INFORMS 2021 Program Book

INFORMS Anaheim 2021

MC15

4 - Selection of the Most Probable Best under Input Uncertainty Taeho Kim, KAIST, Daejeon, Korea, Republic of, Kyoung-Kuk Kim, Eunhye Song We suggest a novel robust ranking and selection under input uncertainty named “selection of the probable best”. Our formulation aims to find the most probable alternative over the posterior of input parameters. The finite support case on the input posterior is considered first. Our theoretical results provide an optimal computing budget allocation (OCBA) scheme derived by a lower bound on the large deviation rate of the false selection probability. Further, several dynamic sampling algorithms which achieve this bound in the limit are presented. Numerical experiments support our findings. 5 - Representation for Conditional Expectation with Application To Dynamic Programming Yi Zhou, University of Maryland, College Park, College Park, MD, United States, Michael Fu, Ilya O. Ryzhov, Steven I. Marcus We consider the problem of estimating a conditional expectation. By formulating the conditional expectation as a ratio of two derivatives, we can apply the generalized likelihood ratio method to represent the conditional expectation using ordinary expectations. We then apply this representation tool in stochastic dynamic programming problems to derive new estimators that can be incorporated into existing learning algorithms, which improves their performances. 6 - Wafer Lot Scheduling in a Real-world Semiconductor Photolithography Bay Patrick Deenen, University of Technology-Eindhoven, Eindhoven, Netherlands A semiconductor wafer fabrication facility (fab) consist of many different bays, one of these is the photolithography bay. Since this bay often forms the bottleneck of the fab, scheduling lots in this area is of key for the overall fab’s performance. A real-world case study at Nexperia is presented, to demonstrate the benefit of scheduling opposed to current dispatching operations. A simulation model which accurately represents represents the real fab. This model is used to analyze different production control techniques and access their performance (1) the throughput of the photolithography bay, (2) the realization of the operational due-date of jobs and (3) the WIP balance of the downstream bays. MC15 CC Room 201C In Person: Statistical Methods for Contemporary Business Applications General Session Chair: Gourab Mukherjee, University of Southern California, Los Angeles, CA, 90089-0809, United States 1 - Personalized Treatment Selection Using Causal Heterogeneity Kinjal Basu, LinkedIn Corporation, Sunnyvale, CA, 94085-4172, becomes the selected treatment for the entire population. However, the effect of a given treatment can differ across experimental units, and a personalized approach for treatment selection can greatly improve upon the usual global selection strategy. In this work, we develop a framework for personalization through (i) estimation of heterogeneous treatment effect at either a cohort or member-level, followed by (ii) selection of optimal treatment variants for cohorts (or members) obtained through (deterministic or stochastic) constrained optimization. Through simulations and real-life experiments, we show the efficacy of the method. 2 - Estimating Promotion Effectiveness in Email Marketing: A High- dimensional Bayesian Joint Model for Nested Imbalanced Data Gourab Mukherjee, University of Southern California, Los Angeles, CA, 90089-0809, United States We consider a large-scale, cross-classified nested joint model for modeling customer responses to opening, clicking, and purchasing from promotional emails. Our logistic regression-based joint model contains crossing of promotions and customer effects, and allows estimation of the heterogeneous effects of different promotion emails after adjusting for customer preferences, attributes, and historical behaviors. Using data from an email marketing campaign of an apparel company, we exhibit the varying effects of promotions. We conduct Bayesian estimation by using a block Metropolis-Hastings algorithm that not only incorporates nested subsampling to tackle the severe imbalance between conversions and no conversions, but also uses additive transformation-based modifications of random walk Metropolis to scale estimation for large numbers of customers. United States, Ye Tu, Cyrus DiCiccio, Romil Bansal, Preetam Nandy, Padmini Jaikumar, Shaunak Chatterjee Randomized experimentation (or A/B testing) is widely used in the internet industry to measure the metric impact obtained by different treatments. A/B tests identify the treatment variant showing the best performance, which then

3 - Analyzing Consumer Choice of Hybrid Cars: A New Multinomial Probit Model with Spatially Correlated Preference and Response Coefficients Sivaramakrishnan Siddarth, University of Southern California, Los Angeles, CA, 90089-1424, United States We propose a new spatial multinomial probit model that allows different subsets of the preference and choice coefficients to have their own unique spatial structures. We apply the model to vehicle choice data from the Sacramento market in 2008 and show how the estimated parameters can help improve the effectiveness of target marketing programs designed to accelerate hybrid adoption. MC17 CC Room 202A In Person: INFORMS Prize Informs Special Session: INFORMS Prize Informs Special Session Session Chair: Erica Z. Klampfl, Ford Motor Company, Dearborn, MI, 48124, United States 1 - Operations Research at Amazon Transport Tim L. Jacobs, Amazon, Tempe, AZ, 85284-3961, United States, Maurico Resende, Nilay Noyan Bulbul Amazon applies OR and analytics throughout its business, including facility location, inventory management, androuting. Transportation plays a key role in giving our customers a great experience. Though last-mile delivery is perhapsthe most customer-facing mode of transportation, middle-mile transportation is just as critical. We focus on Amazon’smiddle-mile transportation research science team. The Middle-Mile Planning, Research and Optimization Sciences teamof Amazon Transportation Services is built on three pillars: surface research, air science & tech, and pricing & yieldmanagement. We describe their impact on provisioning Amazon’s retail delivery network. Chair: Brian Tarroja, University of California, Irvine, CA, United States 1 - Estimating the Sensitivities of Power System Components to Heat and Drought for Climate-informed Planning Ana Dyreson, Assistant Professor, Michigan Technological University, Houghton, MI, United States, Sean Turner, Ariel Miara, Thushara De Silva, Stuart Cohen, Naresh Devineni, Nathalie Voisin, Jordan Macknick Climate-informed planning for future power systems requires understanding the sensitivity of physical assets including power plants, buildings, and transmission lines to climate stress. Capturing physically-based impacts is difficult in regional- to-national scale capacity expansion and operational power grid models because of computational tractability and the diversity in individual systems. We review representations of hydroelectric, thermoelectric and demand components and their sensitivities to climate change, focusing on the characteristics of the Western U.S. interconnect. We discuss systematic uncertainty characterization moving forward. 2 - Extending Energy System Modeling to Include Extreme Weather Risks Jeffrey A. Bennett, University of Virginia, Charlottesville, VA, United States, Joseph F. DeCarolis, Andres F. Clarens Electric power system planning is supported by energy system optimization models which project future power plant and storage installations in order to meet demand at the lowest possible cost. These models have generally not incorporated the costs of damage to the electric power grid resulting from extreme weather events such as wind or flooding damage from hurricanes or fires caused by drought. In this talk, we present an extended energy system optimization model that incorporates hurricane risks and apply it to the context of Puerto Rico, an island territory of the United States that had its electric grid severely damaged by Hurricane Maria in 2017. When hurricane trends are included, 2040 electricity cost projections increase by 32% based on historical hurricane frequencies and by 82% for increased hurricane frequencies resulting from climate change. MC18 CC Room 202B In Person: Climate Impacts on the Electric Power System General Session

62