2016 INFORMS Annual Meeting Program

WB55

INFORMS Nashville – 2016

2 - IT Processes Improvement

5 - Serial Inventory Systems With Markov-modulated Demand: Derivative Bounds, Asymptotic Analysis, And Insights Yue Zhang, Duke University, 5507 Butterfly Ln Apt 207, Durham, NC, 27707, United States, yueyue.zhang@duke.edu, Li Chen, Jing-Sheng Jeannette Song We consider the inventory control problem for serial supply chains with continuous, Markov-modulated demand. We perform a derivative analysis and develop general, analytical solution bounds for the optimal policy. We further derive a simple procedure for computing near-optimal heuristic solutions. We next perform asymptotic analysis with long replenishment lead time. We show that the relative errors between our heuristics and the optimal solutions converge to zero as the lead time becomes sufficiently long, with the rate of convergence being the square root of the lead time.

Larisa Shwartz, IBM, lshwart@us.ibm.com With the advent of cognitive computing, new generations of cognitive systems and services are being conceived. In IT service management, cognitive approaches used for optimization and automation of IT Service management processes. We discuss an integrated framework for problem resolution that enables an automated discovery of informative phrases from IT incident tickets which are later used to construct knowledge and facilitate automation of IT service management. The effectiveness and efficiency of our framework are evaluated by an extensive empirical study of a large scale real ticket data. 3 - Business-driven Optimization Of It Service Configuration In Public And Hybrid Clouds Based On Performance Forecasting Genady Grabarnik, St. John’s University, Queens, NY, Mauro Tortonesi, Larisa Shwartz Modern Cloud environments are rapidly evolving, leading to a growing adoption of dynamic pricing for virtual resources and of speedier deployment tools and to the emergence of hybrid Cloud scenarios. The need to address these challenges is paving the way to a new generation of Cloud services, capable of adapting to changes in their operating conditions and deployment environments by dynami- cally realigning their configuration. However, their management will require new and more sophisticated tools. This paper presents a new optimization solu- tion for Cloud-based IT services, that tries to address these issues by using queu- ing analysis of the services’ workflows and ILP optimization. Chair: Yue Zhang, Duke University, 5507 Butterfly Ln Apt 207, Durham, NC, 27707, United States, yueyue.zhang@duke.edu 1 - Deep Learning For Newsvendors Afshin Oroojlooyjadid, Lehigh University, 200 West Packer Ave, Bethlehem, PA, 18015, United States, afo214@lehigh.edu, Martin Taká , Lawrence Snyder We study a newsvendor problem in which each demand observation also has a set of features. We propose an algorithm based on deep learning that optimizes the order quantity based on the features. It integrates the forecasting and inventory-optimization steps, rather than solving them separately. The algorithm does not require probability distributions. Numerical experiments on real-world data suggest that our algorithm outperforms approaches from the literature, including data-driven and SVM approaches, especially for volatile demands. 2 - Inventory Repositioning In Product Sharing Networks Xiang Li, University of Minnesota, 2508 Delaware St SE, Apt 473A, Minneapolis, MN, 55414, United States, lixx1315@umn.edu, Saif Benjaafar, Xiaobo Li We study a product sharing network in which customers can pick up a product without reservation, and are allowed to keep the product for as long as they want, without committing to a specific return time or location. We model the periodic inventory re-positioning as a Markov decision process. We characterize the qualitative properties of the optimal policy . 3 - Dynamic Inventory And Price Control In The Face Of Unknown Demand Tingting Zhou, Rutgers university, Newark, NJ, 07102, United States, tingzhou@rutgers.edu, Michael N Katehakis, Jian Yang We study adaptive policies that combat unknown demand in a dynamic inventory and price control setting. Inventory control is achieved by targeting newsvendor ordering quantities for empirical demand distributions learned over time. On top of that, demand-affecting prices are selected in a fashion that balances between exploration and exploitation. When burdened with the task of selecting the most profitable price, bounds for the regret can range between the orders of T1/2and those of T2/3. Simulation studies are conducted as well. 4 - Approximating Optimal Inventory Policies For Assemble To Order Manufacturing Systems Levi DeValve, Duke University, 716 Turmeric Lane, Durham, NC, 27713, United States, levi.devalve@duke.edu, Yehua Wei, Sasa Pekec We study the classical one-period assemble-to-order problem, modeled as a two stage stochastic integer program with recourse. We leverage a primal-dual approach to develop several approximation methods based on newsvendor solutions. We identify co-monotone demand and symmetric hierarchy systems as special cases where a component newsvendor solution is optimal under a constant mark-up assumption, and provide closed-form bounds on sub-optimality for more general cases. Further, we establish closed-form bounds for systems where components serve many products and show asymptotic optimality. WB55 Music Row 3- Omni Inventory Management VIII Contributed Session

WB56 Music Row 4- Omni Predictive Analytics in eBusiness Sponsored: EBusiness Sponsored Session

Chair: Ajit Sharma, Carnegie Mellon University, 5000 Forbes Ave, Pittsburgh, PA, 15213, United States, ajits1@andrew.cmu.edu 1 - Online Assortment Optimization at Scale Deeksha Sinha, MIT ORC, deeksha@mit.edu, Theja Tulabandhula We revisit the problem of assortment optimization in retail and consider a setting where (a) the product universe is large, (b) there are time constraints in offering an assortment, and (c) the user choice model can be personalized. In this setting, we propose an offline algorithm based on the theory of locality sensitive hashing (LSH) that computes approximately optimal assortments quickly. We also perform a sensitivity analysis of the optimal assortment to the choice model used. We then propose an online learning setup where we get feedback on the quality of assortments offered in each round. We come up with new online algorithms based on our offline solutions that can learn the user choice model quickly. 2 - Crowd-driven Competitive Intelligence: Understanding The Relationship Between Local Market Structure And Online Rating Distribution. Dominik Gutt, Paderborn University, Dominik.gutt@upb.de, Philipp Herrmann, Mohammad Saifur Rahman Crowdsourced information, such as online ratings, are increasingly viewed as a critical source for understanding local market dynamics. A key aspect of utilizing online ratings to derive competitive market intelligence is to delineate the systematic relationship between local market structure and distributional properties of online ratings. Using restaurant review data from Yelp.com for 372 isolated markets in the U.S., our empirical findings suggest that an increase in competition leads to a broader range of ratings and to a decrease in the average mean rating in a market. Moreover, we present evidence in support of both the internal and external validity of Yelp’s crowdsourced online ratings. 3 - Networks And Income: Evidence From Individually Matched Income And Mobile Phone Metadata Guillaume Saint-Jacques, MIT Sloan School Of Management, Cambridge, MA, 02142, United States, gsaintja@mit.edu, Eaman Jahani, Pål Roe Sundsøy, Johannes Bjelland, Bjørn-Atle Reme, Sinan Aral, Alex “Sandy” Pentland Measuring the relationship between income and various properties of one’s social network has proven difficult because it requires data on income and social ties to be matched at the individual level. We offer the first large-scale investigation of this question using data that is both large scale and individually matched. How are ego-networks different across income levels? Are there measurable differences in degree, reciprocity, diversity and centrality? We use a dataset of Call Detail Records from an Asian country of over 100M individuals and income surveys sent to over 110,000 individuals. We use location data to control for location effects, rather than rely on it to match incomes.

414