja. 4 - Taylor-ed DPS: Between Dynamics Programming and Brownian Control Problems Itai Gurvich, Associate Professor, Cornell University, Cornell Tech, New York, NY, 10025, United States, gurvich@cornell.edu, Junfei Huang, Anton Braverman We propose an approximation to Dynamic Programs (DP) that is grounded in perturbation techniques that have been recently developed for the performance analysis of queues. These approximations, utilizing the so-called Stein’s Method, work directly with the equations that characterize expected performance. This work seeks to move beyond the performance analysis of queues to develop “Taylor-ing” as a structured tool for dynamic optimization. 340B Applied Probability Contributed Session Chair: Lisha Yu, City University of HongKong, Kowloon, Hong Kong, lishayu2-c@my.cityu.edu.hk 1 - Modeling Correlated Wind Speeds Through Stochastic Differential Equations Rafael Zarate-Minano, Associate Professor, University of Castilla-La Mancha, Plaza Manuel Meca 1, Almaden, 13400, Spain, rafael.zarate@uclm.es, Miguel Carrion, Juan Miguel Morales Decision making tools involving stochastic programming techniques are widely used for the operation of power systems with a significant penetration of wind power. The stochastic nature of the wind speed is usually considered through a set of plausible scenarios generated by means of time series. However, time series have shown limited accuracy to model correlations among wind speeds at different locations. This work analyzes the ability of stochastic differential equations to model correlated wind speeds. Models are constructed and tested on the basis of real-world wind speed measurements. 2 - High Dimensional Linear Regression: Mean Square Error and Phase Transitions Ilias Zadik, PhD Candidate, MIT, Cambridge, MA, United States, izadik@mit.edu In this talk we will discuss some new theoretical results on the topic of sparse high dimensional linear regression with Gaussian matrix and noise. In the literature, some remarkable papers by Wainwright and others have established that with enough samples LASSO and other efficient recovery mechanisms can recover the ground truth vector. On the other hand, the information theoretical results on the topic suggest that the vector is recoverable with much fewer samples than the one LASSO needs. We show that recovery of the ground truth vector is indeed possible with much fewer samples than the one LASSO needs and we prove this recovery becomes possible very suddenly as the sample size increases. 3 - Search in Discrete Locations with Two Search Modes Kyle Y Lin, Naval Postgraduate School, 1411 Cunningham Rd, Monterey, CA, 93943, United States, kylin@nps.edu, Jake Clarkson, Kevin D. Glazebrook An object is hidden in one of several locations. There are two search modes to search each location. The search time and the detection probability depend on both the location and the search mode. The goal is to minimize the expected total time to find the object. We study structure of the optimal policy, and design heuristic policies that deliver near-optimal performance in our numerical experiments. WC18

4 - Stochastic Models for Optimal Resource Allocation and Scheduling of Biomanufacturing Projects Yasemin Limon, University of Wisconsin-Madison, Madison, WI, 53706, United States, ylimon@wisc.edu, Ananth Krishnamurthy We study the resource allocation and scheduling problem common in biomanufacturing operations, where purification projects need to be assigned to scientists with varying skill sets, and projects need to be repeated if the scientist is unable to achieve the desired yield and purity requirements. We develop stochastic models to analyze different resource allocation strategies and their impact on waiting times for projects and total system costs. 5 - Exponential Convergence Rates for Stochastically Ordered Markov Processes with Random Initial Conditions Julia Romanski, Graduate Student, Massachusetts Institute of Technology, 77 Massachusetts Ave, Bldg. E40-103, Cambridge, MA, 02139, United States, romanski@mit.edu, Saurabh Amin, Patrick Jaillet In this paper we extend the results of Lund, Meyn, and Tweedie (1996) on exponential convergence rates for stochastically ordered Markov processes to allow for a random initial condition that is also stochastically ordered. We find an explicit exponential convergence rate for an M/M/1 queue beginning in equilibrium and then experiencing a perturbation of its arrival or departure rates. In order to show the convergence bound we also prove a result on hitting times of differences of independent Poisson processes, known as Skellam processes. Finally, we comment on the applicability of our result to other stochastically increasing processes, such as Jackson networks with changing arrival rates. 6 - Detecting Node Propensity Changes in the Dynamic Degree Corrected Stochastic Block Model Lisha Yu, City University of HongKong, 88 Tat Chee Avenue, Kowloon, Hong Kong, lishayu2-c@my.cityu.edu.hk Many applications involve dynamic networks which are evolutionary networks with timestamps. Studying the evolution of node propensity over time is significant in exploring and analyzing these networks. In this paper, we propose a multivariate surveillance plan to monitor node propensity in dynamic social networks based on the degree-corrected stochastic block model. The method is flexible enough to detect anomalous nodes that arise from different mechanisms. Experiments on simulated and case study social network data streams demonstrate that our surveillance strategy can efficiently detect node propensity changes in the dynamic degree-corrected stochastic block model. 342A Production and Scheduling Contributed Session Chair: Faisal Alfaiz, Oregon State University, Corvallis, OR, United States, alfaizf@oregonstate.edu 1 - The Timing of Investment on Market Share and Profitability with Financing Constraints Kehong Chen, University of Science and Technology of China, Hefei, China, ckh0605@mail.ustc.edu.cn, Xiaohang Yue Whether to invest in market share that can improve the future demand and thereby raise future profits and reduce the likelihood of bankruptcy is a key concern faced by many startups firms when facing with a financing stress. We first identify under which condition the firm with no capital constraint has the incentive to invest on the market expansion. As a benchmark, we then analyze the investing strategy and operational decision of the firm with constrained financing capital. Given the limited borrowed capital, the firm has to make a trade-off between creating short-term profits from the production to avoid bankruptcy and long-term growth from market size investment. 2 - On Finding Process Capacity Yang Bo, Assistant Professor, The Chinese University of Hong Kong, New Territory, Hong Kong, boyang1102@gmail.com, Milind Dawande, Woonghee Tim Huh, Ganesh Janakiraman, Mahesh Nagarajan Recently, Gurvich and Van Mieghem (2015) show that the OM textbook formula to calculate process capacity fails in general. We show that it is hard to compute process capacity exactly and also to approximate it within a reasonable factor. These results are based on a novel characterization, which we establish, of process capacity that relates it to the fractional chromatic number of the associated “collaboration graph”. An important implication is that it is unlikely that we can replace the bottleneck formula with a simple but close approximation of process capacity. On the positive side, we show that capacity can be efficiently computed for processes for which the collaboration graph is a perfect graph. WC19

510