Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SB42

Jarrod D. Goentzel, Massachusetts Institute of Technology, 77 Massachusetts Avenue, E38-650, Cambridge, MA, 02139, United States, Gilberto Montibeller, Corinne Carland Malaria poses one of the greatest challenges to global health. Accurate diagnosis, with rapid diagnostic tests (RDTs), is critical because presumptive treatment of malaria wastes resources and increases the risk of drug resistance. In this project, we developed a value model for the agents along the RDT kit supply chain in Uganda. The model, based on multi-attribute value theory, enables an understanding of the motivations of each agent in the supply chain. The analysis identified a new package of incentives that proved to be robust against the variation of relevance of agents in the chain. 2 - New Methods for Improved Decision Making in Unconventional Field Development Andrew Beck, University of Texas-Austin, Austin, TX, 78757, United States, J. Eric Bickel We developed a decision support tool for Statoil’s US Onshore portfolio. This tool includes an asset valuation model and price and production uncertainty models. Because of high dimensionality, finding an optimal dynamic development schedule based on learning is difficult. We combined the Least Squares Monte Carlo algorithm with forward simulation to replicate an annual cycle of drilling, learning, and adaptation. By finding the set of alternatives that yields the highest net present value, the user ends up with an expected NPV for the asset, and a recommended year 1 development plan. 3 - Safer Skies in Spain David Rios Insua, ICMAT-CSIC and Royal Academy of Sciences, Nicolas Cabrera 13, Madrid, 28049, Spain, Javier Gómez, C sar Alfaro, Verónica Elvira, Pablos Hernández-Coronado, Fran Bernal As required by the International Civil Aviation Organization, nations must develop a so-called State Safety Program (SSP) to promote a proactive approach to safety oversight and management at country level. SSPs support strategic decision-making and resource allocation to areas with higher aviation safety risks. The Spanish Aviation Safety and Security Agency (AESA) developed a novel methodology with full use of numerous Decision and Data Sciences methods and implemented it in an R-based tool. This has allowed AESA to better support its safety decision making and attain considerable savings. n SB42 North Bldg 227A Joint Session ISim/APS: Simulation Optimization II Sponsored: Simulation Sponsored Session Chair: David Eckman, Cornell University, Cornell University Co-Chair: Shane Henderson, Cornell University 1 - Sequential Inferential Optimization via Simulation Under Input Model Risk Eunhye Song, Penn State University, 310 Leonhard Building, University Park, PA, 16802, United States Optimization via simulation (OvS) is subject to model risk when the input distributions to the simulator are estimated based on finite real-world data. We sequentially fit a Gaussian process metamodel to the conditional means of solutions as a function of the input model parameters and the design parameters to assist OvS and define a solution to be inseparable from the real-world optimum if the expected improvement (EI) of the solution’s conditional mean is less than a threshold specified by a user. At each iteration, we compute EIs of all solutions to form a set of inseparable solutions from the optimum. We simulate a solution that can best reduce the size of the set until a stopping criterion is met. 2 - Simple Top-two Sampling Algorithms for Best-arm Identification Daniel Russo This talk considers the optimal adaptive allocation of measurement effort for confidently identifying the best among a finite set of options or designs. I propose top-two sampling, a template for designing algorithms that adaptively allocate measurement effort. Four simple Bayesian algorithms designed via this template are shown to have strong performance in numerical experiments, and a unified analysis establishes each satisfies strong asymptotic optimality properties. 3 - A Knockout-tournament Approach to Large-scale Indifference-zone Ranking and Selection Ying Zhong, City University of Hong Kong, 22 Cornwall Street, KLN Tong, KLN, Hong Kong, L. Jeff Hong Ranking and selection is similar to a sport tournament where the goal is to find a champion. Traditional R&S procedures typically compare the best with all competing K-1 alternatives. Motivated by the form of knockout tournaments typically used in tennis grand slams, we propose a R&S procedure where the best only needs to survive log_2 K comparisons. We show that the procedure works well for large-scale problems and is particularly suitable for commercial cloud services.

n SB40 North Bldg 226B Statistical and Optimal Learning Sponsored: Applied Probability Sponsored Session Chair: Ilya O. Ryzhov, University of Maryland, College Park, MD, 20742, United States 1 - The Diverse Cohort Selection Problem John P. Dickerson, University of Maryland, 3217 A.V. Williams Building, University of Maryland, College Park, MD, 20742, United States How should a firm allocate its limited interviewing resources to select the optimal cohort of new employees from a large set of job applicants? How should that firm allocate cheap but noisy r sum screenings and expensive but in-depth in-person interviews? We view this problem through the lens of combinatorial pure exploration (CPE) in the multi-armed bandit setting, where a central learning agent performs costly exploration of a set of arms before selecting a final subset with some combinatorial structure. We present new algorithms with theoretical guarantees and provide experimental validation on real data from a hiring process at a large US-based graduate program. 2 - Data-driven Decision-making Using Variational Bayes Harsha Honnappa, Purdue University, West Lafayette, IN, United States We consider a canonical problem where an agent uses Bayesian updating to learn to take optimal actions from independent and identically distributed (i.i.d.) observations of the state of a stochastic system. Given these observations, beliefs over unknown state parameters are summarized by a posterior distribution, allowing decision-making which, while Bayes-optimal, is typically computationally intractable. In this talk, we present asymptotic and empirical analyses of two algorithms for learning optimal decision-rules, which we call naive variational Bayes (NV) and loss-calibrated variational Bayes (LC) algorithms for data-driven decision-making. 3 - Index Policies and Performance Bounds for Dynamic Selection Problems David Brown, Duke University, Fuqua School of Business, 100 Fuqua Drive, Durham, NC, 27708, United States, James E. Smith In this work, we study heuristics and performance bounds for dynamic selection problems. Examples of dynamic selection problems include assortment planning with demand learning, design of clinical trials, and screening a pool of applicants (e.g., for admission to colleges). We study a number of index-based policies, including one based on a Lagrangian relaxation of the problem. We characterize the performance of this Lagrangian index policy compared to an optimal policy. We also develop performance bounds that combine Lagrangian relaxations with information relaxations and show that these new performance bounds outperform standard Lagrangian relaxation bounds. 4 - A New Rate-balancing Algorithm for Ranking and Selection Ye Chen, 1305 Mathematics Building, University of Maryland, College Park, MD, United States, Ilya O. Ryzhov Recent years have seen a surge of interest in the study of optimal convergence rates in ranking and selection, based on the large deviations approach of Glynn & Juneja (2004). We present a new algorithmic and computational approach that adaptively learns the optimal allocations derived by Glynn & Juneja (2004) without any tunable parameters. Our proposed BOLD (Balancing Optimal Large Deviations) method assumes that the samples come from a known distributional family, but is applicable and quite easy to implement for multiple families (in particular, it does not require normality). We report our progress on the theoretical analysis of BOLD and present preliminary numerical results. n SB41 North Bldg 226C Joint Session DAS/Practice Curated: Decision Analysis Practice Award Competition Sponsored: Decision Analysis Sponsored Session Chair: Saurabh Bansal, Penn State University, State College, PA, 16801, United States Co-Chair: Michael C. Runge, USGS Patuxent Wildlife Research Center, 12100 Beech Forest Rd, Laurel, MD, 20708, United States Co-Chair: Michael Fitch, Chevron, San Ramon, CA, 9458, United States 1 - Modeling the Value of Agents in Supply Chains of Malaria Rapid Diagnostic Test Kits with Decision Analysis

45