INFORMS 2021 Program Book

INFORMS Anaheim 2021

MD25

3 - Optimal COVID-19 Containment Strategies Hyun-Soo Ahn, Professor, University of Michigan, Ross School of

apply them to an industry dataset containing nearly 5 million messages. We find that service interactions are characterized by strong customer-agent dependency and the centrality of the process’s cross- and self-excitation attributes. Finally, we use our models in a data-driven simulation to improve upon contact center routing algorithms, yielding significant decreases in wait times. MD24 CC Room 205A In Person: Emerging Topics in Data-Driven Supply Chain and Revenue Management General Session Chair: Divya Singhvi, MIT, Cambridge, MA, 02139-4230, United States 1 - Joint Assortment Optimization and Customization under a Mixture of Multinomial Logit Models: On the Value of Personalized Assortments Omar El Housni, Cornell Tech, New York, NY, 10044, United States, Huseyin Topaloglu We consider a joint customization and assortment optimization problem under a mixture of MNL models. A firm faces customers of different types, each making a choice according to a different MNL model. In the first stage, the firm picks an assortment of products to carry subject to a cardinality constraint. In the second stage, a customer of a certain type arrives into the system. Observing the type of this customer, the firm customizes the assortment that it carries by, possibly, dropping products from the assortment. We study the complexity of this problem, present tight bounds on the value of customization and design a novel algorithm that gives $\Omega(1/ \log m)$-approximation to the problem, where m is the number of customer types. The problem has obvious connections to assortment optimization under a mixture of MNL models, which can only admit a $O(1/m)$- approximation. 2 - Math Programming Based Reinforcement Learning For Multi-echelon Supply Chain Management Divya Singhvi, MIT, Cambridge, MA, 02139-4230, United States, Pavithra Harsha, Ashish Jagmohan, Jayant Kalagnanam, Brian Quanz Reinforcement Learning has lead to considerable break-throughs in diverse areas such as robotics, games and others. But the application to RL in complex decision making problems remains limited. Many problems in Operations Management are characterized by large action spaces and stochastic system dynamics. These characteristics make the problem considerably harder to solve for existing RL methods that rely on enumeration techniques. To resolve these issues, we develop Programmable Actor Reinforcement Learning (PARL), a value iteration method that uses techniques from IP, SAA and optimal discretization of continuous random variables. We then apply our algorithm to real-world inventory management problems with complex supply chain structures and show that PARL outperforms state-of-the-art RL and inventory optimization methods in these settings. MD25 CC Room 205B In Person: Innovation and Design Management General Session Chair: Sidika Tunc Candogan, University College London, London, E14 5AA, United Kingdom 1 - Agile Development is Not (Always) A Panacea: An Experimental Study Evgeny Kagan, Johns Hopkins Carey Business School, Baltimore, MD, 21202-4673, United States, Tobias Lieberum, Sebastian Schiffels We experimentally study the effects of Agile project planning techniques on performance in two tasks: (1) a creative task reflective of product innovation with an open solution space and limitless creative possibilities, and (2) a search task reflective of business model innovation, in which subject search through a finite (but complex) solution landscape. Our results suggest that Agile techniques significantly improve performance in the first (creative) task, but harm performance in the second (search) task.

Business, Ann Arbor, MI, 48109, United States, John M. Silberholz, Xueze Song, Xiaoyu Wu

Policymakers rely on mathematical models to plan non-pharmaceutical interventions (NPIs) such as lockdowns to combat COVID-19, weighing health benefits against economic costs. Many such models have been created, but they vary in forecasts and recommendations. We find an NPI policy (how to change restrictions based on the current pandemic status) optimized wit h a single model can perform poorly (more than double the cost) when evaluated with a different model. We optimize across multiple models and find policies that all models find effective. The most effective policy varies significantly by state, due to differences in the NPIs selected by states and the response of citizens to those NPIs. 4 - To Catch A Killer: A Data-Driven Personalized and Compliance- Aware Sepsis Alert System Zahra Mobini, The University of Texas at Dallas, Richardson, TX, 75080-3021, United States, Mehmet U.S. Ayvaci, Ozalp Ozer In this study, we develop an alert system for early detection of sepsis. Our system personalizes alerts to individual patients and accounts for caregivers’ compliance behavior. Integrating predictive approaches with prescriptive ones in an MDP framework, our system determines when to alert for sepsis. We find that personalized alerts are essential for capturing the heterogeneity of sepsis risk among patients, while compliance-aware alerts are necessary when caregivers’ compliance varies during a patient’s hospital stay. Using data from a large hospital system in the US, we back test and validate our alert policy. On average, our system detects 22% more sepsis cases and triggers alerts 39 hours earlier (ranges 29-53) than the existing alert system. This time difference matters, as every hour of delay in providing proper sepsis treatment can increase mortality by up to 8%. Management General Session Chair: George Ball, Indiana University, Kelley School of Business, Bloomington, IN, 47405-1701, United States 1 - Need for Speed: The Impact of Website Performance on Online Retail Nil Karacaoglu, Assistant Professor, Fisher College of Business, Ohio State University, Columbus, OH, United States, Santiago Gallino, Antonio Moreno The share of e-commerce sales is rapidly increasing and so is the relevance of website performance. Weleverage novel retail and website performance data to investigate how website performance impacts onlinesales. The impact of speed and waiting time has been studied in various offline services. We extend thisliterature to online services and quantify the impact of website speed on brands. We estimate sizable adverseeffects of website speed slowdowns on online sales. 2 - Bad Things Come to Those Who Wait: Firm Stock Ownership, Recall Timing, and Stock Market Penalties George Ball, Indiana University, Kelley School of Business, Bloomington, IN, 47405-1701, United States, Jessica Darby, Dave Ketchen, Ujjal Kumar Mukherjee Firms often delay the decision to recall faulty medical devices long after they become aware of a defect. We examine how stock ownership of two key actors CEOs and institutional investors influences the speed with which medical devices are recalled. We then examine if the stock market penalizes firms differently based on recall decision-making speed and if this penalty varies with recall severity. We collect time-stamped data on 2,196 medical device recalls across 50 public medical device firms from 2002 to 2015. We find that firms with greater CEO and institutional investor ownership stakes recall medical devices more slowly. We also find that delaying recalls magnifies the stock market penalty attributable to the recall, indicating that bad things in the form of stock penalties may come to those who wait too long to initiate recalls. 3 - The Co-production of Service: Modeling Service Times in Contact Centers Using Hawkes Processes Andrew Daw, Marshall School of Business, University of Southern MD23 CC Room 204C In Person: Service and Quality Operations

California, Los Angeles, CA, 14853-3801, United States, Antonio Castellanos, Galit Bracha Yom-Tov, Jamol Pender, Leor Gruendlinger

In customer contact centers, a successful service interaction involves a messaging dialogue between a customer and an agent. In this talk, we propose, develop, and compare new stochastic models for service co-production in a contact center. Our models distinguish between the role of the customer and of the agent, reflect the service process’s dynamic evolution over time based on its own history, and include additional behavioral and operational aspects. To evaluate our models, we

79