2015 Informs Annual Meeting

MD19

INFORMS Philadelphia – 2015

MD19 19-Franklin 9, Marriott

3 - Optimal Resource Allocation Algorithms for Cloud Computing Siva Theja Maguluri, Postdoctoral Researcher, IBM TJ Watson Research Center, 1101 Kitchawan Road, Yorktown Heights, NY, 10598, United States of America, smagulu@us.ibm.com Jobs arrive at a cloud computing system according to a stochastic process and request resources like CPU, memory, etc and need service for a random amount of time. These jobs need to be scheduled on servers. The jobs are first routed to one of the servers when they arrive and are queued at the servers. Each server then chooses a set of jobs from its queues so that it has enough resources to serve all of them simultaneously. We present an optimal load balancing and scheduling algorithm. MD21 21-Franklin 11, Marriott Stochastic Models in Healthcare Sponsor: Health Applications Sponsored Session Chair: Sait Tunc, UW-Madison, 3233 Mechanical Engineering Building, 1513 University Avenue, Madison, WI, 53706, United States of America, stunc@wisc.edu Co-Chair: Oguzhan Alagoz, UW-Madison, 3242 Mechanical Engineering Building, 1513 University Aveneue, Madison, WI, 53706, United States of America, alagoz@engr.wisc.edu 1 - Robustness of Markov Decision Processes for Medical Treatment Decisions Lauren Steimle, University of Michigan, 1205 Beal Avenue, Ann Arbor, MI, 48109-2117, United States of America, steimle@umich.edu, Brian Denton Markov decision process (MDP) models are frequently used to study optimal policies for treatment of patients with chronic diseases. However, these models can be sensitive to estimates of transition probabilities and rewards. We discuss an approach for quantifying robustness of MDP-based policies with respect to parameter uncertainty. We illustrate our findings based on a model for optimal treatment of blood pressure and cholesterol in patients with type 2 diabetes. 2 - Ambulance Emergency Response Optimization in Dhaka, Bangladesh Justin Boutilier, University of Toronto, 5 King’s College Road, Toronto, ON, M5S 3G8, Canada, j.boutilier@mail.utoronto.ca, Moinul Hossain, Timothy Chan Dhaka, the capital city of Bangladesh and the tenth largest city in the world, does not currently have a centralized emergency medical service (EMS) system or 9-1- 1 type number. As a result, patients experience restricted access to healthcare. To address this problem, we have developed a novel data-driven robust location- routing model that can be applied to Dhaka and other developing urban centers. The model uses traffic data collected via GPS to construct an uncertainty set for travel times. 3 - Score Based Anticipative Transfer Requests in the Intensive Care Units Yasin Ulukus, University of Pittsburgh, Pittsburgh PA, United States of America, myu1@pitt.edu, Gilles Clermont, Guodong Pang, Andrew J. Schaefer The efficient operation and management of ICUs is critical to providing high quality of care while managing costs. We construct a new Transfer Score to estimate readmission and death probabilities. We further show that an anticipative transfer request policy combined with effective use of clinical markers can significantly decrease transfer delays without increasing the capacity. We present a Markov Decision Process (MDP) model for the transfer request problem and solve it via approximations 4 - Optimal Breast Cancer Diagnostic Decisions under the Consideration of Overdiagnosis Sait Tunc, UW-Madison, 3233 Mechanical Engineering Building, 1513 University Avenue, Madison, WI, 53706, United States of America, stunc@wisc.edu, Oguzhan Alagoz, Elizabeth Burnside Breast cancer overdiagnosis issue becomes more severe every year, a recent study approximates the annual cost of overdiagnosis to the United States as $243 million. We propose a large-scale MDP model to determine the optimal diagnostic strategy under the consideration of overdiagnosis by incorporating cytologic grade into the traditional breast cancer diagnostic decision problem.

Application of Nonlinear Optimization using Sequential Linear Programming Techniques with Xpress Sponsor: Computing Society Sponsored Session Chair: Zsolt Csizmadia, Principal Engineer, FICO, FICO House, Starley Way, Birmingham, B37 7GN, United Kingdom, zsolt.csizmadia@gmail.com 1 - State of Optimization in Advanced Process Control Process Control forms the backbone as well as the driving agent for almost all of process industries today. Smart algorithms combined with superior computational capabilities allow us to automate processes in a controlled fashion while optimizing environmental, safety and economic performance. This talk discusses recent advances in commercial Advanced Process Control technology by harnessing the latest developments in the optimization community along with the challenges going forward. 2 - Modeling Recursive Formulae in Xpress using Variable Eliminations Libin Varghese, Lead Modeling Developer, FICO, 1500 Broadway, Suite 1101, New York, NY, 10036, United States of America, LibinVarghese@fico.com Modeling a deposit pricing problem, that optimizes rates for a multiyear period, involves handling of various recursive formulae that link each time period to the next. We shall focus on how we modeled the problem in the Mosel modeling language using the new variable elimination feature of Xpress-Nonlinear and the performance improvements achieved. 3 - A New Optimality Measure for Sequential Linear Programming Methods Rishi Amrit, Shell International, Houston, TX, United States of America, R.Amrit@shell.com The KKT conditions are regarded as the definite first order optimality conditions for nonlinear programming though regularity conditions relatively rarely hold in practice. The convergence of nonlinear optimization algorithms based on first order approximations often focus on the progress made rather than the solution properties. We introduce a new optimality measure derived from the KKT conditions and explore the connection between the convergence of first order methods and the new measure. MD20 20-Franklin 10, Marriott Stochastic Models and Analysis for Cloud Computing Cluster: Cloud Computing Invited Session Chair: Yingdong Lu, IBM Research, 1101 Kitchawan Rd, Yorktown Heights, United States of America, yingdong@us.ibm.com 1 - Model Based Autoscaling of Hadoop Clusters Parijat Dube, IBM, 1101 Kitchawan Road, Yorktown Heights, United States of America, pdube@us.ibm.com, Li Zhang, Andrzej Kocut, Anshul Gandhi We develop novel performance models for Hadoop workloads that relate job execution time to various workload and system parameters such as input size and resource allocation. We employ statistical techniques to tune the models for specific workloads, including TeraSort and Kmeans. The tuned models are used to determine the resources required to successfully complete the Hadoop jobs as per the user-specified execution time SLA. 2 - Navigating the Amazon Cloud Aaron Yan, Data Scientist, Gravitant, 11940 Jollyville Road, #325N, Austin, TX, 78759, United States of America, aaron.yan@gravitant.com, Ilyas Iyoob Selecting the right level of reservation in the cloud is a tricky problem, especially when there are multiple reservation levels. In this paper, we explore the optimal levels of reservation for a portfolio of cloud servers that satisfy the CapEx and OpEx budget. The team has developed a web application that solves this problem and demonstrates the savings incurred from choosing the correct reservation pricing models. Zsolt Csizmadia, Principal Engineer, FICO, FICO House, Starley Way, Birmingham, B37 7GN, United Kingdom, zsolt.csizmadia@gmail.com

238