2016 INFORMS Annual Meeting Program

TB48

INFORMS Nashville – 2016

TB48 210-MCC Optimization and Statistical Learning Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Dimitri Papadimitriou, Alcatel-Lucent, Bell-Telephonelaan, Brussels, BC, 00000, Belgium, dimitri.papadimitriou@alcatel-lucent.com Co-Chair: Dimitri Papadimitriou, Bell Labs, Copernicuslaan 50, Antwerp, 2018, Belgium, dimitri.papadimitriou@nokia.com 1 - Statistical Learning Approaches For Stochastic Optimization (SLASO) Suvrajeet Sen, USC, s.sen@usc.edu, Yunxiao Deng SLASO is a new distribution-free concept for Integrative Analytics, bringing together both Predictive and Prescriptive Analytics. We will illustrate the power of this concept by demonstrating how multiple activities, such as sales, marketing, and production planning can all work from the same data sources, thus helping to coordinate decisions from a variety of groups within an organization. The SLASO framework helps cross-fertilize members of an analytics team so that tools such as regression (Statistics), linear programming (Optimization), variance reduction (Simulation) and others can be viewed from an integrative perspective, rather We will introduce how SLASO can be used for Time Series Applications by studying a single echelon inventory model in which the demand data are time dependent and stochastic. To minimize the expected cost of holding inventory plus lost sales, we formulate a Stochastic Linear Programming model which combines statistics as well as optimization by adopting SLASO framework. With other two methods (Newsvendor and Demand Forecasting), this approach performs better in both Back-Testing and Stress-Testing during validation analysis. 3 - Bayes-optimal Entropy Minimization For Active Learning In Conjoint Analysis Stephen N Pallone, Cornell University, 290 Rhodes Hall, Cornell University, Ithaca, NY, 14853, United States, snp32@cornell.edu, Peter Frazier, Shane Henderson Choice-based conjoint analysis is a method for preference elicitation where the user is offered a set of alternatives and chooses the preferred option. The rate at which we learn depends on the alternatives offered. We model the user’s preferences through a linear classifier. Under certain noise assumptions, we prove a linear lower bound on the posterior entropy of this linear classifier, and show entropy pursuit can attain the bound when alternatives can be fabricated. Further, we explore an information theoretic variant of the knowledge gradient policy that selects comparative questions to greedily minimize interaction information, and numerically compare this policy with entropy pursuit. 4 - Learning Uncertainty Sets And Automation Of Robust Formulation Dimitri Papadimitriou, Bell Labs, dimitri.papadimitriou@nokia.com Machine learning shares deep connections with robust optimization as they both perform by adding uncertainty in the model, formulating the optimization problem, and exploiting mathematical programming. Machine learning applies to i) model uncertainty by automating processing of noisy or aleatory data to produce perturbation sets, ii) automatically derive robust formulation but also adapt decision rules associated to adjustable variables, and iii) learn about the behavior of resolution algorithm(s) and tune its execution to improve its performance. We illustrate them on three network optimization problems, the multi-commodity network flow, facility location and hub location problem. than the current lens of disciplinary stove-pipes. 2 - Time Series Applications Of SLASO Yunxiao Deng, USC, yunxiaod@usc.edu

2 - Effective Teaching For Mastery Of Boolean Constraints Scott Stevens, Professor, CIS & Business Analytics, James Madison University, James Madison University, MSC 202, Harrisonburg, VA, 22807, United States, stevensp@jmu.edu, Susan Wright Palocsay Mixed integer optimization problems often include constraints involving Boolean variables—variables that can equal only 0 or 1—representing the truth values of some logical propositions. The pure compound propositions of either “all Ai are true” or “at least one Ai is true” for some collection of Boolean variables Ai are easy to represent as linear expressions, but students frequently find propositions of the form (pure compound 1) implies (pure compound 2) to be baffling. We introduce a technique that allows any such proposition to be written as one or more linear constraints by applying three simple rules: decomposition, translation, and compression, and then provide evidence of efficacy. 3 - A Review Of The Literature For Operations Management Education Susan Wright Palocsay, James Madison University, Harrisonburg, VA, United States, palocssw@jmu.edu, Michael Busing Operations management (OM) has maintained a close association with OR/MS as its scope has broadened from manufacturing to service processes. In response, OM curricula have undergone significant revision with a corresponding increase in OM pedagogical studies. We will present a summary of OM educational publications and discuss OM teaching trends. Moderator: Maria Esther Mayorga, North Carolina State University, 111 Lampe Drive, Campus Box 7906, Raleigh, NC, 27695, United States, memayorg@ncsu.edu 1 - Topics For Phd Students Maria Esther Mayorga, North Carolina State University, memayorg@ncsu.edu This session will serve as a panel discussion on topics of interest for PhD students nearing graduation. Topics include: - deciding on industry versus academia - how to prioritize objectives towards then end of the PhD Process - work/life balance when pursuing tenure - networking to achieve a desired faculty position - how to position yourself when pursuing the market - networking at conferences such as INFORMS 2 - Panelist Maria Esther Mayorga, North Carolina State University, memayorg@ncsu.edu 3 - Panelist Karen T Hicklin, University of North Carolina at Chapel Hill, Chapel Hill, NC, 27599, United States, khicklin@email.unc.edu TB51 213-MCC Matchings and Assignments with Societal Impact Sponsored: Public Sector OR Sponsored Session Chair: Tina Rezvanian, Northeastern University, 360 Huntington Avenue, Boston, MA, 02115, United States, rezvanian.t@husky.neu.edu Co-Chair: Ozlem Ergun, Northeastern University, 453 Meserve, 360 Huntington Avenue, 360 Huntington Avenue, MA, 02115, United States, o.ergun@neu.edu 1 - Two-sided Stable Matching In Markets With Multiple Periods Many industries demand mechanisms for job assignments that are sustainable over multiple periods of time. It is a well-known fact that the core stays the same under any stable matching algorithm and therefore no such matching can guarantee sustainability and perfectness as a dominant strategy for participants as same participants stay unsatisfied. We introduce a popularity measure that identifies a connection between stability of the match and fairness for those unsatisfied. We use this trade off to employ negotiation schemes in our proposed algorithm. By analyzing network structures of the large-scale markets we are able to acquire near-optimal solutions. Tina Rezvanian, Northeastern University, Boston, MA, United States, rezvanian.t@husky.neu.edu, Ozlem Ergun TB50 212-MCC Panel: Topics for PhD Students Sponsored: Minority Issues Sponsored Session

TB49 211-MCC Educational Research in OR/MS Sponsored: Education (INFORMED) Sponsored Session

Chair: Susan Wright Palocsay, James Madison University, Harrisonburg, VA, United States, palocssw@jmu.edu 1 - Batching In Higher Education

Jan Riezebos, University of Groningen, Nettelbosje 2, P.O. Box 800, Groningen, 9700 AV, Netherlands, j.riezebos@rug.nl, Iris F Vis Models for optimal batch sizes in industrial context, such as economic lot sizing, cannot directly be applied to the context of higher education. However, in higher education, batching is even more important than in industrial contexts, as it is not just the economical impact that has to be considered, but also various effects on learning. We explore possible ways to extend OR models for batching in higher education.

280